AI is fine for assembling the deck. The mistake is letting it do the analysis.
- Workflow that works: export the raw metrics to a sheet (Meta, Google, TikTok exports, last 30 days), write 3 to 5 bullet observations yourself first, the ones you could defend in the meeting. Then hand the model your numbers plus your bullets and ask it to structure slides: headline, one chart, your observation, one recommended action. AI formats, you think.
- Guardrails: paste actual numbers and never let the model recall metrics from memory, require that every slide headline be a conclusion not a label ('CTR fell 18% on prospecting' instead of 'CTR overview'), and aggregate spend data before it touches the model if you would not email it.
- Tooling: ChatGPT or Claude plus a slide template (Google Slides API or python-pptx for the layout) covers about 80% of it. The remaining 20% is your commentary on why the numbers moved. If the deck takes 4 hours now, this cuts it to 90 minutes without dropping quality.
Rule of thumb: AI drafts the deck, you write the three bullets that matter. If you cannot write those bullets without the model, the problem is the analysis, not the tooling.