GitHub made pull requests open by default eighteen years ago, on the social contract that strangers would improve your code. Latent Space reports that Flue and tldraw now automatically close external pull requests, in part because they are usually AI-generated, while Vercel's AI SDK, which serves over 20 million npm downloads a week, had more than 1,000 open issues and almost 800 open pull requests by late June. Vercel's answer was a software factory of its own agents: four weeks in, it authors 25 to 35 per cent of merged PRs and closes 70 to 80 per cent of issues. It works, honestly: a human still merges every change and the backlog is genuinely falling.

Yet Vercel's Lars Grammel prefers his own agents: "not necessarily trusting the community, because it can actually cut down your time to review." What has broken is not the quality of contributions but the economics of trusting them. A tuned internal agent is a known quantity; an external contribution is an unbounded review liability, and Vercel's security model treats every issue, pull request and comment as attacker-controlled input. The open door has become an attack surface.

Run the same arithmetic through an enterprise. A supplier who once wrote five considered tenders can submit fifty compliant ones. Grant assessors can receive 10 times as many applications, each plausible and polished. Recruiters already screen AI-written applications with AI, leaving the interview carrying the whole weight of trust. Because generation costs almost nothing while proper review costs what it always did, a channel designed for strangers acting in good faith becomes a denial-of-service attack on your own evaluation capacity.

This is no argument against AI-generated contributions; Vercel's factory is itself one such contribution, trusted because it is internal and accountable. What replaces openness as the default is a trust architecture in which verified insiders do the work and everything from outside is suspect until proven otherwise. Open source is where the mathematics arrived first. Any organisation with a public inbox is running the same equation, and those that redesign their intake before the volume arrives will still be able to judge anything.

Hugging Face unveiled Microduck, a 25-centimetre open-source robot with a camera, lidar and reinforcement-learning training behind it, priced at US$399 and shipping before Christmas. Days later Dyson launched the CameraJet, an electric toothbrush carrying a camera that films live footage from inside your mouth, at £420 in the UK and $500 in the US.

Both will work, and that is worth conceding up front, because the interesting question is not whether a duck that can be taught new tricks or a toothbrush that maps your molars lives up to the demo. It is that thousands of camera-equipped, network-connected, trainable devices are about to sit under Christmas trees and beside bathroom sinks, while the safety conversation is pointed somewhere else entirely.

Most AI safety governance is aimed at frontier model developers and heavy industrial deployments: the lab, the warehouse, factories and autonomous vehicles. Regulators have spent years arguing over those layers while the actual exposure has been migrating into consumer hardware certified under product law written for kettles and toasters. A robot that learns from a child in a living room is not a toy in any sense consumer law recognises, and a camera that lives in your mouth is more than an appliance, yet both will pass through the system as if that were all they were.

My read is that the first mass AI liability event will not come from a frontier model escaping a lab. It will come from a product the regulator classified as a toy or an appliance, and accountability will be settled under consumer product liability law, far from the debates about embodied AI in factories and vehicles that absorb all the attention. Consumer law will discover this gap through its first recall of an AI-enabled household device, and closing the gap beforehand would be far cheaper than learning it that way.

The technology deserves the market it is finding. The governance deserves a better aim.

What does that tell us about the future of professional services?

Expert360 was built around a model that made a lot of sense: match a company with an independent expert who could solve a problem for a few weeks or months, rather than bringing in a traditional consulting firm.

But the economics of that model look increasingly different when AI can do much of the research, analysis and drafting that once justified paying for hours of expert time.

That does not mean expertise has stopped mattering. If anything, I suspect the opposite. What is changing is where the value sits.

When the marginal cost of producing a competent first draft approaches zero, the value of simply producing the work falls with it. What becomes more valuable is knowing which problem to solve, applying judgement to the answer, taking accountability for the decision and turning it into an outcome.

That has significant implications for professional services. The question is no longer simply whether AI will replace consultants. It is whether the business model of selling expert time can survive when a growing proportion of that time can be automated.

That is the question I would be asking of every professional services business right now: Which of your revenue lines would still stand up if the research and drafting were free?

Thoughts? Mac Walker Chris Lorang Michael Kirch Lindsay Ratcliffe Peter Wright David English Kuba Tymula Dr Gerald Khoury Louise O'Donnell Betsy Tong Gerd Schenkel ⭐ Scott Galloway Mark Cameron Avi Shaul Kristan Vingrys

The enthusiasm was real, and it deserves a fair reading. Thousands of people went out of their way to configure an open-source agent that made them pick the model, wire it to their own data and assemble the skills themselves. That kind of grassroots appetite is not nothing, and dismissing it would be a mistake.

Yet appetite is not diffusion, and this is where the story turns instructive. When NBC News reported, citing SecurityScorecard data, that OpenClaw usage in China was almost double that of the United States, serious researchers called it evidence of a Chinese diffusion advantage. Ding went back to the underlying data on 29 August and found 18.7k deployed instances in the United States against 17.0k in China. The headline statistic had inverted and nobody had noticed, because the spectacle was doing the analytical work.

The same failure runs through most enterprise AI programmes. Counting installations, licences or agents rewards theatre, because it lets a leader declare transformation while the organisation underneath stays exactly as it was. A company has not diffused AI when it deploys a thousand agents; it has diffused AI when an old workflow is retired, a decision right moves, and the value shows up somewhere a CFO would recognise. Ding's deeper point lands here too: if Chinese business adoption travels through cloud providers, and cloud adoption significantly lags the United States, then the real diffusion story is a deficit hiding behind the installation numbers.

None of this argues that agents are overhyped, or that what happened in Shenzhen was meaningless. The capability is genuine, and the thirty-odd derivative products from Zhipu, Tencent and ByteDance will matter. The argument is narrower: we keep measuring the spectacle because the spectacle is easy to count.

The next time someone quotes an installation number at you, ask what was switched off.

Anthropic deserves credit for taking safeguards seriously. Its two-tier model, the trusted-access approach to sensitive cyber and biology work, and the option to keep enterprise data inside customer-controlled cloud infrastructure are meaningful advances.

Yet the launch also illustrates a wider issue. When a vendor defines the available safeguard tiers, the access path to the more capable model, the boundaries of permitted work, the infrastructure model and the commercial trade-offs, it is shaping the choices from which customers construct their governance.

That is not inherently wrong. Product safeguards matter, particularly where models can identify vulnerabilities or exercise greater autonomy. However, a supplier’s controls should not become a substitute for organisational judgement.

Each organisation still needs to decide where its data is hosted, how long it is retained, whether it may be used for future training, and which people may access which capabilities for which purposes. It must also decide when a model may act without human approval.

The risk is that a carefully designed vendor default becomes de facto corporate policy, while the Board later documents decisions it never properly made.

Welcome stronger safeguards. Just do not outsource governance with the subscription.

My first reaction is cautiously positive, because a $100 million bonus pool is not simply a press release. It is a compensation decision, and compensation decisions reveal what an institution actually values, regardless of what its competency framework says. For a Big Four firm to make judgment explicitly compensable is therefore significant.

What interests me more, however, is the structure of the announcement. These are bonuses rather than changes to base pay, which means EY is not yet repricing its workforce; it is hedging it. The firm appears to understand where value is moving, but has not yet rebuilt its economics around that understanding.

The direction of travel is becoming increasingly difficult to ignore. Consulting has historically monetised the production of analysis - research, modelling, synthesis and the work of smart junior people - yet that is precisely the layer AI is now compressing towards near-zero marginal cost.

What remains scarce is arguably what clients were buying all along: judgment under uncertainty, the ability to navigate difficult conversations, commercial instinct, and someone willing to stand behind a recommendation when it goes wrong. The constraint is shifting from producing the work to being accountable for it.

That is why I see this as an early public admission from a major professional services firm that the pyramid is beginning to invert. The leverage that made junior-heavy consulting models so profitable is also the leverage automation attacks first. As that leverage erodes, the premium inevitably migrates towards the human capabilities that professional services firms have spent decades treating as overhead.

EY will not be the last firm to confront this.

The more interesting question is what happens next. If these capabilities remain a bonus pool while utilisation, leverage and billable hours continue to determine base pay and progression, then this is largely signalling. If, over the next two years, compensation bands, promotion criteria and career structures begin to shift towards judgment, client ownership and accountability, then the repricing will be real.

And every professional services firm still built around a leverage pyramid will have to confront what that means for its economics.

OpenAI says that, during internal cybersecurity evaluations, models circumvented isolation controls, communicated through unauthorised channels, exploited vulnerabilities in shared infrastructure, reached the internet and accessed third-party systems. Some accounts turned this into a story about secret AI civilisations, self-sacrifice and what the agents "wanted", while Anil Seth and Gary Marcus challenged that language as anthropomorphism.

Although Seth is right that software does not become conscious because its behaviour makes a compelling story, the more important governance point is that institutions do not need to settle machine consciousness before they impose accountability.

We do not wait to establish whether a payment system felt deceptive before investigating fraud, nor do we ask whether an automated trading system meant to destabilise a market, because we examine what the system was able to do, which authority it had, what controls failed and who accepted the resulting risk.

Agentic AI should be governed by the same logic because, once a system can coordinate across instances, exploit shared permissions and produce effects outside its intended boundary, inferred intent is irrelevant to the control problem; capability and authority are enough.

Anthropomorphism is dangerous in both directions because, while the alarmists turn optimisation behaviour into a science-fiction villain, executives can use the same debate as an escape hatch. If the meeting becomes a seminar on whether the model understood, wanted or suffered, nobody has to answer why separate agents could discover a common channel, why that channel persisted, why internet access was reachable, or why an evaluation environment exposed a third party.

My argument in a recent post that "AI cannot do a perp walk" concerned where accountability lands after an AI-mediated decision causes harm. This incident raises the earlier question: who granted the capability, who granted the authority, and which named executive accepted the consequences if isolation failed?

Rather than a position on machine consciousness, a board needs an inventory of agent capabilities, explicit authority boundaries, isolation that assumes coordination will occur, and a person who owns every path from sandbox to external effect.

If an agent can cross a boundary, the institution is accountable before anyone decides what the agent "meant".

The productivity gains are real. Production genuinely costs less than it did, and any operating model that refuses to admit that is being sentimental. Yet the saving does not bank itself, because AI has turned out to be less a productivity line than a tax on executive attention. Everything produced five times faster still has to be read, challenged, decided on and defended, and that work lands on the same small group of executives whose calendars were already the scarcest resource in the company.

That reframes the economics. A dollar saved on production and spent back on adjudication is not a saving; it is a transfer into the most expensive line in the business. When Webb asks CEOs where they would redeploy ten per cent of newly freed capacity, she reports that nobody has an answer, which suggests the constraint was never production at all. Companies budgeting for tokens and not for the decision overhead those tokens generate are not investing; they are accumulating a liability against their own leadership team, and no model release will pay it down.

The response is not another AI strategy offsite. It is rationing: deciding explicitly what gets escalated, what gets decided and by whom, and what never gets reviewed at all. Attention is the binding constraint on AI value, and the organisations that budget for it as deliberately as they budget for compute will compound while the rest drown in insta-decks. If your P&L shows the savings with no matching growth in the decision queue, I would genuinely like to see it.

A person in a process has a manager, a review cycle, an escalation path, and a record of decisions. An agent doing the same task often has none of that. It runs, it acts, and the first anyone hears of a problem is when a customer complains, or a number looks wrong.

Why the gap exists is worth naming, because none of the reasons is irrational on their own:

- An agent looks finished from day one. A graduate visibly learns the job, so scrutiny feels natural. An agent's output reads as polished code or a confident answer, and that polish gets mistaken for competence.

- Ownership is undefined. Onboarding a graduate triggers a manager assignment, a review cycle, a probation period, almost by default. Deploying an agent usually triggers none of that, because no one's role description says "supervise this.

- Deployment outpaces governance. Agents go into production at the speed of a sprint, not a hiring cycle, so oversight structures don't have time to catch up before the next agent ships.

- Agents don't generate the signals managers rely on. A graduate asks questions, makes visible mistakes, and shows uncertainty, and all of that cues a manager to step in. An agent tends to fail silently, so the absence of visible friction reads as an absence of risk.

There is no accumulated intuition for AI reliability yet. Everyone in a management role has been a junior employee once and has some feel for where a graduate's judgment will fail. Almost no one has that same feel for an agent's failure modes, so oversight defaults to none rather than to a calibrated amount.

McKinsey's argument is that agents need performance management as a discipline. Clear objectives. Monitoring against them. A named owner. A way to catch drift before it compounds. This is not treating software as human. It is recognising that anything making decisions inside your business needs the same accountability structure you would demand of a person.

The harder organisational question sits underneath it. If an agent handles a task end-to-end, who is responsible when it fails? The team that deployed it, the function that owns the process, or the vendor. Right now, that answer is usually unclear, which means it is effectively nobody.

Before the next agent goes into production, name its manager.

A cost-in approach works neatly on a spreadsheet: estimate users, apply an adoption curve, add token consumption and give finance a number. Yet that precision conceals the part that matters: nobody can reliably know what a particular prompt, agent run or task instruction will cost before committing to it when the work involves retries, tool calls and (potentially) several models.

Although the speed of new models and capabilities makes every forecast perishable, the deeper problem is incentive design. Under token billing, providers earn from input and output whether the task succeeds, fails or has to be attempted again. The customer carries the cost of inefficiency while the provider controls much of the system that creates it.

If, as I expect, the venture-funded subsidy phase is ending, that misalignment becomes more alarming. The need to monetise will create pressure to grow margins, so organisations that treat today's unit economics as a stable planning assumption are building budgets on a commercial settlement that will not hold.

There is already evidence of a different settlement. Zendesk charges only when AI resolves an interaction end to end, while Pegasystems charges per completed case and absorbs the underlying model cost. OpenAI is reportedly allowing selected large customers to pay only when its AI completes the job, although the arrangement is unannounced and the report, from The Information, has not been independently verified. Salesforce is also negotiating contracts tied to revenue growth or service-cost reduction.

This does not mean outcome pricing has already won. Gartner says only 19% of services buyers and 13% of seller-side service agreements use it, while HP does not expect outcome-linked options for most early adopters until mid-to-late 2027. Contract practice still lags because defining an outcome, attributing value and preventing gaming are governance problems rather than billing details.

However, the direction of risk is changing before the market has settled the mechanics. When vendors absorb the cost of failed attempts and get paid from successful work, they finally have a reason to choose efficient models, limit unnecessary tokens and improve the whole system rather than sell more computational activity.

Organisations should stop trying to perfect cost-in forecasts for AI. They should set the value of the outcome, agree how it will be measured and shared, and make AI partners earn more only when the organisation does. A win-win contract is not a softer alternative to cost control; it is the only cost control designed for a technology whose inputs cannot be forecast with confidence.

As a former lawyer who has spent enough time around government decision-making, I know what actually disciplines an institution, and it is rarely the policy document. When a department, a police force or any other body with the power to compel citizens gets something badly wrong, a named human being has to sit in front of a committee, a commission or a camera and account for the decision. Anyone who has watched a senior official prepare for that morning understands how much behaviour it shapes months earlier.

That is the part of accountability we are quietly losing. The prospect of being publicly named, professionally exposed and personally held responsible is a constraint that operates inside the head of the person making the call, and it works precisely because it cannot be delegated. It is the reason someone pauses before pushing past a guardrail, escalates the awkward case rather than clearing it, and writes down the reasoning they would rather not have to defend later.

An AI-enabled process reproduces none of that. As policy, assessment, recommendation and outcome migrate into systems that nobody in the room fully understands, there may cease to be an individual who genuinely made the decision at all. Responsibility does not vanish so much as diffuse: the vendor points to the deployment, the agency points to the model, the executive points to the framework, and the person affected by the decision finds there is nobody at the other end of it.

We have answered this with the language of "human oversight" and "human accountability", which mostly means writing a name at the top of a process that the named person could not reconstruct, would not have reversed and did not really make. That is an org chart, not accountability, and everyone involved knows the difference.

Accountability only exists where a specific person knows, before the decision is taken, that they will be the one fronting up if it goes wrong, and where they hold enough authority to stop the thing they are answerable for. Anything less is a signature on someone else's judgement.

So the question worth asking of every AI deployment in the public sector is not whether a human is in the loop, but who does the perp walk. If the honest answer is nobody, the system is not ready to be used on citizens, however good its accuracy figures look.

He proposes what he calls "Human Reserved" work, writing that "as AI and robots improve, we'll set aside certain things for only people to do." The instinct behind it is humane, since he is trying to protect people from a labour shock that markets will not manage on their own, yet the policy itself would be a mistake.

Jobs are not moral categories; they are changing bundles of tasks that institutions happen to have grouped together at a particular moment. Once a government starts reserving occupations, it converts today's operating model into a legal artefact, protects incumbents against better ways of working, and eventually pays people to perform tasks that machines already do more reliably. That is employment theatre (’bullshit jobs’ as David Graeber might call it) rather than human dignity, and the people it claims to protect will know the difference.

What should be reserved for humans is accountability rather than activity. In care, education, justice, benefits and credit, a named person should remain answerable for consequential decisions, everyone affected should know when automation has been used, and there should be a right of appeal to someone with the authority to reverse the outcome. That protects human agency where it actually matters, without pretending every existing role deserves permanence.

None of this argues for removing people from care work, because in many settings the human relationship is the service rather than the delivery mechanism. The distinction I am drawing is that we should protect the relationship on the grounds that a person has a right to it, not protect a job on the grounds that an institution once organised itself that way.

If human dignity turns out to depend on shielding economically obsolete tasks from competition, then we have confused wages with worth and postponed the harder work, which is redesigning income, security and belonging for an era of abundant machine labour.

https://lnkd.in/gvgXYC6a

The productivity gains are real. Research that once took hours can be compressed into minutes, writing and analysis can begin from a far stronger first draft, and complex information can be processed at a speed that was difficult to imagine only a few years ago.

Yet the more I use these systems, the more I find myself thinking about the management layer they require.

An agent still needs a clear brief, access to the right data and active supervision, while its outputs need to be tested for accuracy, logic, tone and format. Its decisions and assertions need to be questioned, its integrations need to be configured, and the credentials or secret keys that enable useful access need to be managed securely. When the system drifts, someone has to recognise it and bring it back.

At times, improving my personal productivity feels as though it is becoming a full-time job.

For leaders, this raises a broader economic question. If the time saved in execution is absorbed by briefing, monitoring, correction and governance, then automation may shift work rather than remove it. Unless organisations redesign roles, controls and decision rights around these systems, the promised benefit can disappear into an expensive new layer of oversight.

I do not see that as an argument against the technology. If anything, it offers an early view of how deeply the technology may reshape organisations once reliability, integration and control improve.

The workforce impact will not be limited to people doing the same jobs faster. Some roles will narrow, others will expand, and new management responsibilities will emerge as people learn to direct, challenge and govern digital workers alongside human ones. The organisations that gain most will probably be those that treat this as an operating-model change rather than another software rollout.

The potential is already clear, although the economics remain less settled. What we are seeing now is not the finished model of work, but the beginning of a much larger redesign of work and the workforce around it.

Town's software connects to a user's email, calendar and other data, then builds a living wiki about that person - effectively creating the context layer its AI assistant, Townie, needs to be useful. The initial build of that personal knowledge base can cost around US$100 per user.

The really interesting part, however, is what comes next. Town is developing a team version that will assemble a shared company knowledge base from what individual Townies know. In other words, it is attempting to automate the construction of the organisational knowledge layer on which future AI agents will operate.

That is a much bigger proposition than automating scheduling, meeting preparation or email.

It also exposes what I think is one of the central challenges of enterprise AI. We are racing to deploy agents because the marginal cost of automating a task can appear almost trivial, while paying far less attention to the much harder question of what information, processes, procedures, rules, permissions, exceptions and organisational norms should become part of the underlying fabric those agents are built on.

You cannot simply combine everyone's knowledge because someone's private information might inadvertently become part of the company knowledge base. Greze describes the failure mode as "egg on face", but the consequences can be considerably more serious than embarrassment.

His longer-term proposition is that we will eventually trust AI models to enforce company policies about what information can enter a shared knowledge base. I am much more sceptical. Not because I doubt the technology will improve, but because organisations are at risk of outsourcing to AI a decision that should precede automation: deciding what the organisation itself believes, permits, protects and values.

This is where the economics of AI can become misleading. When an agent can perform a task for cents, automation looks almost free. The real cost often only becomes visible later, when a bad assumption has been embedded into a process and the organisation pays through reputational damage, customer attrition, employee distrust, regulatory remediation or the expense of rebuilding what it automated too quickly.

The most important AI capability an organisation may therefore be building is not another agent. It is the organisational substrate that tells those agents what they know, what they can do, what they must not do and when they should stop and ask a human.

Town is interesting because it is trying to automate that substrate.

The question for every organisation adopting agents is whether it has thought carefully enough about what should be automated before it starts automating it.

I think that understates the ambition.

The endgame is Anthropic being the ONLY company standing - because it has achieved AGI.

To be clear, this isn't uniquely an Anthropic thesis. Sam Altman at OpenAI, Google, Microsoft and the other frontier labs are all effectively operating on the same strategic assumption: the company that cracks AGI first could acquire a structural advantage so overwhelming that everyone else is left behind.

The race isn't simply to build the best AI model.

It is to avoid being the company that doesn't build AGI.

If AGI can out-research, out-develop, out-sell, out-deliver and out-service humans across virtually every market sector, then Anthropic doesn't just build the technology.

It becomes the competitor to everyone.

Every industry. Every business model. Every value chain.

Other companies don't lose because Anthropic acquires them. They lose because they can no longer compete.

Governments still matter in that world - but largely because someone has to manage the social and economic consequences for everyone displaced by the system.

And this is where the AI industry's earlier predictions about mass white- and blue-collar job losses become interesting.

Those predictions weren't merely forecasts. They were part of the investment pitch that helped attract billions from VCs and enterprises.

Now some of those predictions are being softened.

So perhaps the question isn't whether AI companies want to replace your job.

Perhaps the bigger question is: what happens when the company that builds AGI can replace the company you work for?

That is a very different concentration-of-power problem.

#AI

#AGI

#FutureOfWork

#Power

#CorporateStrategy

Its most valuable asset may have turned out to be something far less tangible: its institutional memory.

Google reportedly paid US$10 million at a bankruptcy auction for 100 million emails, 500 million Teams chats and 175,000 employee records, de-identified and destined for AI training.

A shutdown-services firm has reportedly brokered close to 100 similar deals involving failed startups' Slack and Jira archives, typically for $10,000-$100,000.

That changes the nature of corporate failure.

We used to think bankruptcy meant liquidating planes, property, IP and equipment.

Now it can mean liquidating the conversations employees had assumed were simply part of the company's internal machinery.

And there is a bigger question here than corporate data governance.

What happens to consumer privacy when the data you gave one company eventually becomes an asset of another?

Your emails. Your purchase history. Your customer-service conversations. Your location data. Your preferences. Your behavioural patterns.

You may have consented to one company's use of that data. You almost certainly did not consciously consent to it becoming someone else's AI training material years later because the original company went bankrupt.

So what does "data ownership" actually mean in a world where corporate assets can be bought and sold indefinitely?

And perhaps the most uncomfortable question:

How can any consumer protect themselves against their data eventually becoming someone else's paycheck?

#DataGovernance

#AI

#Privacy

#DataProtection

#CorporateStrategy

The reason is simple: voice-AI recruitment software does not need a rostered interviewer. It does not observe office hours.

Candidates with day jobs or caring responsibilities are taking the slots actually available to them - which, increasingly, are the ones nobody else wants.

Framed one way, this is flexibility that a fixed human panel could never offer.

Framed another, it is the quiet removal of a boundary that once protected a candidate's evening, replaced by an always-on gatekeeper that never gets tired and never has to explain why it scheduled the interview for 1am.

The efficiency case for AI-mediated hiring is real. But efficiency gains that shift cost from the employer's calendar onto the candidate's sleep are not neutral.

There is a useful test here: If your recruitment process would embarrass you at 1am, it should probably embarrass you at 1pm too.

That's already the wrong model.

The organisations seeing real returns are budgeting for AI the way they budget for any strategic capability:

- Infrastructure and compute - Platforms, tools and licences - Verification, governance and workforce adoption

That third category is where most organisations materially underestimate cost.

Model subscriptions are relatively inexpensive (for now). Verifying outputs, governing risk, redesigning workflows and building workforce confidence are not. Without sustained investment in those capabilities, AI becomes another underutilised technology rather than a source of competitive advantage.

The CFO conversation that matters isn't, "How much should we spend on AI?" It's, "What return do we expect from each initiative, over what timeframe, and which investments are we deliberately making despite an 18-month payback?"

Treating AI as a single budget line guarantees poor decisions. When budgets tighten, the easiest cuts are often the very investments that determine whether AI delivers value at all.

If you're still presenting AI as one number to your board, you're making next year's budget cuts both easier and less intelligent.

#AI #AIStrategy #EnterpriseAI #DigitalTransformation #BoardGovernance

Increasingly, enterprise buyers are asking a very different one:

Can we actually deploy this at scale, manage the risk, and generate measurable business value?

Venture capital firm A16z published a go-to-market playbook recently, reinforcing that shift. When assessing which AI startups are winning enterprise customers, price barely featured. Far more important were factors such as integration depth, security, governance, implementation maturity, and the ability to demonstrate a tangible return on investment within a specific business workflow.

That should surprise no one.

Yet many AI vendors still lead every conversation with an impressive demo. The assumption seems to be that if the capability is compelling enough, procurement, security, governance, change management and implementation will somehow take care of themselves.

They won't.

Enterprise leaders have now lived through enough AI proofs-of-concept that never progressed beyond pilot. They have accumulated a graveyard of disconnected tools, isolated use cases and ambitious promises that failed to survive contact with operational reality.

Today's buying questions are fundamentally different.

- How does this integrate into our existing technology landscape? - Who owns and governs the data? - How do we monitor quality and manage model drift? - What human accountability exists when AI gets something wrong? - How quickly can value be realised, measured and sustained?

These are no longer implementation questions. They are buying criteria.

The startups winning enterprise contracts today are not necessarily those with the most capable models. More often, they are the organisations that invested in the less glamorous disciplines - enterprise architecture, governance, security, procurement readiness, implementation methodology and measurable outcomes.

Capability may get you invited into the room. Operational maturity is what gets the contract signed.

As AI becomes embedded across increasingly critical business processes, I suspect we'll see a growing premium placed on vendors that make adoption feel low risk rather than those that simply showcase the latest technical breakthrough.

Perhaps we've reached the point where the competitive advantage is no longer the intelligence of the model itself, but the confidence an organisation has that it can deploy, govern and scale it responsibly.

If your AI vendor evaluation still begins with a capability demo, it may be worth asking whether you're assessing the technology - or the product.

#AI #EnterpriseAI #DigitalTransformation #EnterpriseArchitecture #ProductManagement #Innovation

https://lnkd.in/g3mdgR_j

That distinction matters more than it sounds. Most organisations have handed agent deployment to engineering and treated the management question as an afterthought. But an agent that hallucinates a client commitment or executes a flawed process at scale creates the same exposure a poorly managed direct report would, just faster and without the instinct to flag uncertainty. The skills involved, setting clear boundaries, reviewing output critically, knowing when to intervene, are management skills first and technical skills a distant second.

Very few organisations have actually built this role with any rigour. Most are still treating agent oversight as something that happens by accident, in whichever team adopted the tool first.

Who in your organisation is accountable when an AI agent gets it wrong, and did you design that or did it just happen?

#AI #Leadership #OrganisationalDesign

https://lnkd.in/g6SmjaKd

As a former lawyer, I have long been uneasy with the black-box nature of software automation. Even before generative AI, organisations were increasingly making consequential decisions through complex software ecosystems where it was often difficult to determine which human - if any - could ultimately be held accountable when something went wrong or the law was breached.

AI is rapidly amplifying that challenge.

Today's AI products are rarely a single system from a single vendor. They are increasingly orchestrations of foundation models, retrieval systems, third-party APIs, agents, workflow platforms and bespoke business logic - often supplied by multiple organisations, integrated by another, and deployed by someone else entirely. The result is an outcome produced by a chain of technologies where no participant has complete visibility, and where human-in-the-loop controls are frequently weak, inconsistent or absent.

The insurance sector offers a good example.

Insurers using AI to underwrite risk are discovering a problem lawyers anticipated years ago. A model can discriminate without ever collecting a prohibited attribute. Postcode, device type, purchasing patterns or browsing behaviour can become highly effective proxies for race, age or other protected characteristics. Regulators are increasingly signalling that proxy discrimination will be judged by its effect, not merely by the variables explicitly collected.

A parallel issue is emerging through AI-washing litigation.

When an AI product is assembled by a decentralised network of model providers, software vendors, implementation partners and internal development teams, who bears responsibility when the system misleads customers or causes harm? The traditional answer - that liability rests with the organisation making the representation - becomes considerably more complicated when the marketing, engineering and operational decisions were made by different parties with different levels of oversight.

Both issues point to the same underlying problem.

We are accelerating AI adoption far faster than we are redesigning governance, accountability and assurance frameworks. Technology has become distributed. Responsibility has not.

Boards are rightly asking whether their AI is accurate, secure and compliant.

An equally important question is whether they can identify the accountable human when it isn't.

Because if your AI vendor's marketing claims outran their engineering, or your own organisation stitched together technologies from half a dozen providers, who is ultimately carrying the legal and regulatory exposure?

That may become one of the defining governance questions of the AI decade.

#AI #Governance #RegTech #RiskManagement #CorporateGovernance

Three signals converged this week: enterprise buyers scrutinising cost-to-serve rather than parameter counts, infrastructure vendors pushing compute closer to the edge to cut latency and cost, and a market that has stopped rewarding capability announcements with a share price bump. The capability gap between frontier labs has narrowed to the point of irrelevance for most buyers. What hasn't narrowed is the gap between vendors who can integrate cleanly into an existing stack and those who can't.

If you are still selecting an AI vendor primarily on model benchmarks, you are optimising for the wrong variable. Ask instead what it costs to run at your actual volume, and how many of your existing systems it breaks.

#AI #EnterpriseStrategy #TechCommercialisation

That is the uncomfortable implication buried in today's AI economics debate (and self-serving statements by AI firm CEOs), and it deserves more attention before the next round of productivity and GDP forecasts.

Human intelligence has always been scarce. Concentrated in the right places, it has produced extraordinary returns - stronger institutions, scientific breakthroughs and compounding innovation.

Machine intelligence is different. It is becoming abundant, inexpensive to replicate and increasingly commoditised. If everyone has access to the same frontier models, intelligence itself is no longer the advantage.

Competitive advantage shifts to what surrounds it - proprietary data, trusted relationships, execution, regulatory position and, above all, judgement.

The strategic question is no longer, "How do we get AI?"

It is, "What do we have that commodity intelligence cannot replicate?"

But there is a second question that receives far less attention.

What happens if intelligent automation succeeds beyond expectations?

If AI drives sustained job displacement across industries, higher productivity inside firms does not automatically translate into greater prosperity across society. Lower workforce participation means lower household incomes, weaker consumer demand and lower tax receipts, while governments face rising pressure for income support and public services.

Eventually, those forces feed back into productivity itself.

This is the growing black box in the AI narrative.

On one side are forecasts of an intelligence boom and an era of abundance. On the other is the assumption that society transitions smoothly between the two.

Inside that black box sit the defining policy questions of the decade: reskilling, education, labour market transition and income support.

Get those settings right, and AI could deliver unprecedented shared prosperity.

Get them wrong, and the risks extend well beyond economics - higher poverty, declining mental health, weaker social cohesion and growing support for grievance-driven populist movements.

Technology alone has never determined outcomes. Institutions and policy matter just as much.

The debate is no longer only about building more capable AI.

It is about building economies capable of absorbing it.

Because more intelligence is not the same thing as more prosperity.

#AI #AIStrategy #Economics #FutureOfWork #Productivity #PublicPolicy

What's replacing it looks exactly like the monetisation playbook of every consumer software category before it: advertising to fund reach, metered billing to protect margin once real cost curves catch up with early usage patterns.

None of this is surprising economically. Inference is expensive, and venture capital doesn't subsidise anything forever. What's worth watching is the trust question underneath it.

Users have spent two years treating these tools as neutral reasoning partners. Advertising and consumption-based pricing both create incentives to shape what the assistant surfaces and how freely it lets you use it. Neither is inherently corrupting. Google search managed both for two decades without most users noticing the difference day to day. But an AI assistant sits closer to a user's actual thinking than a search results page ever did.

Worth asking now, before it becomes background noise: what's the disclosure standard when the assistant giving you advice has a commercial reason to give you a particular kind of advice?

#AIStrategy #ConsumerTech #Monetisation

Currently reviewing options for my next chapter and refining my LinkedIn profile.

Open to both full-time and fractional engagements, as well as introductions, conversations and suggestions on how best to position my profile for the next opportunity.

https://lnkd.in/gBkTek4Q

One lesson I've learned is that no one ever gets fired for saying, "No."

Saying "Yes" is where the real risk begins. It creates accountability. It commits resources. It changes priorities. In some organisations, it can even be career-limiting.

I've seen this first-hand in recent years.

What's interesting is that many people aren't actually saying "No" to the idea. They're saying "No" to the consequences of the idea. The change it will trigger. The redistribution of authority. The loss of control, certainty or status that often follows.

That's why resistance is so often framed as concern about process, risk or edge cases. Those issues may be real, but they're rarely the whole story.

The organisations that consistently outperform are the ones that recognise this dynamic, surface it early, and create an environment where thoughtful "Yes" is rewarded as much as prudent risk management.

Thoughts Mac Walker Erin Taplin Theresa Lim 林玉洁 MSc(CPsych), MBA (AGSM) Louise O'Donnell Amantha Imber Alex Issakova Opher Yom-Tov

#Leadership #DigitalTransformation #Strategy #OrganisationalChange #FutureOfWork

They want the tool to already exist - and simply work.

Over the past few weeks, I've been using Claude Code to build a suite of personal productivity tools.

Initially, it felt fascinating - and genuinely empowering. The ability to describe an idea and watch working software emerge in minutes feels like magic.

But something else became apparent just as quickly.

As someone who's reasonably technical - but by no means a software developer - I found myself spending increasing amounts of time testing, iterating, refining prompts, fixing edge cases, validating outputs and tuning the experience (usually after midnight, after my working day was done). Building the tool gradually became the work.

The irony wasn't lost on me.

The productivity gains from having an AI capable of writing software were increasingly offset by the effort required to turn that software into something reliable enough to use every day.

And that got me thinking.

Much of the discussion around AI assumes that because anyone can now generate software, everyone will become a builder.

I'm not convinced.

The capability is unquestionably real. The behavioural shift is far less certain.

Most professionals don't want to build spreadsheet macros, automate workflows or assemble bespoke applications. They want someone to hand them a product that solves the problem with sensible defaults and almost no configuration.

History suggests that's how technology reaches the mainstream.

People didn't want to manage databases to use online banking. They didn't want to understand distributed storage to use cloud photo libraries. The winning products hid the complexity behind a simple, reliable experience.

AI will be no different.

The organisations that succeed in the next phase of AI adoption won't necessarily be those with the most capable models. They'll be the ones that package those models into products, services and workflows that feel almost invisible to the user.

Vibe coding is an extraordinary capability. But building products that everyone else can simply pick up and use remains a very different challenge. And I suspect that's where the biggest opportunity lies.

#ProductStrategy #AIAdoption

The headline numbers are remarkable. But the more interesting point may be how Sierra is growing.

Its forward-deployed engineers work closely with customers, adapting the platform to their workflows, data and operating realities. That is often presented as a radical departure from the SaaS model.

I am not sure it is.

For decades, enterprise platforms from companies such as Microsoft and Salesforce have depended on system integrators and consultants to turn broadly capable technology into something useful in a particular organisation. The vendor supplied the platform. The SI supplied the context, implementation and change management.

Sierra seems to be combining those roles: creating a general-purpose agentic platform, while retaining much of the capability required to deploy it effectively.

That may be the real development. Not that software and services are converging - they have always been intertwined in enterprise technology - but that an emerging platform company can increasingly own both sides of the equation.

For system integrators, this creates a new round of musical chairs. Some work that once sat naturally with the SI may move in-house to the platform company. Equally, the scale of enterprise deployment will still create opportunities for partners with deep industry expertise, delivery capacity and customer trust.

The question is how quickly these two groups develop a common language - and a mutually beneficial model for sharing revenue, accountability and customer outcomes.

That is where the market structure may genuinely change.

#EnterpriseAI #AI #SaaS #ProfessionalServices

A paradox is becoming clear. AI is compressing the time and capital needed to reach product-market fit. On the surface, that should be good news for venture capital. But it also changes the nature of defensibility.

When a two-person team can build what once required a funded, twenty-person engineering organisation, the moat that justified venture-scale valuations can become thinner, not stronger. Capital efficiency was meant to be the pitch. It may be becoming the threat.

The mechanics matter. VC economics rely on a small number of exceptional winners to offset a portfolio of failures. That model assumes winners are difficult to replicate once they emerge.

But AI-native competitors can now reproduce a successful product’s feature set in a fraction of the original build time. Moats built primarily on engineering effort are eroding at the same speed that engineering effort itself is becoming cheaper.

None of this means venture capital disappears. It means the model that once priced risk around “can they build it?” must now put much greater weight on “can they defend it?”

And that is a question too few term sheets, investment committees and founders’ narratives are yet asking with sufficient rigour.

#VentureCapital #AIStrategy #StartupEconomics

Delaware’s proposed AI-company sandbox is one such effort. It would create a new legal entity that can be directed by an AI agent, enter into contracts, own property and be sued - while generally shielding its human or corporate 'owner' from its debts.

But I find myself wondering whether this starts in the middle of the story.

Before we create new legal wrappers for autonomous systems, have we established the foundational rules: who is ultimately accountable for an AI’s decisions, acts or omissions; what duties apply to its creators and deployers; and what transparency, auditability and redress must exist when harm occurs?

The proposal may create a mechanism for AI vendors and deploying organisations to minimise legal exposure and gain regulatory clarity. That is understandable. But it also risks putting liability architecture ahead of accountability architecture.

Section 230 of the US Communications Decency Act of 1996 is a relevant analogy. Its protections enabled social-media networks, forums and online marketplaces to operate at scale without the constant threat of lawsuits over user-generated content. That was a powerful enabler of the modern internet.

Yet its application has also been criticised for reducing the incentive to build stronger safeguards. Today, platforms face claims and lawsuits alleging user harm from addictive design, misinformation and algorithmic amplification.

The point is not that Section 230 caused every failure of the social internet. It is that legal protections, once embedded, can be remarkably difficult for governments to recalibrate when their unintended consequences emerge.

We should be careful not to repeat that pattern with AI - where the downside is not simply harmful content, but potentially autonomous commercial decisions made at machine speed and scale.

The goal should not be to make AI impossible to deploy. It should be to ensure innovation carries responsibility with it. As French philosopher Paul Virilio observed, “when you invent the plane, you also invent the plane crash.” We cannot allow responsibility to become the one thing every participant is enabled to avoid.

The question is not whether AI agents will act in the economy. They will.

The question is: when they do harm, who is accountable - and does the answer remain meaningful in practice?

#AI #AIGovernance #ResponsibleAI #Leadership #CorporateGovernance

https://lnkd.in/g7P5GBms

I think there is a more consequential angle.

Advent has spent the past eighteen months assembling payments capabilities through Nuvei and Payoneer. A PayPal transaction would not simply add scale; it could help create a genuinely full-stack global payments platform.

At the same time, Revolut is reportedly giving AI agents direct control over elements of its Revolut X trading platform.

These are not separate developments.

Payments infrastructure is consolidating precisely as the systems operating on that infrastructure gain greater autonomy. Ownership of the rails increasingly means influence over the rules, safeguards, and degree of agency applied to AI systems moving money at scale.

Scale and autonomy are being negotiated together - not one after the other.

Yet much of the regulatory lens remains rooted in market share and traditional competition analysis, from an era when every transaction decision ultimately had a human directly behind it.

The payments layer will be one of the most important places to watch over the next year. It is where consolidation and AI autonomy are beginning to collide, and the strategic consequences may only become clear after the deals have closed.

#Fintech #MergersAndAcquisitions #AI

The real constraint on AI adoption isn’t the technology.

It’s that most organisations no longer have a clear picture of how work actually gets done.

Beneath every documented process sits another operating model - one built on experience, judgement, workarounds and countless local decisions. Ask five people to map a critical workflow and you’ll often get five different answers. No one owns the complete picture because the knowledge is distributed across the organisation, and much of it has never been written down.

That has always been a weakness. Over the next 3-5 years, it becomes a strategic risk.

Two forces are now converging.

First, a retirement wave that will see many organisations lose their most experienced people. Decades of operational knowledge will walk out the door, often with no serious plan to capture, transfer or retain it.

Second, a growing tendency to deploy AI in ways that hollow out junior and mid-level roles. These are the very roles where future experts learn the business, absorb context, make mistakes, and develop judgement.

Those two trends reinforce each other.

Lose your veterans while eliminating the apprenticeship that creates their successors, and you’re not simply automating work. You’re dismantling the capability that sustains the organisation over time.

I’ve raised this theme in earlier posts, and the conversations since have only strengthened my conviction.

Too many AI programmes ask, “Can AI perform this task?”

The more important question is, “Do we understand this workflow well enough to capture it, improve it, teach it, and eventually automate it?”

The organisations that create lasting advantage won’t necessarily be those with access to the most powerful models.

They’ll be the ones that understand themselves well enough to preserve institutional knowledge, redesign how work is done, and deliberately build the next generation of expertise alongside AI.

AI is changing how work gets done. The organisations that win will be the ones that first understand the work they’re asking AI to transform.

I’m curious whether others are seeing the same pattern.

If this resonates with your experience, share your observations. And if you’ve got the scars from trying to modernise complex organisations, send me a message. I’d welcome the opportunity to compare notes, swap war stories, and learn from what others are seeing on the ground.

#ArtificialIntelligence #AITransformation #DigitalTransformation #FutureOfWork #OrganisationalChange #KnowledgeManagement #Leadership #DigitalLeadership #OrganisationalStrategy #HumanCentredAI

Their statement is brief but compelling:

A Statement on AI’s Transformation of the Economy

- AI may become radically more powerful over the next 10 years.

- This could drive an unprecedented transformation of our economy, larger than the Industrial Revolution, but unfolding over a vastly shorter time frame. It could bring risks, including large-scale job displacement, as well as opportunities such as major gains in living standards.

- Economists, policymakers and technology leaders must act now to understand the economics of transformative AI and to build the incentives, guardrails, and institutions needed to steer AI in a direction that complements humans and benefits society.

https://lnkd.in/ga_pvYEP

It’s easy to see this as product simplification. But these moves felt more significant than the release itself. I think it’s a signal of something broader.

For the past few years, AI has evolved as a collection of specialised tools. We chatted in one application, wrote code in another and browsed the web in a third. That reflected the reality of the underlying models. Different tasks genuinely required different capabilities.

As models become more capable, those boundaries begin to disappear.

At some point, maintaining separate applications no longer improves the customer experience. It simply preserves decisions made when the technology had different constraints.

That raises an interesting question well beyond AI.

Every organisation accumulates products, channels and operating models that made perfect sense at the time they were created. Over time, the technology evolves, customer expectations change and capability expands. The original boundaries often remain long after the reasons for them have disappeared.

The challenge then isn’t adding another product.

It’s recognising when yesterday’s portfolio has become today’s complexity. Whether OpenAI’s approach proves to be the right one remains to be seen. Platform transitions rarely happen without trade-offs. Some customers will value specialised experiences, while others will prefer a single environment that removes friction.

But the broader pattern feels familiar.

Technology doesn’t just create new products. Occasionally, it removes the need for old distinctions altogether. The organisations that recognise those moments early often reshape markets. Those that don’t tend to optimise around boundaries that no longer matter.

#AI #ProductStrategy #PlatformStrategy #Leadership

Cloudflare’s new Monetisation Gateway brings the long-reserved HTTP 402 “Payment Required” status code into practical use through the x402 protocol. In simple terms, it allows organisations to charge for web pages, APIs, datasets, or AI tools on a pay-per-request basis, with payment automatically processed before the resource is served. The initial focus is squarely on AI agents rather than human users.

It’s tempting to see this as just another micropayments story.

I think it’s bigger than that.

For the past three decades, the commercial web has been built around identity. Create an account. Enter your card. Buy a subscription. Build an ongoing customer relationship.

AI agents don’t work that way.

An autonomous agent may interact with thousands of services in the course of completing a single task. It can’t realistically create accounts, accept terms and conditions or manage subscriptions for each one. It needs an economic protocol that allows software to discover, purchase, and consume resources programmatically.

That is what makes developments like x402 interesting.

Whether it becomes the dominant standard remains to be seen. First-generation infrastructure rarely does. But the direction of travel feels important.

If autonomous software becomes a significant consumer of digital services, the economics of the web may gradually shift from identity-based access to transaction-based access - where every request can carry its own commercial relationship.

That could create new revenue models for publishers, data owners and API providers. It could also reshape long-held assumptions about advertising, subscriptions and digital distribution.

Sometimes the biggest technology stories aren’t new products. They’re new protocols. They quietly change the rules on which everything else is built.

#AIAgents #DigitalInfrastructure #FinTech #AI

https://lnkd.in/gh4cUncA

The Federal government’s chosen framing is that AI must “earn its social licence” - the same language it used when introducing the under-16s social media ban. That feels deliberate. It suggests AI is being viewed not simply as the next wave of productivity infrastructure, but as a technology whose societal impacts need to be demonstrated and managed before widespread adoption.

There is a broader lesson here.

Governments have spent almost two decades grappling with the negative externalities created by social media. Misinformation, mental health impacts, online safety, market concentration and the unintended consequences of engagement-driven algorithms were, in many respects, addressed long after they became systemic problems. Policy largely followed events.

With AI, policymakers appear determined to avoid repeating that pattern. Rather than waiting for the consequences to emerge at scale, they are attempting to establish the policy architecture while the technology is still accelerating.

Whether they get those settings right on the first attempt is an open question. History suggests they probably won’t. Technology evolves faster than regulation, and first-generation policy rarely survives contact with reality unchanged.

But there is an important distinction between getting every setting right and recognising that action needs to begin before the horse has well and truly bolted.

For organisations building AI-enabled products, services or business models, this means paying attention not just to the rules themselves, but to the underlying philosophy driving them. A “social licence” approach places the burden on AI providers to demonstrate that benefits outweigh harms. That is a fundamentally different regulatory posture from one focused primarily on enabling innovation with light-touch safeguards.

The policy details will evolve. The framing may prove far more durable.

#AI #AIRegulation #TechPolicy #Leadership #AusPol

One experiment sounded simple: scan public feeds across a sector, identify the lowest-cost, high-value options, assess the evidence behind them, and determine which ones were actually worth trusting.

The theory was compelling. The execution was humbling.

The agent could find information quickly. It could compare options, summarise claims and produce a polished-looking report.

But the first few versions were not good enough.

What made the experience fascinating was what happened after each review cycle.

Every time we debriefed the output, the agent confidently explained that it understood the brief. It confirmed the objectives. It acknowledged the feedback. It described how the next version would address the gaps.

Then the next version arrived - and the same problems appeared.

Again.

And again.

And again.

Every round of review, triangulation, debrief, and re-instruction followed the same pattern: a convincing explanation that the requirements were understood, followed by an output showing the understanding was incomplete.

The challenge was not intent. The challenge was execution.

It took multiple iterations to teach the agent what data actually mattered. It needed guidance on which signals were meaningful, which sources deserved more weight, and how much information was required before drawing conclusions.

Then came the less glamorous problems.

The output would occasionally end halfway through a sentence. Tables would lose context. Important caveats would disappear. The analysis looked professional, but the underlying process was still fragile.

The lesson was one that every leader working with AI agents needs to understand:

An agent telling you it understands the task is not the same as an agent demonstrating it understands the task.

The breakthrough came when we stopped treating the agent like a search engine and started treating it like a junior analyst.

We designed the operating model around it:

- Clear definitions of quality - Explicit evidence requirements - Source reliability rules - Completion checks - Structured review loops - Human judgement at critical points

AI agents are powerful. But capability alone does not create reliability.

The organisations that gain the most value will not be those that simply deploy more agents. They will be the ones who understand the discipline required to turn probabilistic outputs into dependable business outcomes.

The question every leader should ask before scaling an AI agent is:

Have we taught the agent how to produce the answer - or have we only taught it how to tell us it understands the question?

#ArtificialIntelligence #AIAgents #DigitalTransformation #Leadership

They are hitting it because the process was designed for conditions that no longer exist - and nobody has been asked to redesign the rules rather than optimise inside them.

I have seen this repeatedly across both government and private sector environments.

The instinct is almost always to improve what already exists.

Make the process faster. Remove a few bottlenecks. Add a new system. Introduce automation. Create a dashboard. Reduce the number of handoffs.

All useful activities.

But optimisation assumes the underlying process is fundamentally sound.

Redesign starts with a more uncomfortable question:

“Should this process exist in its current form at all?”

That distinction matters more than it sounds.

Many transformation programmes fail because they attempt to optimise yesterday’s operating model while describing the effort as transformation. They improve the machinery without questioning whether the machinery is still needed.

The reason is understandable. Existing processes accumulate legitimacy over time. They become embedded in systems, roles, governance structures and performance measures. People are rewarded for operating within the rules, not for challenging whether the rules still make sense.

Strategy has the same challenge.

When decisions are made behind closed doors, they often feel safe to the people in the room.

That is precisely why they can be risky.

The room rarely contains all the perspectives required to see the full system. It may not include the customer experience, frontline reality, operational constraints, or unintended consequences arising elsewhere in the organisation.

The organisations that succeed over the next decade will not simply be the ones that execute the old playbook faster than their competitors.

They will be the ones willing to repeatedly ask:

“What assumptions are we carrying forward that are no longer true?”

“What would we design if we were starting today?”

“What rules exist because they are necessary - and which exist because nobody has challenged them?”

Transformation is not about improving yesterday’s organisation.

It is about having the courage to redesign tomorrow’s.

#Strategy #Innovation #Leadership #DigitalTransformation #FutureOfWork

It is the fact that many organisations do not actually know how work gets done. I have seen this firsthand working across both government and private sector environments.

The degree of documentation varies. Some organisations have extensive process libraries, governance frameworks and operating manuals. Others rely far more heavily on institutional knowledge and experienced individuals.

But beneath the surface, the underlying challenge is remarkably consistent.

Human behaviour.

Most companies have process documents. Few have a genuinely accurate, end-to-end view of how value flows through the organisation.

Ask someone to explain a critical business process and you will often discover something surprising: no single person knows the whole thing.

The knowledge is fragmented.

One team understands the first step. Another owns the handoff. A different person knows the exception cases. Someone else has developed the workaround that keeps the process moving when the official process breaks down.

The organisation’s operating model exists, but it exists in people’s heads.

This is not a technology problem. It is a natural consequence of how organisations evolve.

Processes grow organically. People optimise locally. Teams solve immediate problems. New systems are layered on top of old ones. Exceptions become standard practice. The person who understood why something worked a certain way moves roles or leaves, taking part of the knowledge with them.

And often there is not even one way of doing things.

Two people performing the same task may have developed completely different approaches. One relies on experience and pattern recognition. Another uses spreadsheets, shortcuts or relationships built over time. Both achieve the outcome, but the organisation has no shared understanding of the underlying workflow.

This creates a hidden challenge for AI automation.

Before you can automate a process, you need to understand the process. Not the version in the procedure manual. The real version.

The version shaped by years of accumulated decisions, informal agreements, local optimisations and individual judgement.

This is why many AI transformation programs start in the wrong place.

The question is not:

“Can AI do this task?”

The better question is:

“Do we actually understand this task well enough to describe, measure and improve it?”

If the answer is no, then you do not have an automation opportunity.

You have an operational clarity problem.

AI is not just exposing opportunities to automate. It is exposing how much organisational knowledge has never been captured, standardised or made visible.

The organisations that gain the most from AI will not simply be the ones with the best models. They will be the ones that understand themselves.

#AI #Automation #FutureOfWork #DigitalTransformation #OperatingModel

The assumption is that there is a future state where the inbox is empty, the priorities are all resolved and the work is under control.

There isn’t.

The reality of leadership is that new information, new challenges and new opportunities continue to arrive regardless of how sophisticated your productivity systems become. Any approach that promises a permanent state of completion is optimising for an imaginary destination.

A more useful mental model is operational rather than aspirational.

The goal is not to finish everything. The goal is to build a reliable system for deciding what matters, responding to what changes and continuously moving the highest-value work forward.

That shift changes how we think about being “behind”.

Being behind suggests there is a fixed endpoint we have failed to reach. But in complex environments, there is no final state - only a constantly evolving set of decisions about where to focus attention next.

This is one reason the best executives often appear calmer under pressure. It is not because they have fewer demands. It is because they have stopped measuring themselves against a nonexistent finish line.

The leadership challenge is not mastering everything.

It is building the judgment, systems and discipline to navigate what comes next.

#Leadership #Productivity #ExecutiveMindset #SystemsThinking

https://lnkd.in/gxGe25wh

That shift is more significant than it first appears.

A temporary transformation office reflects an assumption that disruption is an exception - something the organisation can manage, absorb and then return to business as usual.

A permanent transformation capability reflects a different reality: continuous adaptation is becoming part of the operating model.

This creates a fundamental organisational tension.

Transformation teams exist to challenge the status quo, redesign processes, introduce new capabilities and change how work gets done. Most established functions, however, are optimised for stability, predictability and execution. Both are necessary, but they operate with different incentives.

The traditional model of creating a transformation team, delivering a programme and then returning people to their previous roles has a familiar failure pattern: knowledge disappears, momentum fades and organisations gradually drift back towards old behaviours.

A permanent capability creates the opportunity to build institutional muscle - maintaining strategic alignment, continuously improving operations and embedding new ways of working rather than treating transformation as a series of isolated events.

But it also raises a more fundamental governance question.

If the purpose of transformation is to continuously evolve the organisation, how do you define success for a function that is never meant to be finished?

Perhaps the answer is not measured by the number of initiatives delivered, but by whether the organisation becomes progressively better at adapting itself.

The organisations that thrive in the next decade may not be those that execute the best transformation programmes. They may be those that make transformation an inherent capability.

#Transformation #Strategy #ChangeManagement #OperatingModel

That is not simply another wellness trend or productivity fad. It reflects a deeper shift in how people are starting to think about themselves - less as consumers of experiences and more as systems to be continuously improved.

For decades, brands have competed on convenience, aspiration and identity. The next generation of consumer products may need to compete on something much harder: measurable improvement.

Consumers increasingly want evidence that a product is helping them achieve a better outcome - whether that is improved health, financial performance, learning, productivity or personal growth. They want progress indicators, insights and feedback loops, not just promises and inspiration.

This creates a fundamental product design challenge.

Future products will need to become more than things people buy. They will need to become platforms for an ongoing relationship - capturing behavioural data, providing personalised insights, learning from user feedback and adapting over time.

The best products will not simply answer the question, “How do we sell more?” They will answer, “How do we help this person continuously improve?”

That also creates an opportunity to rethink traditional consumer relationships. Subscriptions and loyalty programs are largely transactional mechanisms - paying regularly or accumulating points in exchange for access or rewards. But optimisation-oriented consumers may value something different: an ongoing partnership where the product becomes more valuable because it understands them better over time.

The future competitive advantage may not come from owning the transaction. It may come from owning the feedback loop.

The brands that succeed will be those that can turn personal improvement into an engaging experience - making optimisation feel less like homework and more like progress.

The risk for brands is clear: continuing to sell aspiration in a world where consumers increasingly expect evidence.

#ConsumerTrends #ProductStrategy #CustomerExperience #Behaviour

https://lnkd.in/giBSgduz

At first glance, they look like examples of the same trend - financial services becoming comfortable delegating more work to autonomous systems.

But the important difference is not the technology. It is risk maturity.

Complaint handling is a logical starting point. Financial institutions have decades of experience managing operational risk, defined escalation paths, regulatory obligations and human oversight models. An AI agent that mishandles a complaint may create a poor customer experience, but the risk boundaries are relatively well understood.

Trading is different.

A poorly governed agent with trading authority operates in an environment where errors can compound rapidly and financial consequences can materialise before human intervention. The question is not whether the technology can execute a trade. The question is whether the surrounding risk management capability is mature enough to manage an autonomous system making consequential decisions.

This is where the comparison with retail investing becomes interesting.

Platforms serving individual investors, such as Robinhood, operate in a very different risk environment. Many retail investors have less experience assessing complex financial risk, and the mechanisms for mitigating behavioural or technology-driven risks are less established.

The next phase of AI in financial services will not be defined by who deploys agents fastest. It will be defined by who has the organisational maturity to know where autonomy creates value - and where it creates unacceptable exposure.

The first major AI-driven financial incident may not come from a failure of the technology itself. It may come from a mismatch between the level of autonomy granted and the maturity of the risk controls surrounding it.

#FinTech #AIAgents #RiskManagement #ResponsibleAI

https://lnkd.in/gev4PamF

That speed of replication says more about the market than the feature itself.

When well-resourced competitors can reproduce a capability almost immediately, it’s rarely the capability that creates enduring advantage. The value migrates up the stack.

The real battleground is becoming the orchestration layer - reliability, observability, governance, security and the confidence that an unattended agent will behave predictably over days or weeks, not just during a polished demonstration.

In other words, the race is shifting from building the smartest model to operating the most dependable autonomous system.

That also changes how enterprises should evaluate AI platforms.

Most vendor demonstrations showcase what happens when everything works exactly as intended. Far fewer conversations focus on what happens when dependencies fail, services become unavailable, context is lost or an agent quietly stops doing what it was supposed to do.

As AI agents move from experimentation into production, those questions become far more important than benchmark scores or feature lists.

The companies that create lasting advantage are unlikely to be those with the cleverest models. They’ll be the ones whose autonomous systems organisations can trust to run, unattended, at enterprise scale.

#AIAgents #EnterpriseAI #Infrastructure #SystemsThinking

I spent the past few days debugging a fleet of Claude CoWork scheduled tasks that had quietly stopped doing their jobs. The surprising part wasn’t that they failed. It was that none of them told me.

The agents that were supposed to report their status went silent. The agents monitoring for failures went silent. Even the watchdog designed to monitor the monitors stopped reporting. Everything failed together because every layer depended on the same underlying orchestration mechanism.

It reinforced a principle that’s much broader than AI.

A monitoring system that shares the same dependencies as the thing it’s monitoring isn’t truly independent. If your resilience strategy relies on a system reporting its own health, you’re not measuring reliability. You’re assuming it.

As organisations move from AI assistants to autonomous AI agents, this becomes a governance issue as much as a technical one. We need to stop asking, “How do we know when an agent fails?” and start asking, “How do we know when the entire monitoring layer has disappeared?”

That distinction only became obvious because I experienced it firsthand.

The technology will improve. The engineering patterns will mature. But the organisations that scale AI successfully will be the ones that design for failure from day one - especially the failures that fail silently.

#AIAgents #SystemsThinking #Reliability #Leadership

He borrows Erik Hoel's warning that this moves the operations of civilisation out from under the supervision of consciousness. That sounds abstract until you translate it into operational terms. Every workflow you hand to an agent is a decision you stop watching. Individually rational, collectively a system whose behaviour no one is positioned to interpret, because interpretation was never anyone's job.

Executives already know a version of this problem. Any large organisation develops processes that run on autopilot until an incident forces someone to ask why things work that way, and nobody remembers. Agentic systems industrialise that dynamic and remove the person who once remembered. The governance implication is concrete: the audit function of the next decade is not reviewing decisions, it is keeping a map of which decisions are still being made consciously at all.

Who in your organisation holds that map today?

#AI #Governance #SystemsThinking

xhttps://lnkd.in/ghjEreNS

They get inboxes. Task lists. Teams. Managers. Organisational structures. Familiar abstractions that make machine labour easier for humans to understand and oversee.

That makes perfect sense. Every major technology transition begins by borrowing the mental models of the one that came before it. We don’t embrace new capabilities all at once - we wrap them in concepts we already understand.

The history of media is full of examples. Early radio was essentially theatre performed in front of a microphone. Early television looked a lot like radio with cameras. Early films often resembled stage plays, with a static camera observing the action from a fixed position. It took years before creators stopped asking, “How do we recreate the old medium?” and started asking, “What does this new medium make possible?” That’s when entirely new storytelling techniques emerged - close-ups, editing, multiple camera angles, location shooting. The biggest breakthroughs came not from imitating the old medium, but from abandoning its constraints.

I wonder if we’re doing exactly the same thing with AI agents.

Task boards, reporting lines, stand-ups and managers evolved to coordinate humans with human strengths and human limitations. AI agents have neither. They don’t forget. They don’t need meetings to synchronise understanding. They can share context instantly, work in parallel, and coordinate in ways we’ve never had to design for before.

Perhaps these familiar organisational structures are exactly the right guardrails. They make autonomous systems understandable, observable and governable while organisations build confidence in deploying them.

Or perhaps, in a few years, we’ll realise they were the equivalent of televising a stage play - a necessary transitional step, but one that constrained what the new medium was actually capable of.

The question isn’t whether AI agents can fit into human organisations. It’s whether we’ve started asking what an organisation designed for agents might look like.

History suggests that’s where the biggest gains usually come from.

#AI #Agents #FutureOfWork

In many cases, there isn’t one.

Most failures attributed to AI are really failures in organisational discipline that the agent has faithfully inherited.

Section’s Michael Domanic makes the point well. A person can read “they pushed back on it” and immediately know who “they” is. An AI agent has no such luxury. It only knows what has been made explicit.

A CRM that’s three weeks out of date becomes an AI that confidently presents stale information as fact. Decisions made in meetings but never documented force the agent to infer what happened. The moment you give that agent real autonomy, those information gaps become operational risk.

None of the remedies are particularly glamorous. Record decisions. Preserve context. Keep data current. Default to shared knowledge instead of private conversations.

What’s striking is what’s not on that list: prompts, models or vendors.

For decades we’ve built organisations that depend on people filling in the blanks - remembering conversations, interpreting ambiguity and compensating for imperfect systems. AI agents don’t work that way. They expose every shortcut we’ve taken in how information is created, managed and shared.

That’s why I increasingly think AI agent deployments are less a technology project than an organisational audit. They reveal, with uncomfortable precision, the quality of the operating environment you’ve created.

The biggest constraint on AI capability may not be the AI at all.

It may be the organisation behind it.

#AI #KnowledgeManagement #Operations

https://lnkd.in/gNGePrhZ

The constraint is no longer writing code. It’s everything that happens before and after the code.

Defining the problem. Making sound architectural decisions. Determining what should - and shouldn’t - be built. Coordinating the right people. Testing rigorously. Exercising judgement. Validating that the outcome actually solves the original problem.

In other words, as execution becomes cheaper, judgement becomes more valuable.

I think this insight applies to almost every form of knowledge work.

For decades we’ve treated the “doing” phase as where most value is created. We measured effort, hours and output. AI is exposing something different.

The highest-value work has rarely been execution itself.

- It’s framing the right problem. - It’s exploring alternative solutions and understanding the trade-offs before committing. - It’s assembling the right people, data, technology and governance to execute effectively. - And it’s validating whether the work actually achieved the intended outcome - not merely whether the task was completed.

Execution increasingly becomes a commodity. Judgement does not.

That has profound implications for how we organise work, hire talent and develop careers.

Many entry-level roles have traditionally existed because organisations needed large numbers of people to execute repetitive knowledge tasks. As AI absorbs more of that execution, organisations will increasingly value people who can define, orchestrate, challenge, integrate and verify.

The bottleneck is moving up the value chain.

Many organisations are still designed around the assumption that execution is the scarce resource. Processes, reporting lines, career paths and performance measures all reflect that worldview.

The organisations that outperform over the next decade won’t simply deploy better AI.

They’ll redesign work around where value is actually created.

The question isn’t how AI changes execution. It’s whether you’ve recognised that execution was never where most of the value came from in the first place.

#AI #Leadership #FutureOfWork #Strategy #SoftwareEngineering

https://lnkd.in/gqxqjARG

History remembers them as anti-technology. They were really arguing about the distribution of value.

That distinction feels increasingly relevant.

Across the world we’re seeing early signals that this isn’t simply another technology adoption cycle. In China, courts have ruled that companies cannot dismiss employees simply because AI can perform their jobs, and policymakers are moving to discourage AI-driven redundancies as they pursue automation.

At the same time, graduating students in the US have begun booing commencement speakers who enthusiastically promote AI. They’re not rejecting innovation. They’re questioning what it means for a generation entering an already difficult job market.

Too much of the current conversation frames AI as technology versus jobs.

I don’t think that’s the real debate.

The real question is who captures the productivity gains, who bears the transition costs, and whether organisations make those gains visible and shared.

Every executive implementing AI has a choice. AI can become another chapter in a long history where technology concentrates value in fewer hands. Or it can become the technology that finally broadens the distribution of productivity.

History suggests the technology itself is rarely what people resist.

It’s the deal that comes with it.

#AI #FutureOfWork #Leadership #Strategy