AI consultant NDA basics: what to have signed before anyone watches your work

An NDA with an AI consultant has to cover more than a standard template does, because the consultant does not only read your documents. They watch the work, record screen shares, get logins, and may run what they collect through AI tools of their own. The basics are five clauses: a definition of confidential information that includes what they observe and record, a rule on which AI tools and subcontractors may see it, a ban on training or reuse, return or deletion at the end, and separate terms for any personal information.

Everything else in a normal NDA still applies. Those five are the ones a template written for a sales meeting usually leaves out.

Why an AI engagement is different from an ordinary one

A bookkeeper or an IT contractor sees your systems too, so the instinct is to reuse whatever NDA you already have. The difference is in how the work is done.

Diagnosing which jobs to automate means watching them. That means recorded screen shares of real orders, real invoices and real customer emails, because the process as it survives in a business is never the version written in the manual. It means read-only access to the software you pay for. It often means sample exports so a build can be tested on real records rather than invented ones.

Then there is the second question an ordinary NDA never had to answer: what the consultant’s own tools do with it. A consultant who pastes your supplier price list into a general chatbot to summarise it has disclosed it to a third company, under that company’s terms. Nothing about that is unusual in 2026, and a template NDA written before these tools existed does not mention it.

The five clauses to check

1. What counts as confidential

Most templates define confidential information as material you disclose, sometimes only material marked “confidential”. Neither covers the recording of a screen share, the notes somebody takes while watching a Barrie dispatcher, or the process map built from them.

Look for wording that includes information the consultant observes, records or derives during the engagement, whether or not it is marked. The standard exclusions are normal and fair: information that is already public, that they already had, that they developed independently, or that they received lawfully from somebody else. Those should stay.

2. Which tools and people may see it

This is the clause most likely to be missing. It should say which categories of AI tool may process your information, on what kind of account, and whether a subcontractor may see it.

A reasonable version names business or enterprise tiers whose terms exclude training on customer data, and says consumer tiers are not used for your material. If the consultant works with a developer or a second firm, the clause should say that person is bound by the same terms before they see anything. You are not trying to ban tools. You are trying to know where your information goes, which is the same question the privacy rules ask.

3. No training, no reuse

Your process documents, recordings and exports are used for your engagement only. They are not used to train or tune a model, not used as examples for other clients, and not published as a case study without your written approval.

Watch for “anonymised” language. A clause that lets a consultant reuse your material once identifying details are removed is common. In a small town like Collingwood, or a narrow trade, a process description with the name taken out can still identify you. If you accept that clause, ask to approve anything before it is used.

4. What happens at the end

The NDA should say what the consultant returns or deletes when the engagement ends, and that includes the copies people forget: screen recordings, meeting transcripts, sample exports sitting in a shared drive, and the history of any AI tool they used on your material.

Deletion “on request” is weaker than deletion at the end by default. Ask for a written confirmation when it is done, and for any logins they held to be removed on the same day.

5. Personal information is a separate question

An NDA protects business information between two parties. It is not the agreement that governs your customers’ or your staff’s personal information, and signing one does not make a disclosure lawful.

The Office of the Privacy Commissioner of Canada is direct about this. Under PIPEDA you remain responsible for personal information you transfer to a third party for processing, and where that processing happens outside Canada, you are expected to ensure through contractual or other means that the third party provides a comparable level of protection and to limit its use to the purposes of the contract. The broader picture is in PIPEDA and AI compliance in Canada.

In practice that means one more section, or one more short agreement, that says what personal information the consultant will see, for what purpose, where it is stored, who else processes it, and how it is disposed of. Most diagnostic work does not need customer personal information at all, which is the better answer when it is available.

Share less, then sign

The strongest confidentiality term is the information you never handed over. Before any NDA is drafted, decide how little the AI consultant needs you to prepare:

• Read-only access rather than an admin login, and a separate account you can switch off, not a shared password.
• A screen share of the work rather than an export of the database, where watching is enough to understand the process.
• Redacted samples where a build has to be tested on real records: real formats and real messiness, with customer names and numbers removed.
• A date on which every login is removed, written into the plan, not left to memory.

None of this signals distrust. A consultant who has done this before will suggest most of it before you ask.

When to sign it

The first conversation rarely needs one. Keep it to what the business does and which job hurts. The brief you send to several consultants when you work out how to compare AI consultant quotes should also be written so that it needs no NDA: the job, the systems by name, and how often it happens, with no customer data and no prices you would mind a competitor reading.

The NDA belongs before the first moment somebody sees the real thing: the first recorded screen share, the first login, the first sample export. In a structured engagement that point is easy to find. The AI business assessment for Ontario owners on this site, a fixed $999 over ten business days, starts with a kickoff call and moves to watching real jobs and a systems inventory, so the terms should be settled between the first step and the second. Whoever you hire, find that step in their plan and put the signature before it.

Red flags in the paperwork

• One-way in the consultant’s favour. You protect their methods, nothing protects your records.
• A non-solicit or a non-compete tucked into the NDA. Those are commercial terms and belong in the service agreement, where you can see and price them.
• No mention of AI tools at all from a firm whose work is AI tools.
• Ownership of the deliverables decided inside the NDA. Who owns the automation, the prompts and the documentation is a question for the service agreement, and it deserves its own reading.

The same rule, pointed inward

The terms you want from a consultant are the terms your own staff should already be working to: which tools are approved, what may never go into them, and what gets deleted. If you are asking an outsider for those in writing and have not written them down for your own team, that is the larger gap, and it is what an AI policy for a small business in Ontario is for.

This is a practical summary for business owners, not legal advice. Have a lawyer read any agreement you intend to rely on, and a privacy professional where customer personal information is involved.