How to pilot AI without losing the firm

Most firms don't have an AI problem. They have a first move problem.

The pressure to "do something with AI" is real, and it usually produces one of two outcomes: a splashy tool that nobody adopts, or a stalled committee that studies the question for a quarter and ships nothing. Neither moves the business. The firms that get value from AI tend to do something quieter — they pick one painful, repetitive task, run a small pilot, measure it honestly, and keep a person in the loop the whole way.

Here's the framework we use.

Start where the hours are, not where the hype is

The best first use case isn't the most impressive one. It's the one that quietly eats hours every week.

Look for work that is repetitive, document-heavy, and follows a pattern your team could describe in a few sentences. In one engagement, a team generated settlement demands by hand — pulling medical records from a shared drive, deciding what mattered, assembling exhibits into a single PDF, and using past demands as templates. Each one took two to four hours. That's the profile of a strong first pilot: high volume, clear inputs, a recognizable output, and a process the team already understands well enough to check.

Avoid starting with anything that touches legal judgment as its output rather than its input. AI should compress the assembly work and hand a draft to a human, not make the call.

Train on your own work, and keep it where your work lives

Generic AI gives generic results. The leverage comes from training a workflow on a handful of your own examples — your demands, your letters, your intake notes — so the output matches how your firm actually writes.

Just as important: keep the data where it already lives. A common mistake is adopting a tool that requires you to upload sensitive records into a new external system. That creates a second copy of your most confidential material, a new vendor to trust, and a new place for things to leak. Wherever possible, connect the AI natively to the systems you already run — your document store, your case management platform, your phone system — so nothing has to leave its home.

Pilot, then measure the before and after

You cannot estimate AI's value from a slide. You have to run it.

Take a few real examples, run them through the workflow, and time it. If a task that took two to four hours now takes fifteen to thirty minutes, you have a number leadership can act on. If it doesn't, you've learned that cheaply, before signing anything. Either way, the pilot — not the demo — is what tells you the truth. "There's no easy way to estimate this until we actually do it" is the right instinct.

A good pilot has three things: a small, representative sample; a clear measure (usually time, sometimes error rate or cost per task); and a defined finish line so it doesn't drift into a permanent experiment.

Keep a human in the loop — and design where they sit

"Human in the loop" gets said a lot and designed rarely. Be specific about it.

For a drafting task, the AI produces the first draft and a person reviews and approves before anything goes out. For an answering service, the AI handles the calls your team would otherwise miss — greeting the caller, capturing the essentials — and then creates a task or sends a notification so a human picks it up. The point isn't to remove people. It's to remove the grind and let people spend their judgment where it counts.

This is also your safety rail. The reviewer catches the rare bad output before it reaches a client, and that review step is where trust is built with a skeptical team.

Phase it: one department first, then expand

Don't boil the ocean. Prove the pattern in one place, then extend it.

A firm replacing a costly live answering service might start with a single department — say, post-intake case management — before touching the more complex intake process, which has scripts and compliance requirements of its own. Same with document automation: one document type, one team, one clear win. Each successful phase funds and de-risks the next, and it gives the rest of the firm a real example to point at instead of a promise.

Buy a platform you can keep building on, not a one-off

When you do bring in outside help or tooling, weigh the one-off against the platform.

A single-purpose tool solves today's task and stops there. A platform you can keep building on means the next automation — the next bottleneck you find — costs a fraction of the first, because the integrations and the know-how are already in place. For most growing firms, two or three tangible wins plus the ability to keep going beats a shiny, dead-end solution.

The quiet version wins

None of this is dramatic. Pick a real bottleneck. Train on your own examples. Keep the data home. Pilot it and measure. Keep a person in the loop. Expand one phase at a time. That's how you get AI that pays for itself without betting the firm on it — and how you build the muscle to do it again.

What's the best first AI project for a firm?

A repetitive, document-heavy task with clear inputs and a recognizable output — document drafting, intake summarization, or handling missed calls. Save anything requiring legal judgment as the final output for later, with a human reviewing every result.

How do we know if an AI pilot is working?

Measure the before and after on a small, real sample. Time the task, track error rate or cost per item, and set a finish line. If the numbers don't improve meaningfully, you've learned that cheaply before committing.

Is it safe to use AI with confidential client data?

It can be, if you keep the data in the systems you already control and connect AI natively rather than uploading records into a new external tool. Pair that with human review and clear governance on what the AI is allowed to do.