Jump to section
A few months ago I watched a very capable associate spend twenty minutes fixing a factum passage that an AI tool had written for her in eight seconds. The eight seconds felt like magic. The twenty minutes were where the real work lived, because the passage cited a case that said almost the opposite of what she needed. That, in one afternoon, is the whole story of AI in law right now.
The technology is genuinely useful and genuinely dangerous, sometimes in the same document. The firms that will do well are not the ones that adopt it fastest or resist it longest. They are the ones that draw a hard, honest line between the tasks AI can handle and the tasks that are yours by professional obligation. So let me draw that line as I see it, from small-firm practice.
The Tasks AI Handles Well Right Now
Start with the good news, because it is real. AI is very strong at the mechanical, high-volume, low-judgment work that eats your billable day and pays you nothing. First drafts of routine correspondence. Summarizing a long transcript or a dense discovery document into something you can skim before deciding what to read closely. Reformatting, tidying, and reorganizing text you already trust. Turning a messy set of notes into a coherent memo skeleton you then rewrite.
It is also useful for generating ideas to check against. Ask it for ten possible arguments, or the counterarguments opposing counsel might raise, and treat the output as a checklist, not an answer. You will discard most of it. The value is in the two items you had not thought of. Extracting structured data from documents belongs here too, and it is one place a well-built practice tool earns its keep; see our note on pulling data out of PDFs for where that actually works.
Tip. A useful test before you let AI touch a task: if a junior person got it wrong, would you catch the error in seconds or only after it caused harm? The first kind is safe to delegate. The second is not.
Where Confident Answers Become Dangerous
Here is the part the marketing skips. These tools produce wrong answers with exactly the same fluent confidence they produce right ones. There is no tell. A fabricated citation looks precisely like a real one, complete with a plausible neutral citation and a paragraph number that does not exist. Canadian courts have already sanctioned lawyers for filing submissions built on cases the machine invented, and every law society in the country has made its expectations clear.
The danger is worst wherever the answer requires knowing our law specifically. Limitation periods. Whether a given tribunal even has jurisdiction. The current state of a statute after last spring's amendments. Anything provincial, anything recent, anything niche. The training data is overwhelmingly American, often stale, and blind to the registry practice and local rules that decide real files.
A tool that is right eighty per cent of the time is not eighty per cent as good as a lawyer. The wrong twenty per cent looks identical to the right eighty, which is where the risk lies.
Why Judgment Is Still Yours Alone
Strip away the drafting and the summarizing and ask what actually constitutes the practice of law. It is judgment under uncertainty, exercised on behalf of a specific human being, with your name and your insurance on the line. Should this client settle or fight? Is this clause worth the fight it will start? What is really going on in this family, behind what they are telling me?
None of that is a text-prediction problem, and pretending otherwise is how good lawyers get into trouble. The duty of competence, the duty to supervise, and the obligation to protect confidences do not transfer to a vendor because the vendor is clever. If anything, they get heavier, because you now have a fast, tireless source of plausible mistakes sitting inside your workflow. Feeding client detail into a consumer chatbot with murky data practices is its own confidentiality problem; our piece on confidentiality in the cloud applies with full force here.
How to Pilot AI Without Betting the Firm
You do not need a strategy deck. You need a small, boring, honest pilot. Pick one workflow that is low-stakes and repetitive, run it for a month, and measure whether it actually saved time once you count the review. Half of what people try turns out to cost more in checking than it saved in drafting. That is a finding, not a failure.
| Task | Delegate to AI | Keep in human hands |
|---|---|---|
| Client email, routine | First draft | Send decision, tone, promises |
| Long transcript | Summary to triage | Findings you rely on |
| Legal research | Idea generation | Every citation, verified in a real database |
| Contract review | Spotting missing clauses | Whether a term is acceptable |
| Court filings | Nothing unverified | All of it |
Watch out. Before any tool touches client data, read its terms on whether your inputs train its models and where the data is stored. If you cannot get a clear answer, that is your answer.
Set two rules everyone follows without exception. Nothing goes out the door unread by a lawyer. And nothing gets cited that a human has not opened and confirmed. If a tool cannot live inside those rules, it is not ready for your practice. When you are choosing between platforms, the same instinct that guides picking practice management software applies: buy the thing that respects your obligations, not the one with the loudest demo.
A Realistic Twelve Month Outlook
I am not going to predict a revolution, because the honest forecast is duller and more useful. Over the next year the tools will get modestly better at the tasks they already do, the good ones will get more careful about grounding their answers in sources you can check, and the reckless ones will keep hallucinating with a smile. The lawyers who benefit will be the ones who treated it as a fast, fallible assistant from day one, kept a human in every loop that matters, and refused to outsource the part clients are actually paying for.
Adopt what saves you time. Verify everything that carries risk. Keep the judgment, because the judgment is the job, and it is the one thing no model can sign its name to. A tool like A1 CMS can handle the routine so you spend your hours where they count, but the line between help and harm is one you draw and defend yourself. Draw it clearly, and AI becomes a genuine time saver rather than a source of malpractice exposure.
