What to automate first with AI — and what never to

The first AI workflow a team builds is usually the most visible one, not the most valuable one. A clever demo gets shipped, saves nothing, and quietly teaches the whole company that the technology is a toy.

The better question is not "what can AI do". It is "which decision do we make forty times a week, the same way every time, with a text trail?"

Automate repeated judgement, not conversation

The sweet spot has three properties:

• It repeats. Daily or weekly, often enough that a 20% improvement shows up in the month.
• It has a rubric. You can describe what a good answer looks like in a paragraph, which means a model can be checked against it.
• It already has an audit trail. Emails, tickets, transcripts, invoices. Text in, text out.

Qualifying inbound leads, summarising calls into next steps, drafting the first pass of a proposal from a template, sorting support tickets by severity: all of that fits.

Keep these four human

1. The price. A model can draft the number; a person signs it.
2. The apology. Anything that touches a customer's disappointment gets a human voice, or it reads as contempt.
3. The hire. Screening is fine to accelerate. The decision to bring someone into a small company is not a sorting problem.
4. The promise. Commitments about delivery, compliance or money are made by someone accountable for them.

Build the smallest loop first

Do not start with an agent that has seven tools. Start with one input, one rubric, one output, and a human who reviews every result for two weeks.

Measure three things during those two weeks: how often the draft was usable unchanged, how long the human edit took, and how many failures were caught by the reviewer rather than by the customer. If the last number is above zero, you do not have a workflow yet. You have a risk.

The trap of invisible work

AI work is easy to over-report and hard to verify. A team can produce a hundred summaries a day and move nothing, because nobody wired the summary to a decision.

Tie every automation to a number that already exists: response time, close rate, hours per client, error rate. If no number moves, the automation is content, not leverage.

The first build, in one line

Pick the task your team complains about every Monday, write the rubric for a good answer, and put one model call behind it with a human approval step.

Then wait two weeks before building the second one. The teams that compound are the ones that ship one boring workflow that works.