Thought 02 · AI implementation
Small, legible steps beat one giant AI black box
AI makes it tempting to push processes straight toward automation. Real businesses are messy enough that the shortcut is rarely as simple as the demo suggests.
With AI, it suddenly feels much more feasible to push processes straight toward automation. And that makes it very tempting to skip stages.
Many examples from experts ahead of the curve focus on agentic loops in software development, where the agent can see the codebase, read the documentation, run tests, and check its own work.
Real businesses are messier than that. Processes are spread across tools, data is incomplete, exceptions live in people’s heads, and not everything can be verified automatically.
The tempting shortcut
Give the agent all the context you can find, together with very specific instructions, and let it figure it out.
The downsides of this approach are:
- Expensive: advanced models need to run for a long time over a large context.
- Unreliable: it might work some of the time, but not every time.
- Unverifiable: it is hard to confirm that the final output is correct.
- Not auditable: when it fails, you don’t know which decision caused it.
What actually works
Break the process down into small steps. Use deterministic automations whenever possible. When AI is needed, provide only the relevant and curated context for that specific step.
The advantages are:
- Reliable and cheaper: each step is small and well-defined enough that even less capable models can complete it reliably.
- Auditable: when something goes wrong, you can inspect the decisions the AI took at every step.
- Verifiable: small, specific outputs are easier to check against rules, records and source material.
There’s a time and a place for everything, and I’m sure agentic loops will make sense for some use cases. But today, for most companies outside the tech industry, simple and reliable steps should still be the foundation.
The other thoughtA maturity framework for digitalization