A 90-day AI pilot plan for a law firm
By the ChamberOne team 5 minute read
Partners who sponsor an AI trial usually want two things: to avoid a costly mistake, and to know by the end whether it was worth it. A pilot with a start, an end and a decision date gives both. Ninety days is long enough to learn something real and short enough that nobody loses interest. Here is a plan, week by week, for a firm that wants to test one use of AI properly.
Key points
- Choose one bounded, checkable use case and name both a sponsor and a day-to-day lead.
- Write down success measures in week one: time, quality, safety, experience and cost.
- Settle data answers and a one-page review rule before anyone uses the tool on client work.
- Run with a small group including a sceptic, meet weekly and keep a list of every error.
- Decide at the end to stop, adjust or extend, and keep the lessons either way.
Weeks 1 and 2: choose one use case and one sponsor
Pick a single task that is frequent, bounded and easy to check. Summarising long documents for internal use, producing first drafts of routine letters or extracting dates from contracts are common candidates. Avoid anything that goes to a court or a client without review.
Name a sponsor, a partner with the authority to decide, and a day-to-day lead who will run the trial. Without both, pilots drift.
Suppose a twelve-person firm picks the extraction of key dates and parties from commercial contracts. The sponsor is the head of the commercial team, who can say yes or no at the end. The lead is a senior associate who gives two hours a week. Both calendars are blocked before the pilot begins, because a pilot nobody has time for is not a pilot.
Weeks 1 and 2: set the measures before you start
Decide in advance what success means. Write it down now, because after the trial everyone remembers it differently.
- Time: how long the task takes with and without the tool, estimated by the people doing it.
- Quality: how many corrections are needed before the output can be used, and how serious they are.
- Safety: the number of errors caught in review, and whether any reached a client or a court.
- Experience: whether the people using it want to keep going.
- Cost: the price of the tool, plus the time spent checking.
Weeks 3 and 4: settle the rules and the tool
Work through the data questions with the supplier: what is stored, whether input is used for training, where it is processed and who can see it. Write down the answers. Check the guidance from your regulator or bar on confidentiality and AI.
Agree a one-page review rule. A lawyer checks everything before it leaves the firm, citations are traced to sources, and each use is recorded on the file. Tell clients where appropriate.
Check also what your professional indemnity insurer needs to know. Some insurers ask about the use of new technology, and it is better to have the conversation before a problem than after one.
Weeks 5 to 8: run it with a small group
Choose three to six people, including at least one sceptic and one junior. Give them a short briefing on what the tool is for, what it is not for and how to report problems. Ask them to keep a simple log of each use: the task, the time taken, what had to be corrected.
Meet weekly for fifteen minutes. Look at examples of good and bad results, share tips on how to ask better questions and note anything that worries people. Keep a list of every error found.
Give the group one rule of thumb for the log: write it down at the time. Memory of what took how long fades within a day, and a log reconstructed on Friday is mostly guesswork.
Do not count only the tool's best moments. Ask people to log the times they chose not to use it and why. Those entries often show where the tool does not fit the work, which is as useful as knowing where it does.
Weeks 9 and 10: review the errors honestly
Go through the error list. Sort errors by type: invented facts, wrong dates, misread clauses, poor tone, missed points. Ask which would have caused harm if not caught, and whether the review rule caught them.
A pilot with no errors is a warning sign, not a success. It usually means nobody was looking. The aim is to learn how the tool fails in your work so that your checks are matched to it.
Weeks 11 and 12: decide
Hold a decision meeting. Compare results with the measures written in week one. There are three honest outcomes.
- Stop. The benefit was small, the errors were costly or people did not want it. Record why and share it.
- Adjust. The tool helped on some tasks and not others. Narrow the use, change the instructions or try another product, then run a shorter second round.
- Extend. The results met the measures. Expand to another team or add one further use, keeping the same review rule.
What to keep from the pilot, whatever the result
The error list, the rule, the data answers and the lessons about how to ask are valuable even if you stop. They form the foundation of any future trial and show regulators, insurers and clients that the firm approached AI carefully.
Tell the firm what happened, including what did not work. A plain account builds trust more than an enthusiastic announcement.
Keep the pilot report to two pages: what we tried, what we measured, what we found, what we decided and what we would do differently. A short report gets read, and it can be handed to the next team as a starting point.
If the decision is to extend, set the next review date at once and keep the measures. A pilot that turns into a permanent arrangement without a checkpoint tends to lose the discipline that made it useful.
Common reasons pilots fail
- No clear question, so nobody can say whether it worked.
- No sponsor with authority to decide, so the pilot lingers.
- Success measured only by enthusiasm.
- Too many uses at once, so the lessons are muddled.
- No one assigned time to run it.
- Start with one clear, bounded use case.
- Name a sponsor with authority to decide.
- Measure time, quality, safety, experience and cost.
- Give someone time to run it.
- Try too many uses at once.
- Judge success by enthusiasm alone.
- Let the pilot linger with no one to decide.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
How a small law firm can start with AI without a big project
Pick two low-risk uses, set review rules and measure before expanding, with a list of what to postpone.
5 minute readChoosing AI models and firm-level controls: what a practice should decide
Model choice, an off switch and a usage record are three decisions a firm should make deliberately.
4 minute readMeasuring AI return on investment in a law firm
Measure per task against a baseline, count review time, and claim only what you can trace.
4 minute read