Agentic AI tools like Claude CoWork and Clawbots promise autonomous productivity; the reality is rather less refined.
These tools are not ready for prime time. Here is why.
Technical friction
Agentic solutions are, in the main, built for technical users. Business users should not be required to configure API keys, access tokens, or gateways. If it requires manual "secrets" management to function, it is not a business-ready tool.
The black box problem
Logic remains opaque. While you can often choose the underlying LLM, you have little visibility into how the platform tasks that model. The result is significant inconsistency. Selecting the same LLM across different services yields highly inconsistent results for standard instructions.
Broken economics
Opaque credit systems replace predictable SaaS fees, creating budget uncertainty. Platforms require you to buy credits for "tasks," with zero transparency into how tasks are measured. Costs are unknowable before execution, and inconsistent credit-to-task ratios make cross-platform comparison nearly impossible.
The hidden maintenance tax
Refining an agent is a continuous cost sink - every guardrail, clarification, and fix burns credits, with no stable cost for performance. Worse, because these platforms are constantly tweaked by the provider, a working workflow can break overnight, forcing more time and credits to diagnose and fix issues that you did not create, only to have them break again.
The path to the early majority
The credit-for-tasks model may appeal to early adopters who are less price-sensitive, but it is poorly suited to the early majority. Broad adoption requires cost structures that are predictable, legible, and easy to budget for, which points to fixed-fee service models.
That transition is challenging if providers depend exclusively on high-cost frontier models. The path forward is increasingly clear: most mainstream business use cases do not demand cutting-edge capability. Consistent, high-quality outcomes can be delivered using efficient open models such as Qwen, Kimi, and DeepSeek, which materially lower the cost base while maintaining acceptable performance.
The innovator's dilemma
Innovators will soon realise that hosting these models themselves makes the most economic sense. By fixing their infrastructure costs and moving away from per-token pricing, they can offer 80 per cent of the functionality at 50 per cent (or less) of the price. This is where the market will be won: by locking in users who value predictability and cost-efficiency over raw speed or frontier power.
The invisible UI
The winning agentic interface is not a dashboard but an invisible intelligence layer that interprets intent, decomposes instructions and routes each component to a rightsized model.
This enables cost optimisation without degrading outcomes - the agentic form of “Keep It Simple, Stupid”: frontier models (perhaps) for complex reasoning; efficient hosted models for execution.