■ MODERATE RISK ■ Public Service & Government
Not replaced, but demoted from oracle to operator. The instinct you sold at $30k a month is now a dashboard, and the surviving strategist is the one who can decide what to do when the dashboard says something the candidate hates.
“AI analyzes polling data and sentiment. Your 'gut read' on voters was just vibes.”
Our AI replacement risk score — how we score jobs
Strategy work is less glamorous than the movies suggest. It is message testing, crosstab reading, media buy allocation across DMAs, candidate prep, opposition research, coalition arithmetic about which 40,000 voters in which three counties actually decide the thing, and an enormous amount of managing egos — the candidate's, the donors', the county chair who thinks the yard signs are wrong. Add crisis response at 11pm when a decade-old post surfaces.
Automation has moved fast here because the inputs are digital and the feedback is numeric. Models now generate and test hundreds of message variants against synthetic and real panels, cluster voters far past the old soccer-mom archetypes, forecast turnout from voter files, and rewrite the same appeal in eleven registers for eleven microaudiences. Sentiment tooling watches broadcast, social and local news in real time. Ad platforms optimize creative automatically, which quietly deletes the media buyer's judgment. The uncomfortable part is that these systems are demonstrably better than veteran intuition at the things veterans bragged about: predicting which frame moves persuadables, and spotting a district shifting before the yard signs do.
What survives is contested and human. Campaigns are coalitions of people who dislike each other, and someone has to broker that. Someone has to tell a senator that the honest answer polls badly and they should give it anyway, or that they should not. Ethics, legality and reputational limits on tactics are judgment calls with prison-adjacent downside. Relationships with reporters, unions, party committees and donors are held by people, not accounts. And a campaign is a live-fire environment where the model's training distribution shifts weekly. Our risk score of 29 says the analytical core is largely automatable while the political core — persuasion of the people inside the room — is not.
Automatability: our editorial assessment of current and near-term AI capability
This is already biting at the junior end — the analyst who built decks from crosstabs has been the first casualty in cycles since the early 2020s. Through 2030 expect campaigns to run leaner war rooms with generative message factories, squeezing mid-level consultants hardest. By the late 2030s the surviving business is a small number of senior operatives with relationships and risk judgment, sitting on top of largely automated analytics shops.
It can run most of the machinery — targeting, creative production, buy optimization, rapid rebuttal drafting — but not the decisions with legal, ethical or coalition consequences. Campaigns fail on candidate discipline and internal knife fights far more often than on bad segmentation, and those failure modes have no software fix.
Entering is harder than it was. The old apprenticeship ran through data grunt work that now barely exists, so build your way in through field organizing, comms or direct candidate work instead. The senior tier is intact and well paid, but the ladder to it has fewer rungs than it did a decade ago.
Deck building, poll memo writing, message drafting at volume, media plan construction and routine opposition-research compilation. Anything with a repeatable output format and a numeric evaluation is already being handed to a model, and clients who once paid for that volume now expect it bundled.
Assume every cycle now includes fabricated audio or video of your candidate, and build the response playbook before you need it: verification partners, prepared statements, a rapid forensic contact, and a press relationship that will take your call at midnight. Speed of credible denial matters more than the quality of the fake.