A model can build the deck. It cannot decide whether the argument is true, and that is the part clients are paying for.
Key takeaways
Yes, AI can build a client presentation, and the honest description of what it does is narrower than the marketing suggests. Microsoft’s own FAQ for Copilot in PowerPoint says the feature can draft content, transform a Word file into a presentation, generate summaries and answer questions about an existing deck. That is drafting and reformatting. It is not analysis, and it is not judgment about what a client needs to hear.
So the workflow that works is the inverse of the one people try first. You do not ask a model to build a deck about a topic. You write the argument down properly in a document, then ask the tool to turn that document into slides, then spend the time you saved on the three slides that actually decide the meeting.
Microsoft documents creating a presentation from an existing file as a supported path: you open a blank presentation, choose Copilot from the ribbon and ask it to create a presentation from a file, and it drafts slides with speaker notes from that source. The same page is explicit about the limits that matter operationally — one file at a time, and you must have access and permission to use the file you reference. Its wider description of what the assistant is sits in the Microsoft 365 Copilot overview.
Read those constraints as a design brief rather than an annoyance. One file at a time means the deck is only as coherent as the single document you feed it, which pushes the work back to where it belongs: writing a findings memo that already has a spine.
A consultancy finishes a four-week diagnostic. The old habit was to open PowerPoint on the Friday and think in slides. The new sequence is: write a six-page findings memo with numbered findings, each with its evidence and its implication; have the engagement lead red-pen it; then generate the deck from the memo.
What comes back is roughly right in structure and roughly wrong in emphasis, which is the correct place to start. The lead then does three things a model cannot: reorders so the uncomfortable finding lands third rather than last, deletes the four slides that restate the brief, and rewrites every slide title as a sentence with a verb in it. The client sees a deck that argues; the firm spent its hours on the argument.
The failure mode to watch is the opposite pattern, where a partner prompts a deck into existence with no memo behind it. It looks finished, reads plausibly, and collapses the moment a client asks how a number was derived — because nobody derived it.
A consulting deck routinely contains claims: a saving, an uplift, a benchmark, a comparison with competitors. Those are representations, and the Competition Bureau is the federal body that enforces the Competition Act’s deceptive-marketing provisions. A model drafting your slides will produce a confident figure to fill a placeholder if you let it, and confident figures with no derivation behind them are exactly the exposure to avoid.
The practical control is a rule the firm can actually enforce: every number on a client slide traces to a cell in a working file, and every claim about an outcome names its source on the slide or in the notes. If a generated slide carries a figure nobody can trace, the figure comes off. Where the deck effectively advises, the ordinary professional standard of care applies to the person presenting it.
Building a deck from a client file means the file passes through the tool. Microsoft publishes a data, privacy and security document for Microsoft 365 Copilot, and reading it is a five-minute task that settles most of the questions a client’s general counsel will ask. Whatever the vendor, the obligation on your side is fixed: under clause 4.1.3 of Schedule 1 to PIPEDA, an organisation remains responsible for personal information transferred to a third party for processing and must use contractual or other means to provide comparable protection.
The Office of the Privacy Commissioner’s principles for generative AI make the same point in a different register: generative AI does not sit outside existing Canadian privacy law, and the organisation deploying it carries obligations of its own. If the client engagement has a confidentiality clause, check it before the first upload — our sister firm covers what an NDA actually restricts.
Pick the recurring deck, not the flagship one. Monthly steering-committee packs, quarterly business reviews and standard onboarding decks all have a fixed spine and change only at the edges — which is precisely where generation earns its keep. Run it for a full cycle alongside the current method and compare on two measures: hours to first complete draft, and how many slides survived to the final version unchanged.
If the deliverable you care about is a written report rather than slides, the same discipline applies but the leverage is different; and if it is the underlying spreadsheet work, that is a separate problem again.
Can it build a deck straight from a spreadsheet?
Microsoft’s documented path is from a file such as a Word document, with PDF supported for licensed users, and one file at a time. Treat a chart-heavy deck as a two-step job: produce the narrative document first, then generate.
Will our client’s content be used to train the model?
That is a vendor-contract question with a documented answer, not something to infer. Microsoft’s data, privacy and security page is the reference for its own service; for any other tool, get the equivalent in writing before a client file goes near it.
Should we tell the client the deck was AI-assisted?
Ask what the engagement letter and your own regulator require, then default to candour about method when a client asks. What is never acceptable is presenting an unverified generated figure as a finding.
A 30-minute call is enough to tell you whether AI pays for itself here.
