The agent is rarely the first missing piece
When small teams say they want an AI agent, they usually mean something more practical: fewer repeated questions, faster drafts, cleaner handoffs, or a way to stop work falling between tools.
Those are workflow problems before they are agent problems. If the inputs, decision rules, and review points are unclear, an agent only moves the confusion faster.
A telltale sign is that no one can describe where the workflow currently breaks. Ask what happens when a request is unusual, and the answer is usually "someone deals with it" rather than a documented step. An agent added on top of that gap does not remove the gap — it just makes decisions faster without making them clearer.
Start where review is natural
The strongest first projects keep humans close to the work. Draft a reply, prepare a summary, classify a request, or highlight a missing field. Let the team see what changed and correct it before anything becomes automatic.
That is slower than the sales pitch. It is also how teams build confidence without handing judgement to a black box.
Drafting, summarising, classifying, and flagging are the four moves that fit this pattern well. Each produces something a person can accept, edit, or reject in seconds, so a wrong output costs a correction, not a customer.
Sales teams often find their narrow point of friction in follow-up drafting; support teams find it in first-response triage; operations teams find it in reconciling handoffs between two systems that do not talk to each other. The AI layer looks different in each case, but the review pattern is the same.
A worked example: sorting inbound requests
Take a shared inbox that mixes sales questions, support issues, and general enquiries. The instinct is to build an agent that reads each message and replies. A narrower first step is to have the model read the message and suggest a category and a one-line summary, then hand both to the person who already triages the inbox.
That person still decides what happens next, but they stop reading every message end to end to work out which pile it belongs in. After a few weeks you can look at how often the suggested category was accepted versus corrected, and use that number to decide whether the next step — drafting a reply for the highest-volume category — is worth building.
Compare that to jumping straight to an agent that replies automatically. If the category is wrong even one time in twenty, the team is now debugging automated replies sent to customers instead of glancing at a suggested label — a much more expensive way to discover the same gap.
When autonomy earns its place
None of this means agents are wrong for SMEs. It means autonomy is something a workflow earns after it has run as an assistive step for a while — after you can point to a track record of the model's suggestions, not just its plausibility on the day it was demoed.
A support-request classifier that has been correct nineteen times out of twenty over three months, with the remainder caught by the person reviewing it, is close to a case where letting the model act — reply directly to the easy majority, forward the rest — starts to make sense. The workflow earned the agent's autonomy; the agent did not have to be trusted upfront.
This is not a fixed timeline. Some workflows earn that trust in a few weeks of high volume; others take longer if requests are rarer or the cost of a mistake is higher. The pace is set by the evidence, not the calendar.
A better sequence
Map the workflow, choose one narrow point of friction, agree how success will be measured, then build the smallest useful assistive step. If that works, you can decide what deserves more automation.
Teams that skip straight to an agent tend to discover the missing decision rules only after something goes wrong in production, which is a more expensive place to learn them than a workflow map.
None of this is about moving slowly for its own sake. It is about spending the cheap failures first — a mislabelled email, a draft that needs editing — before risking the expensive ones a live, unsupervised agent can produce.