How should a small company run an AI training pilot?

How should a small company run an AI training pilot, in short?

Keep it small, short and decided in advance. Choose a handful of people with a shared, repeatable task, one approved tool, about four weeks and a named owner. Write down what evidence would justify continuing before you start, so the decision at the end is about the evidence rather than enthusiasm.

A small company has an advantage here: the person running the pilot can usually see the work directly. Use that. A pilot that produces a few reviewed work samples and a clear yes, no or change decision has done its job.

What are the steps of a small pilot?

Five steps cover most small pilots. Each is short, but skipping any of them makes the end decision harder.

How many people should be in the pilot?

Enough to see a pattern, few enough that one person can review every piece of work. For most small companies that is three to eight people. Below three, one person's good or bad month dominates; above eight, review time tends to slip and the pilot turns into an unsupervised rollout. These are our working limits, not a research finding.

What does a pilot look like in a real setting?

Maple Row Marketing, a fictional 18-person agency, piloted with five account managers who write weekly client updates. The rule was approved ChatGPT plan only, no client financials in prompts. The standard had three points: accurate to the supplied notes, clear next steps, no invented detail.

Baseline drafts were readable but two invented a deadline. After four weeks of short sessions, the retest on a new client update showed every draft flagging uncertain dates. The agency owner had agreed in advance that this, plus all five still using the routine in week four, would justify adding the design team. She expanded, and kept the same standard.

What are the common pilot mistakes?

The common ones are choosing a group with nothing in common, so no shared task exists; skipping the baseline, so there is nothing to compare; reusing the same task for the retest; and deciding the success criteria after seeing the results. Another is piloting with only enthusiasts, which tells you little about the rest of the team.

Vendor adoption guidance points the same way. Microsoft's Copilot adoption hub describes assigning licences to a first cohort after readiness checks and starting with scenarios tied to team goals, rather than a company-wide switch-on.

Where does Tosheo.ai fit?

Tosheo.ai is being built to fit a pilot of this size: a planned seven-day trial with up to five seats and a card required, a 5-minute diagnostic that acts as a baseline, rubric-graded tasks and a manager view of skill progress. The trial covers the first week, not the whole pilot, so budget for the paid weeks after it. All of this is planned launch scope.

When is Tosheo.ai not the answer?

If your organisation needs a procurement-led pilot with security reviews, SSO and contracts, Tosheo.ai is not set up for that at launch, and an enterprise provider may suit you better. If you have no approved tool yet, the pilot should start with tool and policy choices, which you can do without any training product.

Sources and scope

Microsoft Adoption, Microsoft Copilot adoption hub, checked 16 October 2026 ↗