Brief
Audience, intended decision, and the change in understanding the presentation must achieve.
A product thought experiment for Heim
PowerPoint should help people think before it helps them present.
My response to the question: which AI feature would I remove, and where would I add a model instead?
Read the case ↓01 The feature I would remove
I would remove “create a presentation from a document” as PowerPoint’s primary AI promise.
A report can become a credible-looking deck before its author has made the hard choices: what outcome is needed, what the audience needs to believe, what evidence matters, and which trade-offs must be visible.
That is not a simplification problem. It is an argument-quality problem. The product rewards document conversion over judgement. A polished deck can conceal a weak narrative, untested assumptions, or claims with no clear evidence behind them.
Next: the model I would add ↓02 The model I would add
An evidence and audience challenger, invoked before the deck is shaped.
The author declares the audience and intended decision. The model maps the argument, tests it against the source material, then makes gaps visible. It never silently repairs the story; it proposes questions the author must answer.
01Which material claims lack a source, a date, or an accountable owner?
02Which alternative, risk, or assumption has been omitted from the argument?
03What would a finance leader, clinician, operator, or engineer challenge first?
04Is the desired decision explicit, and is there enough evidence to support it?
03 An evolution, not an anti-AI position
The goal is not fewer models. It is better placement of intelligence in the work.
Argument map
Claims, evidence, assumptions, risks and alternatives made explicit before a slide is designed.
Challenge
A role-aware model asks the questions a sceptical audience is likely to ask.
Craft
The author chooses the narrative and expression; the system protects clarity, provenance and accessibility.
Learn
Questions, decisions and outcomes inform the next version — whether the artefact is a deck or a living web brief.
At each stage, the model improves the quality of the human decision rather than simulating one.
Next: design principles ↓Design principles
Start with the decision
A presentation exists to help a specific audience understand, decide or act — not to turn a document into slides.
Keep the evidence visible
Substantive claims need provenance, recency and an honest account of uncertainty or disagreement.
Keep the author accountable
The model can reveal gaps and test the argument. A human remains responsible for the judgement and the decision.
04 Why this matters for healthcare
In high-consequence work, fluent output is not evidence of a safe or useful decision.
My inference is that the same principle matters for Heim: use models to surface relevant information, structure trade-offs and accelerate operational action—while preserving professional judgement, clear escalation and traceability.
The specific clinical and operational boundaries must be co-designed with practitioners, clinical leadership and users. That is a product question, not something a model should decide alone.
Next: questions to test ↓Questions I would want to test
- Where do frontline teams currently lose the most time converting information into a safe action?
- Which decisions benefit from a challenge layer, and which need a simple operational hand-off instead?
- What evidence, provenance and escalation signals would earn clinicians’ trust?