Incoming mail on a live engagement is four jobs pretending to be one. Automate three of them and you have a firm that answers on time.
Key takeaways
You manage client email with AI by breaking triage into four separate jobs — classify the message to an engagement, extract what is being asked and by when, draft a reply for a person to review, and file any attachments to the right client folder — and by automating the first, second and fourth while keeping a human on the third.
Most firms attempt the opposite. They start with drafting, because it demonstrates well, and get least value from it. Classification and extraction are where a firm stops losing things.
An engagement-linked inbox is worth more than a smart one. The useful output is a label — which client, which engagement, which stage — plus a confidence. Anything the classifier is unsure about goes to a person, and that queue should be short enough to clear before lunch.
This is also the job where firms already have the training data: your existing correspondence is already sorted by matter. Use it.
Every client message either asks for something, provides something, instructs something, or is noise. Extraction should return which of the four it is, what precisely is being requested or supplied, and any date mentioned or implied — a filing deadline, a closing, a board meeting, a renewal.
That output feeds a work list, and the work list is the actual product. A firm that knows every outstanding client ask and its date is running a different practice from one where that knowledge lives in six people's heads. The Law Society of Ontario's Client Service and Communication guideline frames client service as an agreement about what the client may expect and the lawyer's promise to meet it — and you cannot meet expectations you have not captured.
A drafted reply is a first draft, and it should be visibly one. The pattern that works is a draft that pulls the answer from your own materials — the engagement letter, prior correspondence, the document just received — and flags anything it could not source.
Nothing sends automatically. A reply from a professional firm is a communication a client may rely on and may later rely on in a dispute, and an email exchange can carry more legal weight than people assume — see contract formation by email and whether a demand letter sent by email is valid. Neither is a reason to avoid email; both are reasons to keep a person on the send button.
Attachments are where the quiet losses happen. An extraction step that reads the attachment, names it consistently, and files it to the engagement folder removes a step everyone dislikes and nobody records having skipped.
The Law Society of Ontario's File Management guideline is explicit that the principles apply to electronic folders and documents just as to paper, and that lawyers must have systems to store and retrieve client information, maintain and protect confidential information, and check for conflicts and limitation periods. Automated filing serves that; automated filing without a naming convention undermines it.
Any message that changes where money goes gets a hard stop. New banking details, a revised wire instruction, a request to send funds to a different account, a "quick" change of payee — these are the highest-value fraud targets aimed at professional firms, and an automated pipeline that files, summarises and drafts a helpful reply is a fraudster's ideal audience.
Build the rule as a hard exception: classify, then refuse to act, then route to a named person for out-of-band verification. Our sister firm's answer on fake wire instructions sent to a lawyer sets out how the scenario actually plays out.
A four-partner practice returns to 180 messages. Classification assigns 141 to existing engagements and leaves 39 unsure. Extraction produces 63 asks with dates, of which 11 have deadlines inside five business days. Two messages are flagged as banking-instruction changes and stopped.
The partner's Monday is now: clear a 39-item classification queue, look at 11 urgent asks, personally handle two flagged messages, and review drafts for the routine remainder. The inbox has not been answered by a machine. It has been sorted, and the sorting is what was taking the morning.
Client mail is personal information and often confidential. The Office of the Privacy Commissioner's generative AI principles apply to a firm using such tools, and the OPC's PIPEDA requirements in brief is the baseline to design against — particularly limiting collection, safeguards and accountability.
For lawyers, the Law Society of Ontario's security and data protection resources are the place to start on where client correspondence may be processed. Accountability inside the firm should be named, not assumed — see appointing a privacy officer.
Can it answer routine questions on its own?
Even a genuinely routine answer from a professional firm carries the firm's name. Keep a person on the send, and make the review fast rather than optional — a good draft that takes ten seconds to approve is the win.
What about the reception side — calls and appointments?
Different problem, different page. See professional practice owners for the front-of-house version.
How do we stop it filing the wrong client's document?
Confidence thresholds and a rule that a document is never filed to a client folder on a low-confidence classification. A misfiled document is a confidentiality incident, not an admin error, which is why the threshold for filing should be stricter than for labelling.
A 30-minute call is enough to tell you whether AI pays for itself here.
