An AI pilot should answer a business question, not simply demonstrate that a tool can produce an impressive response. Does it make an agreed HR task more dependable, faster or easier to explain after the necessary human checks are included? A ninety-day plan gives the organisation a manageable period in which to test that question.
Choose a bounded use case with a clear output and a safe fallback. Examples include drafting general training material or organising an already prepared, non-sensitive exception log. Avoid starting with automated hiring, pay or disciplinary decisions, where the effect of a wrong answer is much harder to contain.
Days 1 to 15: define the problem and baseline
Describe the current workflow, its owner and the people who use its output. Measure how long a representative sample takes, what corrections are required and what a satisfactory result looks like. Include the review work that already happens, rather than comparing a checked human result with an unchecked AI draft.
Write a short hypothesis, such as “assisted drafting will reduce total preparation and review time without increasing material corrections”. Choose an outcome the team can measure and identify the circumstances in which the tool must not be used.
Days 16 to 30: establish controls and the test sample
Agree permitted inputs, access, provider terms and the review procedure. Use synthetic or appropriately anonymised material where possible. Define who can approve an output for real use and how staff report errors. Keep an approved version of the prompt or task instructions so that the test is reasonably repeatable.
- Name the process owner and independent reviewer.
- Choose a representative set of normal and difficult cases.
- Record the baseline answer or assessment criteria.
- Define a stop condition for an unacceptable error or data exposure.
- Keep the normal process available throughout the pilot.
Days 31 to 60: run alongside the normal process
Collect results without letting the tool send messages or change employment records automatically. Record generation time, review time, necessary corrections and the final outcome. Track failure cases such as missing context, inconsistent instructions and plausible statements that are not supported by the source material.
For an illustrative calculation, suppose a manual task takes 40 minutes and the assisted version takes 10 minutes to prepare plus 20 minutes to review. The measured saving is 10 minutes, not 30. If the output needs another 15 minutes of correction, that example no longer shows a time saving.
Days 61 to 75: assess quality and cost together
| Measure | Decision relevance |
|---|---|
| Total preparation and review time | Whether the workflow saves effort overall |
| Material corrections | Whether errors are acceptable for the task |
| Unusable outputs | How often the fallback process is required |
| Provider and training costs | Whether adoption is economically reasonable |
| Staff and user feedback | Whether the process is usable and understandable |
Separate minor editing from errors that change meaning or affect a person. A tool that produces elegant prose but introduces an incorrect policy rule may create more risk than a slower draft. Describe limitations honestly, particularly when the sample is small or does not include the difficult cases expected in real use.
Days 76 to 90: make an adoption decision
Choose to adopt, revise or stop based on the recorded evidence. If adopting, define the permitted scope, review owner, budget and next review date. If revising, state what needs to change and which part of the evaluation must be repeated. Stopping a weak use case is a useful outcome when it prevents a larger unsupported investment.
Monitor the workflow after launch because a provider update or a new document type can change performance. Retain a fallback and keep responsibility with the organisation's accountable staff. Trinity's AI Workforce Hub and Systems Development service can help connect a technology trial to a defined operational problem and a practical evaluation plan.
