How AI actually automates a workflow, and where it shouldn't

Every business is being told to "do something with AI". Most of what is on offer is a chat window and a hope. The useful version is quieter: a workflow that already exists, with a model doing the one step that needs judgement, and everything else handled by the boring parts nobody puts in a demo.

Here is what that looks like end to end, using a request that arrives by email.

The eight steps

1. A request arrives. An email, a form, a message in Teams, or an event from another system.
2. The model understands it. It classifies the request ("this is a change of address", "this is a document request") and extracts what matters: who, what, which account, by when.
3. Retrieval. The workflow fetches the information it needs, and only from sources the requester is entitled to see. This is where retrieval-augmented generation (RAG) belongs, and it is a permissions problem before it is an AI problem.
4. Business rules. Your policy, in code, decides what is allowed. A model does not decide whether a refund over a threshold needs a director; a rule does.
5. The system of record is updated. CRM, back office, ticketing, document store.
6. A task is created. Assigned, due, visible.
7. Human approval. Anything consequential waits for a person. The workflow shows them what it found and what it proposes; they approve, amend or reject.
8. Audit trail. Every prompt, retrieval, action and decision is logged to a store you own, with retention set by your policy.

Steps 3 to 8 are deterministic. The model only touches step 2, and occasionally drafts the text a person will send in step 7. That is the point: the model is a component, not the system.

What should not be automated with a model

• Decisions with legal or similarly significant effects on a person. Whether someone gets a loan, a job or a service. The law expects meaningful human involvement, and so should you.
• Anything you cannot evaluate. If you cannot write down what a correct answer looks like for a hundred real examples, you cannot tell whether the model is doing it right.
• Steps that were never a judgement call. Starters and leavers, document chasing, monthly reporting: these need rules and integrations, not a language model. Automating them with AI adds cost and a failure mode.

Where the value is

Not in the model. In the retrieval boundary (what may be seen), the rules (what may happen), the approval step (what may be executed) and the log (what can be proved afterwards). Get those right once and every workflow after the first is cheaper, because they share the same runtime.

That is also why the first workflow should be small: one process, one owner, one month, and a measurement of the time it gave back before you automate the next one.