AI Agents: Brilliant at Clear Tasks, But Not for DIY Builders

A new employee asks when something is unclear. An AI agent doesn't. It works with what it gets — and when in doubt, delivers convincingly sounding wrong results.

That's why AI agents already work excellently at some tasks — and fail spectacularly at others. Data reconciliation, document review, rule-based analysis: here the agent can verify its own output. Try, check, correct — in a loop, until it's right. No breaks, no careless mistakes, no motivation dip on Friday afternoon.

Where the result is verifiable, AI becomes reliable.

But between "an agent could do this" and "the agent does it reliably" lies a gap that most people underestimate. Specifying the task cleanly. Connecting to the right data sources. Handling error cases. Integrating the whole thing into existing workflows without breaking them. That requires people who understand both the technology and the business processes in detail.

The interesting part: the formats for this have existed for a long time. Product Requirements Documents, Scope-of-Engagement documents, process descriptions — every industry has its own. Ethan Mollick at Wharton Business School recently put it succinctly: all of these formats work surprisingly well as a foundation for agentic work.

The catch: most companies don't have them. Or they're not current. Or not precise enough for an agent to verify its own output. No room for interpretation, no "everyone just knows that" assumptions.

Most don't fail because of the AI — they fail because they've never properly documented their own processes.

For clearly specifiable tasks, the technology is ready. The question is whether the organisation is too.