AI intake for Ontario law firms: what the LSO rules let an agent say

Key takeaways

• AI intake for Ontario law firms runs on rules that never mention AI. The Law Society of Ontario's conduct rules were last amended in June 2022, and the closest they come is a line asking lawyers to understand the technology they use.
• The Law Society's own practice note supplies the test: using generative AI "is akin to receiving assistance from a non-licensee employee". Read that way, By-Law 7.1 settles most of it. An intake agent may not accept a client for the firm, and may not give legal advice.
• The duty of confidentiality starts before anyone is a client, so the agent should take the names for the conflict check and stop before the story.
• On a property management line over the summer, our own agent offered a rule of tenancy law four times in 2,147 graded calls. On a law firm's line, each of those would be the advice the by-law forbids.

What Ontario has published

In February 2026 the Oregon State Bar answered the question head on. Its Formal Opinion 2026-208 asks whether lawyers may use chatbots with clients and potential clients, and concludes "Yes, qualified." The Law Society of Ontario has published nothing that specific. Its Rules of Professional Conduct are current to June 28, 2022, and that last amendment added commentary to rule 3.1-2 on understanding "the benefits and risks associated with relevant technology". No rule names a chatbot or a voice agent.

What Ontario does have is an April 2024 white paper on licensee use of generative AI, which names intake chatbots as a risk, and a practice note on professional obligations that tells lawyers to treat the tool like a non-licensee on staff. Hold that analogy against By-Law 7.1, which lists what a lawyer may never let a non-licensee do, and you have a working rulebook for an intake agent. We read all of it on 7 October 2026.

We have not built intake for a law firm. We build and run the two voice agents on a Toronto property management and brokerage operation's phones, and a separate grader reads every call they take. The misses in those reports are the clearest evidence we have of where an intake agent drifts, and a law firm's line is where that drift costs the most.

Four times our agent stated the law

Between 1 June and 10 September 2026 the property management agent took 2,147 graded calls. On four of them, according to the call reports, it offered a rule of tenancy law that nobody had put in its instructions.

It told one caller that a phone call counted as notice. Under section 43 of the Residential Tenancies Act, a notice of termination is a signed document in the Board's approved form. It told another that any move out date would work with sixty days' notice, when section 44 also requires the date to fall on the last day of a rental period. It gave a firm notice period for removing an occupant who was not on the lease, and it told a fourth caller their notice was valid without having seen it. The grader flagged two of these as errors and could not settle the other two, which is why only two of them appear among the flagged calls in our count of the agent's mistakes.

Four calls is a small number, and it is still the one that matters most if this agent ever moves to a law firm. A caller asking whether their notice is valid is asking for legal advice, and a confident answer in the firm's voice is what section 6(1)(a) of By-Law 7.1 says a lawyer must not permit a non-licensee to give. The white paper names the same risk: a client could read the tool's answer "as being tailored to their specific legal situation and rely upon it."

Each was the agent trying to be helpful on a question its instructions did not cover. That is the failure an intake agent has to be built against.

The call, step by step

Say what is answering

Commentary 5.1 to rule 6.1-1 asks that a non-lawyer be "identified as such" when dealing with the public. An agent should say in its first line that it is the firm's automated intake and not a lawyer.

Disclosure costs something, and we have seen it. A caller reporting a flood hung up right after the agent said it was automated, before giving an address. A firm's agent needs a route to a person for the caller who will not talk to a machine, and for an arrest or a hearing tomorrow that route should be offered first.

Take the names, then stop

The duty in rule 3.3-1 does not wait for a retainer. Its commentary extends confidentiality to "anyone seeking advice or assistance", whether or not the lawyer ever acts, and warns that information taken "on an informal or preliminary basis" may stop the firm acting for someone else in the same matter. An agent that lets a caller tell the whole story before the conflict check has taken that risk on the firm's behalf.

So the order of the questions matters: the caller's name, the other side's names, the kind of matter in a few words, a callback number. Then the conflict search, and the facts wait for a person.

The names have to be right. On the property management line the agent misheard a name, number or email on 19 calls and failed to capture a core detail on 110. There, a misheard surname means a callback to the wrong number. On a law firm's line it means a conflict search on the wrong name, which comes back clean. Every party name should be spelled back to the caller, and a name nobody could spell is a reason for a person to call back.

One more question belongs before the story: whether the caller already has a lawyer for this matter. Rule 4.1-2 bars means "intended to influence a person who has retained another lawyer or paralegal" to switch, and an agent built to convert inquiries will try unless it is told to stop.

Never sound like a yes

Section 4(2)(a) of By-Law 7.1 says a lawyer "shall not permit a non-licensee to accept a client on the licensee's behalf." Acceptance does not need a signed retainer. Rule 1.1-1 counts as a client anyone who, having consulted the lawyer, "reasonably concludes" the lawyer has agreed to act, and one of the white paper's concerns about intake chatbots is a lawyer and client relationship forming "without the licensee's knowledge".

Agents drift toward reassurance. Ours told two callers it had all their details and would pass them on when it had taken none. On an intake line the same reflex comes out as "we'll take care of this" or "it sounds like you have a case", and either can be enough for a caller to believe the firm has taken them on. The script should say what happens next, who calls back and roughly when, and nothing that sounds like a promise about the matter.

Write the legal question down

Most callers will ask for advice, because that is why people call a lawyer. The agent's job is to take the question down in the caller's words and say a lawyer will answer it, every time, however simple it sounds. Words like notice, deadline, served, court date and limitation should send the call down that path rather than into anything the model thinks it knows. All four of our tenancy answers were about notice.

Follow up the way a person would

Automated follow up is where intake agents earn their keep, and where rule 4.1-2 applies again: no means that "take advantage of a person who is vulnerable or who has suffered a traumatic experience and has not yet had a chance to recover." A sequence that texts someone three times the day after a car accident is that, whatever the template calls it. Personal injury and family intake need a slower cadence and a person deciding when to chase. Consent for the messages is a separate question, covered for brokerages in our CASL post.

Know where the transcript lives

The technology commentary to rule 3.1-2 asks a lawyer to weigh a tool's risks while "recognizing the lawyer's duty to protect confidential information". In practice that means knowing where every call ends up. Retell, the platform our agents run on, stores transcripts, recordings and logs unless the agent is set otherwise, and by its own API reference keeps them forever unless someone sets a retention period between 1 and 730 days. On the property management line, call summaries also post to a team chat channel, which is fine for a leaking tap. A prospective client's account of an assault belongs in the firm's practice management system, under the firm's permissions, and nowhere else.

What supervision means for an agent

Rule 6.1-1 makes the lawyer responsible for the whole practice, and section 4(2)(e) of By-Law 7.1 asks for review of a non-licensee's work "at frequent intervals". The white paper goes further for client-facing chatbots and recommends avoiding them unless controls ensure "predictable and reliable outputs".

Our property management agent took about 21 calls a day over the summer. At that volume a spot check is not frequent review. We found the four tenancy answers because a grader reads every call and a person reads the grader. A firm running an intake agent should expect to read its flagged calls every week, and to change the instructions each time one of them turns out to be a legal answer.

That is the build we would propose to a firm. The agent answers, says what it is, takes the names, puts the conflict search inputs into the practice management system, writes the question down and books the lawyer's callback, and every call is graded and every flag read. It is what our page for law firms describes, and the Clio side of it is in Clio automation for a Canadian firm.

This is a reading of the rules for building software, not legal advice. Check the current text on lso.ca before relying on any of it.