How do we get a sceptical partner to agree to try AI?

Most AI conversations inside a firm fail because they happen in the abstract. The partner says it makes things up, the enthusiast says it saves hours, and nobody can settle it because nothing has been tested. The way through is to pick one narrow task the partner can check in ten minutes, run it on a real matter with the firm's own papers, and let them judge the result against what they already know.

Work out what the scepticism actually is

"I don't trust AI" is rarely one objection. It is usually three or four, and they need different answers.

• Accuracy. They have read about fabricated authorities and assume everything is guesswork. This is answered by output that names its sources and by a tool that says when it cannot find something.
• Confidentiality. They do not want client papers sitting on a server somewhere, feeding a system used by other firms. This is answered by the supplier's contract and data arrangements, not by a demonstration.
• Liability. They are the one who signs the letter and holds the practising certificate. This is answered by being clear that nothing goes out unchecked and that the tool does not send, file or serve anything.
• Cost and distraction. They have sat through software projects that consumed months and delivered little. This is answered by keeping the first step small and reversible.

Ask which of these is the real one. If the answer is confidentiality, no amount of impressive drafting will help. If the answer is accuracy, a contract clause will not help either. Match the response to the objection.

Pick a task with a result they can check quickly

The wrong first task is anything long, subjective or hard to verify. A sceptical partner cannot assess a strategy memo in a few minutes, and will not try.

Good first tasks share three features: the partner already knows the answer or can find it fast; the output is short; and every statement can be traced back to a page or a source.

Suitable candidates include:

1. A chronology of a file the partner knows well, so they can spot what is missing or out of order.
2. A list of the key dates and figures on a matter, with the working shown, checked against the diary entries already made.
3. A summary of a long document the partner has read, to see whether the emphasis is right.
4. A set of questions a reviewer should ask about a file, compared with the questions they would have asked.

The point is not to impress. The point is to let an experienced person apply their own judgement to something small enough to evaluate honestly.

Use a real matter, not a demonstration

Supplier demonstrations use documents chosen because they work well. Partners know this, which is why demonstrations rarely move anyone. Use a live or recently closed file from the firm's own work, with its usual scanned exhibits, incomplete attachments and poorly named PDFs.

Before doing that, check the data position: who the supplier is, where the information is held, who inside the firm can see it, and whether the client's retainer or any data processing arrangement raises a question. If the firm is not satisfied on those points, a closed matter or an anonymised set of papers is a sensible starting point, and the partner will usually accept the test on that basis.

Then let them mark it. Do not explain away the misses. A partner who finds two errors and sees them acknowledged will trust the exercise more than one who is told everything was perfect.

Agree in advance what a yes and a no look like

Decide the test before you run it, or the discussion will drift. Write down, in a few lines:

• The task being tested and the file it will run on.
• Who checks the output and how long that check should take.
• What would count as a failure: a wrong date, an unsupported statement, a missing document, a source that does not say what the output claims.
• What happens if it passes: who else tries it next, on what, and when the firm reviews it again.

This matters because it converts a question of belief into a question of evidence, and because it gives the partner a genuine veto. People agree to tests they are allowed to fail.

Be honest about what is still the fee earner's job

The strongest argument to a cautious partner is not that AI is reliable. It is that nothing changes about who is responsible. The fee earner still reads the sources. The supervisor still reviews the draft. The firm's diary still governs its deadlines. The file still records what was done and who checked it. If someone cannot explain the reasoning in the output, it does not go out.

Framed that way, the decision is smaller than it looks. The firm is not handing work to a machine; it is adding a step that produces a first pass with its sources attached, which a qualified person then checks as they would check a trainee's note.

Where Alesis fits

Alesis is an AI assistant for UK law firms, made by Alesis.ai Limited and used through the web browser, with one conversation for a matter. It answers questions about a matter from the matter's own papers and names the page each answer came from; where the papers do not say, it says so, and when it cannot find support for a point it says what is missing instead of guessing. The firm's information is held in the UK and processed only in the UK and the EU, kept apart from every other firm, and never used to train anything for anyone else. Alesis is funded by credit rather than a subscription: the first firm a person sets up starts with £25 of free credit, no card needed, which is usually enough to run the kind of test described above.