Skip to content
Is there an AI for this?

Field note · Customer education

AI pilot plan for small business: a four-week buying test

Choose a first AI project, run a four-week trial, and decide whether to expand. A practical plan for owners who need useful work and clear costs.

Source
Editorial guidance with 1 retrieved source documents for context. Recommendations are our judgement.
Evidence
none on this page — it links to the pages that hold it

By Is There an AI for This? · Editorial guidance · Reviewed · How we use evidence

01Start with a queue of work you can count

Ask a colleague to show you the last ten examples of a repetitive task. Choose something with a visible beginning, a finished output and a person who can judge correctness: answering delivery questions, preparing invoice records or finding an approved policy. Write the task as a sentence: when this arrives, this person produces this output in this system.

Our suggested first-project filter is repeatability, available examples, reversible output and a willing owner. Score each candidate from zero to two on those four dimensions. Use the scores to expose missing information, then discuss them with the person doing the work. An impressive score cannot compensate for an owner who has no time to review the trial.

02Give each week a specific deliverable

  1. 01Week one: measure the current methodTime 30 representative tasks, including corrections and handoffs. Record volume, completion time and common mistakes. Ask the owner to write an expected answer for each sample. Reserve some examples for a final check that the supplier cannot rehearse.
  2. 02Week two: prepare one controlled workflowGive the trial only approved material and the minimum access it needs. Begin with drafts that a person checks. Agree the spending ceiling, the reviewer, the support contact and the exact action that switches the trial off.
  3. 03Week three: compare completed workAlternate comparable tasks between the current process and the assisted process. Include time spent reading the draft and correcting it. Record rejected outputs as failures rather than quietly removing them from the results.
  4. 04Week four: make the buying decisionReview the reserved examples and the live workload together. Produce a one-page decision with measured benefits, errors, total running cost, unresolved risks and the proposed next group of users. Extending the trial is a separate decision with a new deadline.

03Estimate value before negotiating the price

Keep capacity and cash separate. Fifteen hours may help the team clear a backlog, but payroll will not necessarily fall. Record what the freed time will be used for and who will check that it happened. If the intended benefit is shorter customer waiting time, measure that directly alongside staff time.

04Answer these questions before the first real task

  • What material is approved for the trial, and who can authorise adding another folder or mailbox? Start with the smallest useful collection.
  • Who checks each output, and what happens when that person is absent? Reserve their time in the plan instead of treating review as spare work.
  • What is the full trial budget, including setup, staff review, support and any usage charges? Require an alert before the spending limit is reached.
  • Where will the results be saved, and can the existing process resume without reconstructing missing work? Rehearse one failed or interrupted task.

05Agree what earns an expansion

  • RECOMMENDATION

    For a small business with no dedicated AI operator, we recommend expanding only after a named owner can run the workflow, explain the errors and recover from a failed task.

    BasisThis recommendation depends on limited internal operating capacity and the measured workflow above. Vendor privacy documentation is reading material for the data review, not proof that a particular trial is approved.

An illustrative acceptance gate is a 20% reduction in median handling time, no material increase in the agreed error categories, all critical errors caught before use, and successful recovery of an interrupted task. Choose the threshold before looking at the results. A small trial cannot establish that rare failures will never happen.

Stop if confidential information reaches an unauthorised person, a critical error bypasses review, costs exceed the approved ceiling or nobody owns the exception queue. Record what failed and return to the existing process. A smaller scope, better source material or a conventional form may be the next proposal. Buying more licences is not the default response.

06Related

Use cases

Stacks

More field notes

Ask this as a question

  1. 01We run a 35-person business and know little about AI. How do we choose one useful first project and judge a four-week trial?

07Evidence

1 record
  1. 01
    Tier 2official vendor legal documentationVendor legal and technical documentation

    Enterprise privacy at OpenAI | OpenAI

    https://openai.com/enterprise-privacy

    FreshRetrieved 19 Sep 2026sha256:925098213d77

    1 record

08Primary-source reading

Consult these primary documents when checking a proposed product. Reading links remain available when evidence has not been verified; they do not establish that your planned use is approved or that a recommendation has been proven.