The question we get asked most often is not "can AI do this?" It is "where should we start?" — and the answer determines whether the project ships in six weeks or dies in a steering committee.

Pick the process, not the technology

Teams tend to start from a capability: we want an AI agent. That is backwards. Start from a process that has four properties.

  • High volume. It happens dozens of times a day, not twice a quarter.
  • Rule-shaped but text-heavy. There is a right answer, but reaching it means reading something unstructured.
  • Tolerant of review. A human can check the output cheaply, at least at first.
  • Owned by one team. Cross-functional processes multiply the number of people who can veto a launch.

Invoice coding, support triage, and document intake keep appearing on that list for a reason. They are boring, frequent, and someone already knows what "correct" looks like.

Write down the current cost before you start

You cannot prove an automation worked if nobody measured what it replaced. Before any build, capture three numbers:

  1. Volume. Items per week.
  2. Handling time. Minutes per item, measured rather than estimated. Ask someone to time twenty of them.
  3. Error rate. How often the current manual process gets it wrong, and what that costs when it does.

That third number is the one teams skip, and it is usually where the return actually lives. A process that takes four minutes and is wrong two percent of the time may be far more expensive than one that takes twelve minutes and is never wrong.

The scoping checklist

QuestionWhy it matters
Where does the input arrive?An inbox, a form and an API are three different projects
Who owns "correct"?If nobody can adjudicate, you cannot evaluate
What happens on low confidence?A queue must exist before launch, not after
What is the rollback?Every automation needs an off switch someone trusts
Who watches it in month three?Unowned automations decay silently

If you cannot answer all five, the process is not ready — regardless of how good a fit it looks technically.

Start with assist, not autonomy

The fastest path to a working automation is one that drafts rather than decides.

Phase 1  Model drafts, human approves every item      → measures accuracy
Phase 2  Model auto-approves above a confidence bar   → measures coverage
Phase 3  Human reviews exceptions and samples only    → measures drift

Phase one feels like it defeats the purpose. It does not. It produces the labelled dataset you need to justify phase two, and it earns the trust that makes phase two politically possible. Teams that skip to phase three ship faster and get switched off faster.

A reasonable first scope

For a first project, aim for something you can describe in one sentence, that touches one system of record, and that a single person can validate. Something like:

Read incoming support email, classify it into one of nine existing categories, attach the three most relevant help-centre articles, and route it — with a human approving every routing decision for the first four weeks.

That is narrow enough to build in weeks and real enough to matter. The ambitious version of the same project — an agent that reads the email, answers it, issues the refund and updates the CRM — is the second project, and it will go far better for having the first one running underneath it.