The room

It sounds like the tidy-up session. It is not. Look at what the four design levers are actually doing: they help a person who can read the slide. The accessibility lever changes who can read the slide at all, and in a room of twelve that is not a rounding error.

Run the ladder back in your head. Headline 27 to 48. Visual evidence to 64. Coherence to 73. Cue the hero number to 82. Then nine points, from one thing: the text finally clears the contrast standard. Every lever before tonight was a judgement call you could argue about. This one is arithmetic with a published threshold.

A speaker note is a glance, not a text. You look down for about a second and a half. Anything that needs a second look is not a note, it is a document, and it belongs in the appendix.

• The cue. Three or four words that restart your brain. "EMEA is the whole gap." Not a sentence, a hook.
• The transition. One fragment naming where you are going next, because the joins are where people lose the thread of an argument, not the slides.
• The one number you must not misquote. Written out exactly, with its unit and its period. This is the only place verbatim text belongs.

Everything else that Copilot generated, delete. You are not deleting information, you are moving it out of the one place where reading it costs you twenty points.

Then read your own notes back with one question: could I say this without looking down twice?

Generating notes for all slides in one pass is the genuinely useful version of the feature, and Microsoft states it works best under 50 slides. If your deck is longer than that, the honest read is not "the feature is broken", it is that your deck is two decks.

Two practical habits that follow from the limit:

One more thing worth knowing before you present from a laptop you do not own: notes live in the file, so they travel with an export only if the export format carries them. A PDF handout does not.

Take the list and cut it to the three questions that would actually embarrass you. Not the three most likely. The three you would rather not be asked, which is a different and much shorter list.

Then do the thing almost nobody does: build each answer as a real slide and put it in the appendix. Titled, sourced, laid out properly. Session 2 taught the appendix as a destination rather than a dumping ground, and this is what it was a destination for.

The difference in the room is large and specific. "I will follow up on that" ends a thread and loses momentum. Jumping to a slide that already answers it does three things at once: it answers, it shows you had already thought about it, and it moves the meeting on.

One honest limit: a model anticipating audience questions is generating plausible objections, not reading your stakeholders. It is very good at finding the unsupported claim on the slide. It cannot know that the finance director has been burned by this vendor before. You still supply the room.

WCAG 2.2 success criterion 1.4.3 sets a contrast minimum: 4.5:1 for normal text, and 3:1 for large text, where large means 18pt, or 14pt bold. That is the whole rule you need to carry.

Two details that people get wrong and then argue about:

• Ratios are never rounded. 4.499:1 fails. There is no "close enough" band, which is the entire point of having a number rather than an opinion.
• Large text is a real exemption, not a loophole. Your section headers may legitimately sit at 3:1. Your body copy, your chart labels and your source lines may not, and those are exactly where AI-generated decks put pale type.

Session 1 introduced this and told you it would become a gate. This is the gate.

Titles and alt text are obvious once said. Reading order is not, and it is the one that quietly breaks AI-generated slides.

A screen reader does not read your slide the way your eye reads it. It reads elements in the order they were added, not in the visual order they ended up in. So a slide where you dropped the chart in first, then added the headline above it, then squeezed a caption in beside it, reads as: chart, headline, caption. Which is not the argument you built.

This matters more after a generation pass than after hand-building, because you did not add those elements and you have no idea what order they went in. Check it, do not assume it.

Say this out loud in your own words the first time someone assumes otherwise: no AI deck tool claims WCAG conformance for its output. Not one. This is not a gap someone is about to close for you.

Canva states it plainly in its own help documentation: "No automated accessibility checker can guarantee accessibility compliance as it requires manual testing to ensure it works with all assistive technologies." That is a vendor telling you, correctly, that the last mile is yours.

Gamma has completed no formal screen-reader audit and makes no specific compatibility claims. That is not an accusation, it is the state of the documentation.

So this is a gate you own, permanently. Four items, one checker run, every deck, for the rest of your career. It takes about six minutes on a twelve-slide deck once you have done it twice.

Not a full re-review. Four minutes, three things, every time.

If you are presenting from someone else's machine, send a PDF as well and stop worrying. A PDF loses your notes and your animations, and it keeps your layout, which on a strange laptop is the better trade.