Find the Fit—Leverage AI in a Law Practice, Then Test to Trust

By Brian Long

In the first two articles we talked about how to think about AI and how to read the bill. A large language model is a language processor — it takes language in and gives transformed language back. And the meter it runs on is the token, the little chunk of text every prompt and every response is measured in.

In this article we are going to make practical application where the stakes are unmistakable: a small law practice. If you don’t run a firm, stay with us anyway — we are working through law as an example precisely because it’s unforgiving. We will answer two questions every business owner eventually asks as they are the same in any field: What do I actually use this thing for? And how do I know the answer is any good? In a law office, getting the second one wrong gets you sanctioned. The discipline that protects a litigator protects everyone.

So, meet our example. Elena is the managing partner of a six-lawyer general practice — some estate planning, some family law, a steady stream of small-business clients, and the occasional litigation matter. She’s heard AI can help. She’s also read the headlines about lawyers filing briefs full of cases that don’t exist. She is, reasonably, both interested and nervous.

Believe AI is a Fit for You

Before we find Elena’s tasks, it’s worth naming the thing that stops most owners from starting. When the U.S. Small Business Administration’s Office of Advocacy looked at why small firms hadn’t adopted AI, the single most common reason wasn’t cost, security, or regulation. It was the belief that AI simply isn’t applicable to their business — cited by roughly 82 percent of the smallest firms (AI in Business: Small Firms Closing In, SBA Office of Advocacy, September 2025). Lawyers are especially prone to this. “My work is judgment, not text,” the thinking goes.

But look at what fills a lawyer’s day. Engagement letters. Client update emails. Demand letters. Discovery summaries. Intake notes. Standard clauses. Billing narratives. The SBA data shows small firms have largely closed the adoption gap with big companies — small-business AI use climbed from 6.3 to 8.8 percent in a matter of months, and the smallest firms now adopt at surprisingly high rates. The report’s underlying economics are blunt: firms that skip productivity-improving tools tend to lose ground to competitors that don’t. The question isn’t whether a law practice has language work. It’s drowning in it. The only question is which pieces are safe and worthwhile to hand off.

Part One: Find the Hidden Recurring Tasks

The way to find them isn’t to brainstorm “AI use cases” but to audit your own week. Walk your calendar and your sent-mail folder and flag every piece of writing that wasn’t the reason you went to law school. Then run each candidate through four simple tests (7 Business Processes You Should Automate First, Distrya):

  • Is it recurring? Something you write weekly or daily, not once a year.
  • Is it language in, language out? Drafting, summarizing, rewriting, extracting — the language processor’s home turf.
  • Can you check it quickly? You need to be able to verify the result without redoing the whole job.
  • Is it low-stakes to start? Build trust on the safe stuff before you ever rely on it for anything that touches a court.

Run Elena’s practice through that filter and the candidates sort themselves into tiers.

Client correspondence and intake. The “we’ve received your documents and here’s what happens next” email. The new-client intake summary turned from messy notes into a clean memo. This is high-volume, easy to check, and low-stakes — the ideal place to start. Microsoft’s Copilot, already sitting inside the Office tools Elena’s firm uses, can draft these from a few bullet points without anything leaving her Microsoft environment.

Document drafting from your own templates. Engagement letters, standard clauses, a first-pass demand letter. The key phrase is from your own templates — give the AI your approved language and your facts, and ask it to assemble and tailor, not to invent. Contract review and revision is one of the functions AI genuinely does well: flagging unusual terms, spotting inconsistencies, comparing a draft against your standard (NCSC, A Legal Practitioner’s Guide to AI and Hallucinations, January 2026).

Summarizing and triage. Condensing a long deposition transcript, a dense contract, or a pile of discovery into digestible insights; identifying which documents deserve a human’s full attention first. The guide rightly files document summarizing and “identifying research angles” among the lower-risk uses — useful, and forgiving of error, as long as a person still reads what matters.

Administrative and marketing language. Billing narratives, scheduling emails, blog posts, the firm’s newsletter. Marketing is, in fact, where small businesses lead big ones in AI use (SBA). It’s also about as low-stakes as language work gets.

Legal research — handle with care. AI research tools can scan case law in seconds, and that’s seductive. But this is where the stakes spike, and we’ll treat it as the cautionary centerpiece of Part Two rather than a starter task.

The rule for picking your first project is simple: start with the single biggest time-drain that’s also easy to check (15 AI Business Use Cases, Product School). For Elena that’s client correspondence — high volume, low risk, instantly verifiable. Prove it there, bank the time, build confidence, then expand outward toward the higher-stakes work as your verification habits harden.

Part Two: Trust, but Verify — Without Losing the Time You Just Saved

The whole point of handing off drafting is to save time. If you then re-do every task from scratch to check it, you’ve saved nothing. The skill is to calibrate your verification to spend exactly as much checking effort as the stakes require, and no more. The legal guide gives us the cleanest framework anywhere for doing this, and it generalizes to any business.

The cardinal rule

Treat every piece of AI output as a first draft requiring review — never a finished product. The guide’s memorable way of putting it: AI legal tools are powerful assistants but terrible supervisors (NCSC). You stay the supervisor. The AI never makes the final call.

Why so firm? Because of how a language processor works — the same mechanics we covered last time. It generates the text that is statistically likely, which means it produces language that sounds right rather than language that is right (NCSC). When its training suggests a citation should exist, it can simply invent one — with a plausible caption, a plausible citation, and a plausible holding. A hallucination doesn’t look like an error. It reads like real authority because it’s assembled from the same patterns as real authority. That’s exactly what makes it dangerous, and it’s why courts have grown willing to sanction attorneys who file it.

Match the scrutiny to the stakes

You don’t check a thank-you email the way you check a brief. The guide sorts uses into risk tiers, and it’s a perfect model for calibrating effort:

  • Low risk — brainstorming arguments, drafting outlines, summarizing documents, internal first drafts. A quick read is enough.
  • Moderate risk — anything that cites a case, paraphrases the law, states a procedural requirement, or gives fact-specific advice. This needs deliberate checking of every specific.
  • High risk — output involving legal interpretation or analysis that could affect someone’s rights, due process, or the outcome of a matter. Maximum scrutiny.
  • Unacceptable — submitting unvetted AI output to a court, relying on unverified citations, or letting the tool make the judgment call. Simply off the table.

The general business translation is the same shape: an internal memo gets a glance, a customer facing message gets read and fact-checked, and anything with legal or financial weight gets every claim verified against a primary source before it goes out.

The manual techniques that actually catch errors

For the moderate-and-up tier, the guide is specific, and the advice is refreshingly old-fashioned (NCSC):

  • Check every citation in a primary source. Read the actual case, not the AI’s summary of it — confirm the quote is accurate, the context is intact, and the holding really says what the draft claims.
  • Verify statutes and rules in official sources, including section numbers, effective dates, and any amendments.
  • Distrust a too-perfect fit. A hallucinated case often lines up suspiciously well with your fact pattern. When something seems too convenient, treat that as a warning, not a gift.

What carries through: trust the prose, check the facts. Names, numbers, dates, quotes, and citations are exactly where a confident-sounding model wanders off, and exactly what a thirty-second lookup confirms or kills.

Faster, semi-automated help

You don’t have to do all of this by hand. The guide endorses a few force-multipliers (NCSC):

  • Citation-checking software can automatically flag suspicious or unverifiable references.
  • Cross-check with a second model. Run a critical question through two different frontier tools and compare; discrepancies surface the shaky claims. Independent agreement between strong models is a meaningful (if not absolute) signal.
  • Ground the model in your sources. Make it work from a document you provide and ask it to quote the supporting line, rather than drawing on its training. The verification techniques you adopt should themselves be checked — even your fact-checking process deserves a sanity test (How to Verify Facts with AI in 2026, Illumination on Medium).

A note on tools: purpose-built legal research platforms increasingly link straight to the underlying cases for exactly this reason, and the more general assistants built into the software you already run — Microsoft’s Copilot, Apple’s on-device intelligence — keep routine drafting close to home. At Synapses Technologies we’re building toward a local AI stack precisely so that sensitive client work can be drafted and checked without leaving your own walls. More on that in a future article.

Make it a system, not a habit

Individual diligence fades under deadline pressure. The guide’s answer is to institutionalize it (NCSC): require a second review for AI-assisted work, with reviewers told AI was involved; keep a reusable checklist (the canonical line: “Have all citations been verified in primary sources?”); maintain an audit trail of issues; and designate one person to stay current on each tool’s known limitations and training-data cutoff. For lawyers this isn’t optional polish — it maps directly onto existing duties of candor to the court, competence, and supervision of nonlawyer assistance. For any business, it’s just how you make quality survive a busy week.

Putting It Together

Notice the loop Elena is now running. Find a recurring language task that’s safe to hand off. Brief it well — the task, role, context, and examples from the first article. Verify at the right level — a glance for the newsletter, a primary-source check for anything that cites the law. Bank the win, and move one notch up the risk ladder as her habits harden.

That’s what proficiency with today’s frontier models actually looks like before anyone touches anything more advanced. It isn’t a single dramatic transformation; it’s a growing list of two-minute wins, each one verified in proportion to what it’s worth. The lawyer who masters this has more time for the judgment that’s genuinely hers. The barrier was never that AI doesn’t apply to the work. It was learning to aim it and to check it.

Once that rhythm is second nature — and only then — it makes sense to talk about letting AI take several steps on your behalf rather than one at a time. That’s where this series heads next.

At Synapses Technologies, helping small and medium businesses — law practices very much included — find the practical, safe fit for AI is what we do. If you’re staring at your own week wondering which task to hand off first, that’s exactly the conversation we’d love to have.

Sources