Treadstone Associates
Article · Staffing, training & knowledge

AI training for office staff at a firm

Teach four things: what may go in, how to check what comes out, when to stop, and who owns the result. Everything else is product training and can wait.

Treadstone Associates · Updated 2026

Key takeaways

  • • The confidentiality rule has to be a list of named systems and data types, not a principle.
  • • Verification works as a step inside the task, not as a value in a policy document.
  • • If you monitor how staff use these tools, Ontario’s written electronic monitoring policy rules may apply to you.
  • • Support staff work under supervision; introducing a tool does not move who is accountable.

The short answer

Most AI training for office staff fails because it teaches the product. Menus change; the four things that matter do not. Teach what may be put into a tool, how to check what comes back, when to stop and escalate, and who owns the output that leaves the firm. A competent administrator can absorb all four in one session and a handful of short clinics.

The four-part curriculum

  • • What may go in. Named systems, named data types, and the explicit list of things that never leave the firm.
  • • How to check what comes out. A verification step written into the task, not a general instruction to be careful.
  • • When to stop. The trigger conditions for escalating to a manager rather than pressing on.
  • • Who owns it. Whose name is on the work when it reaches a client, and what that person needs to have seen.

Write the confidentiality rule as a list

“Do not share confidential information” fails in practice, because at four o’clock on a filing day nobody is running a legal analysis. What works is a list: these systems are approved, this data may be pasted into them, this data never leaves the practice management system, and here is the one-line test for anything not on the list — ask first.

The reasoning behind the list is worth teaching once. The Privacy Commissioner’s generative AI principles frame it as limiting collection, use and disclosure to what the identified purpose requires, and being open about what happens to the information. Staff follow a list; they follow it better when they know why it says what it says. The training obligation is not just good practice — PIPEDA’s own Safeguards principle requires an organization to make its employees aware of the importance of protecting personal information, and to match the safeguard to the sensitivity of the data: a locked cabinet and restricted access for paper, passwords and encryption for anything digital. The same principle covers the exit as much as the entry: the safeguards clause separately requires care in the disposal or destruction of personal information, so nobody can reconstruct it from what gets thrown out — worth saying out loud the first time somebody drafts on a printed page and puts it in the regular recycling.

In a law firm there is a further layer: everyone in the practice is inside the confidentiality obligation. The conduct rules require a lawyer to exercise due diligence in ensuring that each employee, and every other person whose services the firm retains, complies with the confidentiality requirements, which includes the people who never speak to a client.

Verification is a task step, not a virtue

The single highest-value habit to install is checking, and it only sticks when it is part of the task rather than an attitude. Concretely:

  • • Every figure in a drafted document is traced back to the source record, not to the draft.
  • • Every reference or citation is opened, not skimmed.
  • • Every name, date and amount is checked against the file, because these are exactly what a fluent draft gets confidently wrong.
  • • Anything leaving the firm passes a second pair of eyes belonging to whoever’s name is on it.

Teach the specific failure mode too: a model produces text that reads as though it has been checked. Fluency is not evidence. New users over-trust polished output, and that is a training problem with a known fix — make them find a planted error in the first session.

Supervision does not move

Support staff in a professional practice work under supervision, and a tool does not change that. In Ontario a lawyer must assume complete professional responsibility for the practice and directly supervise non-lawyers to whom tasks are assigned, reviewing the work at frequent intervals. Say this out loud in the training session: the point is not that the administrator is untrusted, it is that accountability sits with a named licensee and the workflow has to make that possible.

Monitoring the rollout without tripping over the ESA

Firms often want usage data to see whether the training worked. If you employ 25 or more employees in Ontario on 1 January, you must already have a written policy on the electronic monitoring of employees, in place before 1 March, stating whether you monitor, how and in what circumstances, and the purposes for which the information may be used. Turning on tool-level usage reporting without updating that policy is an avoidable problem, and the drafting requirements are set out here. The policy has its own distribution clock too, separate from the 1 March creation deadline: a new hire must be given a copy within 30 days of becoming an employee, and if you update the policy after a rollout like this one, every current employee is owed a copy of the new version within 30 days of the change.

Training time itself is generally paid work time in Ontario — see the position on mandatory training before you schedule a session at seven in the morning.

A worked example

An 11-person accounting practice with two administrators and a receptionist runs one 90-minute session and four 20-minute clinics over the following month. The session covers the list, the verification steps and the escalation triggers, using the practice’s own documents. The clinics are held over the real work: the tool is open, a live task is on screen, and the manager watches what people actually do.

Three rules go on one page pinned above each desk: approved tools, never-paste list, and who to ask. The firm measures one thing at the end of the quarter — the number of corrections made at review on the work those three people produce. If it falls, the training worked. If it rises, the rollout was too wide, and the fix is scope rather than more slides.

Questions we get asked

What if somebody pastes client data into a public tool?
Treat it as an incident, not a disciplinary event on the first occasion: contain it, record what was disclosed and when, and involve whoever owns privacy in the firm. That record is not optional paperwork — PIPEDA requires an organization to keep a record of every breach of security safeguards involving personal information, however minor, and to produce it to the Privacy Commissioner on request. If it is reasonable to believe the breach creates a real risk of significant harm, the duty escalates: the organization must report it to the Commissioner and notify the affected individual, in each case as soon as feasible. Treadstone Law’s guide to mandatory breach reporting under PIPEDA walks through how that risk assessment is actually made — two of the factors the Act names are the sensitivity of the information involved and the probability that it has been or will be misused, which is exactly why pasting a client’s name into a chatbot and pasting their SIN and account numbers are not the same incident, even though both are mistakes. The OPC’s principles expect openness about what happened to the information, and a firm that punishes the first report will not hear about the second.

Should we buy a generic AI course?
Generic literacy is cheap and mildly useful. The part that changes outcomes is firm-specific: your systems, your data, your escalation path. No external course will contain those.

How do we know the training worked?
Not by a quiz. By the correction rate at review on the work those staff produce, measured before and after. It is the only signal that survives contact with a busy month.

See where AI pays off first in your practice.

A 30-minute call is enough to tell you whether AI pays for itself here.