The honest answer is that genAI is collapsing the parts of agency work that used to justify the headcount, while making the parts that remain more valuable, and most agencies haven't reorganised around that yet.
On the creative side, the biggest shift is that production and ideation are no longer separate phases. A strategist can now generate mood-boards, script variants, storyboards, rough cuts, and voiceover demos inside a single working session. What used to be a two-week loop between strategy, creative, and a production house becomes an afternoon. That sounds like a productivity win, and it is, but it also dissolves the traditional "creative department" handoff structure, the deck-to-comp-to-shoot pipeline that a lot of agency P&Ls were built around.
The second shift is variant volume. A campaign that historically shipped three hero executions plus a handful of cutdowns can now ship hundreds or thousands of personalised variants, different messages for different audiences, channels, languages, moments. This is the part clients actually want, because it maps directly to performance. But it requires a different creative discipline: you're not designing an ad, you're designing a system that generates on-brand outputs within constraints. That's closer to product design than to traditional art direction, and most agencies don't hire for it.
On the media side, the change is more about agents than about images. Bidding, audience construction, budget pacing, and creative-to-placement matching are increasingly handled by model-driven systems inside the ad platforms themselves (Meta's Advantage+, Google's Performance Max, TikTok's Smart+). The agency's value moves upstream, into feeding those systems good creative, good first-party signal, and good measurement, and downstream into interpreting what they did. The middle layer, the "media planner who pulls levers in a DSP," is the role under most pressure.
Research and insights are quietly being rewired too. Synthetic audiences, AI-moderated qual interviews, and instant analysis of open-ended survey data mean strategy teams can test concepts in hours rather than weeks. The risk is that "ask the model what a 28-year-old in Lagos thinks" becomes a substitute for actually talking to one, and the agencies that win will be the ones that use AI to do more primary research, not less.
A few second-order effects worth naming. Junior roles are getting hollowed out, the work that used to train a 22-year-old (resizing assets, building decks, pulling competitive scans, drafting first-pass copy) is exactly what models do well, which creates a real pipeline problem for the industry five years out. Homogenisation is a genuine creative risk, when everyone prompts similar tools with similar references, outputs converge, and distinctiveness becomes the actual scarce resource. IP and provenance are unresolved, clients are increasingly asking who trained the model, what it was trained on, and whether the output is indemnifiable, and "we used Midjourney" is no longer an acceptable answer for a brand campaign. And pricing is the elephant, if a deliverable that used to take 200 hours now takes 20, the hourly model is mathematically dead, which is why outcome and subscription pricing keep coming up even though most agencies haven't figured out how to actually run them.
The agencies pulling ahead aren't the ones using AI to do the same work faster. They're the ones using it to take on work they couldn't profitably touch before, always-on personalisation, real-time creative optimisation, productised offers, and treating the freed-up senior time as room for sharper strategy and craft, not as a margin grab.
