An AI is not a work system

An AI can write a brilliant draft and still not have improved your practice.

It can save time on one task and add review to another. It can produce a correct answer without anyone being clear who should approve it. It can work with information that was not relevant to that assignment. Or it can leave the next step dependent on someone remembering it.

This happens because AI capability and work system are different things.

In recent weeks, we've talked about tools: how they fail, how to verify them, and what to ask a vendor. Now it's time for a question prior to any serious automation: what needs to exist around AI for the work to be repeatable, controllable, and defensible?

The Category Error

When a firm says «we have implemented AI,» sometimes it means they have bought a license, created some prompts, or allowed the team to experiment. All of that can be useful. None of those things, on their own, design how the work actually functions.

A system emerges when the process retains its logic even if the chat, person, or provider changes. Work should not vanish within a conversation: it should be clear what is being solved, who is next in line, what decision is pending, and what outcome is considered valid.

Start with the work, not the model

Consider a consultancy that receives recurring inquiries. The client's email initiates a process. Before asking what AI can do, it's essential to understand what information is indispensable, which decisions alter the path, what actions have an external effect, and where a competent person needs to intervene.

AI can help at various points. What it should not do is invent the organization on the fly.

Three public questions that already improve design

1. What is the observable outcome that this process should produce?
2. What decisions or exceptions cannot remain implicit?
3. What should be reconstructible afterward if something goes wrong?

Answering these three questions well prevents a common mistake: automating a conversation before understanding the work.

Autonomy isn't a yes or no question either

Within the same process, an AI might only prepare a proposal, while another low-impact operation could be executed under rules and supervision. There isn't a single answer to 'how much autonomy does the firm's AI have?' The decision belongs to each operation and must be proportional to its consequences.

The public rule is simple: the greater the impact and the more difficult it is to reverse or verify an action, the stronger the human and organizational control must be.

A serious system also knows when to stop

Real processes have incomplete information, contradictions, out-of-scope tasks, and situations where continuing automatically would be worse than stopping. Requesting data, blocking an action, or escalating an exception is not a failure: it's expected behavior from a well-governed system.

What changes in a small SME

None of this requires copying the bureaucracy of a large corporation. A small business needs the minimum structure to prevent critical work from depending on memory, improvisation, or implicit authority. Simplicity is not about eliminating governance. It's about making explicit only what changes a real decision.

A practical exercise

Choose a process you repeat and describe, in plain language, what outcome it should produce, what decisions matter, where it might get stuck, and what you would need to demonstrate if someone questioned the outcome. Only then ask which part deserves assistance or automation.