Consultants use AI on the production of a deliverable, not the judgement inside it. Where it earns its place, what has to stay human, and who ends up owning the result.
Key takeaways
A consulting deliverable is four things stacked on top of each other: the evidence you gathered, the reasoning you applied to it, the recommendation you are prepared to defend in front of a board, and the artefact that carries all three — a report, a model, a deck, a roadmap. Only the first and the last of those are production work.
AI is production capacity. It shortens the distance between having the evidence and having the artefact. It does not shorten the distance between the evidence and the recommendation, because that distance is the thing the client is actually buying.
That framing has a commercial edge. If your fee rests on the artefact — page count, slide count, hours of formatting — cheap production erodes it. If it rests on judgement, cheap production protects it.
Almost everything a consultancy handles belongs to somebody else, and some of it is personal information. The Office of the Privacy Commissioner has published principles for responsible, trustworthy and privacy-protective generative AI, the first of which is that every party should know and document its legal authority for collecting, using, disclosing and deleting personal information across the life of the system — and that where consent is that authority, it should be specific and meaningful.
The same guidance asks for openness: telling people what personal information is collected, how, when and why, and making sure outputs that could significantly affect someone are identified as having been produced by a generative tool. Underneath sits the Personal Information Protection and Electronic Documents Act, whose fair information principles include limiting collection and limiting use — the two principles an enthusiastic pilot breaks first.
On the vendor side, Innovation, Science and Economic Development Canada maintains a voluntary code of conduct on the responsible development and management of advanced generative AI systems. Whether or not a supplier has signed it, the commitments in it make a serviceable list of questions to put to one.
Contractually, the confidentiality obligation you already owe does not bend for a new tool. The distinction between a confidentiality clause and a standalone NDA is worth understanding before you find out the hard way what a breach of that clause exposes you to.
Ownership is a contract question before it is a technology question. The Copyright Act frames ownership around an author, and section 13 sets out who the first owner is in the ordinary cases — including the rule that a work made in the course of employment generally belongs to the employer, absent agreement. Canadian law has not settled how authorship applies to material generated by a model, which is a reason to write the answer into your contract rather than leave it to a default.
The default that catches small firms most often is the employee-versus-contractor line, because the two are treated differently and consultancies use subcontractors constantly. The clean fix is an express assignment; what an IP assignment clause has to do is a short read and will save an argument at the end of a project.
While you are in the contract, look at your liability cap. Limitation of liability and indemnity clauses are the terms that decide what a bad deliverable costs you, and they are not automatically enforceable. No tool changes that analysis.
A four-person operations consultancy runs a three-week review of a distributor’s warehouse and order desk. Inputs: 22 interview transcripts, three workbooks, 18 policy documents, and a fortnight of the partner’s own site notes.
Names and identifying details are stripped from the transcripts before anything is uploaded, and the workbooks stay out entirely because they carry customer records. The tool produces a coded set of themes with quotes attached. The partner rejects two of the six — one was an artefact of who happened to be interviewed, the other conflated a symptom with its cause. That rejection is the engagement.
The agreed structure goes on one page before a word of prose is written. Sections are then drafted from the team’s own analysis notes, and a final consistency pass catches an order-volume figure stated three different ways across the report.
What did not move: the recommendation to consolidate two shifts, the sizing of it, and whose name is on the cover. If that recommendation is wrong, the claim that follows is against the firm, not the software.
Pick the deliverable type you produce most often and automate only the synthesis and the consistency pass for one full cycle. Leave structure and argument alone. Then look at whether the senior time you freed actually went into judgement or simply disappeared — if it disappeared, the tool was not the problem.
If drafts are already flowing but nothing catches them before they reach a client, our article on building a sign-off workflow so AI drafts never go out unchecked covers the control layer, and there is a companion piece on rolling out AI drafting tools without losing partner buy-in.
Should we tell the client we used AI?
Look at your engagement terms first. Many consulting contracts contain confidentiality and subcontracting provisions that already answer the question, and a client who has restricted where their data may be processed has effectively answered it for you. Silence is a decision, so make it deliberately.
Can AI write the recommendation if we give it all the evidence?
It can write a recommendation. It cannot take responsibility for one, and it has no way of knowing which of your client’s constraints are real. The recommendation is the deliverable; everything else is packaging.
Does this make our work product less defensible?
Only if you cannot say what you did. Keep a short record on the file of which steps used a tool and what a person checked afterwards. That record is the difference between a defensible process and an argument about one.
A 30-minute call is enough to tell you whether AI pays for itself here.
