The only independent measurement
There is surprisingly little independent data on commercial deck generators — most of the numbers come out of vendor announcements. A university study from March 2026 fills exactly that gap, and it tests commercial products too, not just research models.
- 238 test cases built by experts, with an average of 54.1 checklist items each.
- Five domains: academic talks, teaching, economics, conference presentations, advertising.
- An average of 22,000 tokens of background material per input — roughly 34 pages of source.
- Five dimensions: presentation fundamentals, visual design and layout, content completeness, content correctness, and fidelity to the source.
Across the eight systems tested, the overall scores spread between 35.9 and 62.5 points (on a hundred-point scale). The field was led by a note- and source-based system; the best-known commercial tool built specifically for decks scored 49.2.
An average around fifty points on a checklist that human experts put together. That is not a failure — but it is not what the demos promise either.
Where it is good: structure
On the "presentation fundamentals" dimension the systems landed between 53 and 81 points — by a wide margin their strongest area.
This covers the logical build, the agenda, the opening and closing sections, one idea per slide, balanced sectioning. In other words, the layer most people think of as their own work is the one the machine handles best.
There is a reason for that: structure is a language task. Pulling the main ideas out of a thirty-page document and arranging them around twelve slides is the same ability that works in summarising.
Where it is bad: layout and source fidelity
Here come the two pieces of bad news, and the second one is the heavier.
First: visual layout is the main bottleneck. The study calls it exactly that. Even the best system only scored 62.8 points here, and most of the field stayed under 50. This is not about the choice of colours; it is about whether the text fits, whether two elements sit on top of each other, whether the slide has a hierarchy that makes sense.
So it is the slide that looks bad, not the argument.
Second: source fidelity. This dimension measures whether the specifics that end up on the slides — numbers, names, claims — really come from the background material you handed in. Even the best system only reached 45.1 points.
Content correctness lags behind completeness anyway: for the leading system completeness is 67.8 and correctness 56.0. In plain terms: more goes onto the slides than is true.
If you had to sum up what to expect in a single sentence: the machine does build a deck out of your material, but roughly every second specific written on the slides does not come from your material. That is the kind of mistake a client presentation does not forgive.
What the vendors write down themselves
The documentation is worth reading, because more limits are spelled out there than anywhere else. The most important ones:
- Your own template can only go in at the front. The documentation says it in so many words: if your organisation uses a standard template, start from that file and generate afterwards — the machine will use the layouts that are already there. It will not pull one over a finished deck later.
- One output language at a time. Languages cannot be mixed while a presentation is being generated.
- Quality is highest in English, which the vendor documentation states itself, and the design suggestions are officially supported in English only. The newer slide-generation features that have been announced also launch in English.
- Rewriting only affects text boxes, not shapes or other objects, and it rewrites the whole box, not the part you selected.
- There is a size limit on the source document, and the amount of text processed per prompt is capped as well — for the latter no concrete number is published.
- Three separate disclaimers appear in the documentation, all saying that generated content may be inaccurate, misleading or out of place.
That list on its own tells you where the machine belongs in the process.
What nobody measured: whether it is faster
This matters, and it is rarely said out loud. There is no controlled measurement of how much faster a deck gets built with AI.
The best thing available is self-reported data from a large public-sector rollout: participants reported saving an average of 19 minutes a day on building presentations. That is a survey estimate, not stopwatch time — and the same survey also reports that 17% of respondents saw no saving at all.
What we do know is that in experiments on the office day the self-reported saving runs systematically higher than the measured one — the real time savings post goes through that in detail, broken down by type of task.
The recipe
The division of labour follows straight from the measurements — just not the one you would expect.
1. START FROM THE TEMPLATE
Open the company template file and generate inside it.
It cannot be pulled over afterwards.
2. LET IT WRITE THE OUTLINE — THIS IT CAN DO
Twelve slide titles and key points out of thirty pages.
This is its strongest measured ability.
3. PUT THE SPECIFICS IN YOURSELF
Number, name, date, client claim. Source fidelity is
45.1 even in the best system.
4. REVIEW THE LAYOUT SLIDE BY SLIDE
The weakest measured dimension. Text running over,
elements overlapping, sizes nobody can read.
5. ONE SLIDE, ONE CLAIM
Anything you cannot say about the slide in two
sentences is two slides.
Step three is the most important and the easiest to skip. On a generated slide the number looks exactly like all the others — nothing marks it as something the machine added.
One practical trick: ask it to note beside every specific which page of the source it took it from. Anything it cannot give a page number for gets cut. This is the same defence that works with meeting notes, which the meeting notes post covers at greater length.
Common misconceptions
- "It will take care of the design." According to the measurement that is precisely its weakest dimension. Structure it can do; layout it cannot.
- "It works from my own material, so it is accurate." Source fidelity is 45.1 points even in the best system. A large share of the specifics that reach the slide does not come from the material you supplied.
- "I will put the company template on at the end." That is not how it works. You have to start with the template, otherwise the generated layout is what stays.
- "It is the same in any language." The vendor documentation states that quality is highest in English, and several slide-generation features are officially available in English only.
- "It gets done in half the time." There is no controlled measurement for that. What exists is self-reporting — and on this topic self-reporting demonstrably errs upwards.
What to take away
With decks the division of labour works the opposite way from how it is advertised. The line of argument you can safely hand over — the layout and the facts you cannot.
Out of that comes a simple working order: template, outline by machine, specifics by hand, layout slide by slide. Skip the third step and you get a well-built, nicely paced, confident presentation with wrong numbers in it — and you will find that out in front of the audience.