| Advantage | Evidence level | Where it is credible | Where to be careful |
|---|---|---|---|
| Faster ideation | Proven for some tasks | Drafting, clustering, rough routes, visual exploration, low-fidelity previsualization | More options can increase indecision and review cost |
| More creative variations | Proven as a capability | Copy, crops, backgrounds, aspect ratios, product contexts, language variants | Volume is not originality or effectiveness |
| Lower early-concept cost | Emerging / often real | Replacing some rough boards, temp art, or exploratory renders | Final production, clearance, and editing costs remain |
| Rapid prototyping | Proven | Testing a concept before a shoot, building a clickable experience, trying a new visual system | Stakeholders may mistake prototype polish for final feasibility |
| Personalization | Proven as infrastructure | Dynamic creative, product recommendations, local offers, user-generated invites | Privacy, discrimination, consent, and “creepy” relevance |
| Localization and translation | Proven as workflow assistance | First-pass translation, subtitles, dubbing, resizing, version matrices | Native review is not optional; humor and culture are not literal |
| Accessibility | Proven in targeted uses | Captions, transcripts, audio description drafts, alt-text drafts, translation | Verify quality and avoid making disabled users the test population |
| Small-team capability | Proven as access | A founder or local business can explore more directions cheaply | Skills, taste, rights, and time still determine quality |
| Faster testing | Proven for asset supply | More headline/layout/CTA variants in controlled tests | More cells can fragment spend and weaken statistical power |
| Workflow automation | Proven | Routing, naming, resizing, tagging, versioning, reporting | Model changes and silent failures require logs and owners |
| Difficult/impossible visuals | Emerging but credible | Previsualization, imaginary worlds, synthetic locations, historical or fantastical scenes | Physical and human plausibility may still fail |
| Media optimization | Proven as platform behavior | Bidding, placement, delivery, pacing, creative rotation | Platform-reported “better performance” is not necessarily incrementality |
| Customer engagement | Proven in bounded tasks | FAQ, triage, product discovery, lead qualification with handoff | Claims and commitments require grounding and escalation |
| New interactive storytelling | Emerging | User-specific invitations, conversational ads, AI characters, adaptive narratives | Moderation, privacy, latency, cost, and disclosure are product work |
| “AI improves ROI” | Mostly promotional unless independently tested | Can be true in a defined pilot or case | Do not generalize from vendor case studies or platform averages |
Independent productivity studies support the idea that AI can amplify work under the right conditions, but they also show why the use case matters. In the NBER customer-support study, an AI assistant increased issues resolved per hour by 13.8% on average, with gains of up to 35% for lower-skilled and less experienced workers, while the highest-skilled workers saw little or no gain.6 In the Harvard/BCG experiment, consultants using GPT-4 completed 12.2% more subtasks, worked about 25% faster, and received higher quality ratings on tasks inside the capability frontier, while performance deteriorated on a strategy task outside it.7 METR’s randomized study of experienced open-source developers found a 19% slowdown with early-2025 AI tools in its setting; METR later said newer data was too confounded to provide a reliable current estimate.813
The advertising implication is straightforward: evaluate task fit, not an abstract “AI productivity” number.
Follow the evidence
This chapter comes from the September research notebook. Linked sources and qualifications remain attached to the claims.
Open all sources and footnotes →