How to Plan a 30-Day AI Pilot at Work

A sales rep drafts proposals from a standard template, then checks every version before sending it. That task may be suitable for an AI pilot if preparation time and correction effort can be measured before and after the change.

This is a model plan for a 30-day trial. It describes no particular client and promises no company-wide rollout. Its purpose is a decision about one bounded process. The schedule assumes approved data and a tool the company can use. Legal review or integration work may extend it.

Week 1. Choose the work and record a baseline

Walk through recurring tasks with the process owner. Drafting from an approved template, summarising work notes or checking document completeness may be candidates. A person must be able to assess the output before it is used.

Choose one pilot from the available inputs, frequency and ability to catch mistakes. Hiring decisions, payments and contract changes are outside this model trial. Their consequences need a separate assessment.

Record case volume, handling time including review, and corrections. Label estimates as assumptions where measurement is missing. Otherwise the apparent saving can be work shifted to the colleague who fixes the output.

Also write down what must remain unchanged. For a proposal that could include prices and binding terms, which the model must not invent.

Week 2. Set up the tool and data permissions

Start with tools the company already permits. Test the actual input and whether someone can check the result. Being included in the existing software suite does not make a tool automatically easy to adopt.

Before purchasing, verify the current plan, price, billing basis, usage limits and processing terms. This guide does not reproduce vendor pricing because those terms change.

Security checks need to reach the actual workspace configuration. For example, Notion documents AI data handling and subprocessors. The pilot still needs a review of its own permissions and contract. Do not send sensitive material into a service without an approved basis for that use.

Write down permitted inputs, output storage and the reviewer. Use synthetic or appropriately anonymised material for early trials. Name the contact for accidental sharing and unsafe output. The NIST AI RMF puts risk management into design, use and evaluation, rather than treating it as a sign-off after launch. It does not certify the chosen tool.

Week 3. Run the task with its real users

Use a small group who already do the work. A colleague can gather questions and help improve instructions, while the process owner remains accountable for the trial. That support role is not a substitute for expert review.

For the proposal example, supply the approved template and relevant source material. Specify the required output and which facts the model must not fill in. The reviewer compares the draft with those sources and corrects it before anything goes out.

Measure from intake to acceptance, including review time. Keep cases where manual work was better. Opening the tool regularly does not establish that it reduced work.

Pause the trial if data is used outside the agreed permission, an output creates unacceptable risk or review stops being practical. Agree who can return the team to the old procedure.

Week 4. Decide what the evidence supports

Compare equivalent work with the baseline. Count licences, setup, training and correction time. Saved hours are not automatically cash savings, and a short trial may not reveal the long-term return.

Continue when the owner accepts both the useful result and the remaining risks. If a specific obstacle can be fixed, name the change and a new review date. Stop when benefit is absent or control costs too much. An inconclusive return may reflect the tool, missing inputs or a process that does not yet justify automation.

The month should end with measurements, limitations and a decision about scope. Our first-automation guide helps prepare the inputs.

To plan a trial with us, see AI and automation. Bring the process, its owner and a sample of permitted inputs to the first conversation. Tool selection follows that assessment.