AI adoption in a small team: a habit, not a purchase

A license changes nothing; an assistant inside one real task, owned by the person who does that task, changes the week.

The small company decides to adopt AI the way it decides to buy furniture: a budget line, a license for everyone, an announcement, and a wait. The license sits unused in half the seats, and the quarter ends with a subscription and no habit. Adoption was never a procurement decision; it is a working habit, and habits start in tasks, not in announcements.

The start is narrow on purpose: pick two or three real recurring tasks, the ones that repeat every week without anyone arranging them. First-draft replies to common questions, meeting notes turned into actions, quotes assembled from a price sheet. The task chooses the tool, because a tool chosen before a task is a solution looking for a problem, and everyone can feel it.

With the task, access; with access, one plain rule for what must never be pasted in: client names, contracts, credentials, anything a person would hesitate to pin on a noticeboard. The rule should be one sentence everyone can repeat, not a policy document nobody reads, and the reason should be said out loud once: what leaves the company cannot be recalled.

Then practice on live work with review: the assistant's first outputs are drafts, and a person who owns the task reads them before anything ships. The review is where the habit actually forms, because it teaches the team what the tool is good at, where it invents, and what finished looks like. Unreviewed output is not adoption; it is risk wearing productivity's clothes.

Spread happens by demonstration, not by memo: the person whose quotes now take an afternoon instead of two days shows the person at the next desk, and the habit walks across the team on its own legs. One converted task with a visible result recruits more users than any kickoff, because people adopt what they have seen working on work they recognize.

The failure mode is the opposite order: a company-wide license, a kickoff, no rules, and no named task. Within a quarter either nothing changed, which is the cheap failure, or someone pasted client data somewhere it should not have gone, which is the expensive one. Both grow from the same root: the tool arrived before the task, and the boundary did not arrive at all.

We run our own funnel this way: the assistants we built handle real recurring work in our company, and ours say plainly that they are AI. The honest limit is contractual: what each provider does with submitted data belongs to its documentation, read before the rule is written. Guidance, not guarantees: adoption is measured in tasks changed, not in seats bought.