The Four Kinds of AI Turning Up in Campaign Work

Every few weeks a new product lands in the inbox of a campaign director or a community organiser wearing the same two letters. The pitch varies, the price varies, but the claim is constant: this is AI, and your team is falling behind without it. The claim is nearly always true in a narrow sense and useless in every sense that matters, because AI is not one thing. In campaign and community work it is at least four things, and they carry different costs, different risks and different questions.

Sorting a product into the right category is the fastest piece of due diligence available to a small team, and it can usually be done from the sales page alone. It is worth doing before the demo call, because the demo is designed to make the categories blur.

The conversational assistant

The first category is the one most people have already met: a large language model behind a chat window. You type, it drafts — summaries, translations, restructured arguments, the fourth version of a thank-you email while the second coffee is still warm. For small organisations this is where most of the genuinely useful work lives: turning a rambling meeting transcript into three action items, condensing an eighty-page planning report the night before a council deputation, tightening a media release without flattening its voice.

Its defining property is that it does nothing on its own. Every output passes through a person on its way to the world, which is exactly what makes it safe to learn on. The risk sits in what you paste in, not in what the tool might do — so the sensible opening rule is boring: no supporter records, no member details, no internal strategy in a consumer chat window until somebody has actually read the vendor's data terms.

The system that acts

The second category does not wait for copy and paste. Agentic systems take actions: they send the email rather than drafting it, update the record rather than suggesting the update, chain a dozen steps together overnight while nobody watches. The productivity story is real. So is the change in the question you must ask. With an assistant, the question is whether the draft is any good. With an agent, the question is what, precisely, this thing is permitted to touch.

The discipline that follows is least access and a human gate. An agent that reads a shared calendar is a different proposition from one that can message a supporter list, and anything public-facing or irreversible deserves a person standing between intention and execution. Teams that skip the gate tend to discover its purpose retroactively.

The feature that was already there

The third category never arrives as a decision at all. It is the AI folded into software you already run: the spreadsheet offering to summarise, the design tool generating imagery, the email platform proposing subject lines, the video call transcribing itself. This is the cheapest AI to adopt and the easiest to overlook, and it quietly rewires your data flows. A meeting that was once ephemeral now produces a transcript on a vendor's servers; a drafting pane now sends fragments of your writing somewhere on every keystroke. Yesterday's privacy posture does not automatically survive today's product update, so the audit that matters here is not of new tools but of old ones.

The tool built for the movement

The fourth category speaks your language. Purpose-built movement tools — fundraising drafters tuned to progressive values, organising databases with automated supporter journeys, monitoring services shaped around civic questions — are sold in the vocabulary of the work itself, which is precisely what makes them hard to evaluate coolly. A sales page fluent in your own convictions has a way of disarming scrutiny.

The clarifying move is to ask which of the first three categories the product is underneath. Almost every movement tool turns out to be a conversational model with a tuned voice, an agent wired into a database, or a feature bolted onto infrastructure you would recognise from elsewhere. That is not a criticism; it is a map. Name the underlying category and you inherit its questions — what does it touch, what does it store, who reviews the output — and the positioning stops doing your thinking for you.

The model matters less than the habit

Beneath all four categories sits the question of which model to use, and it matters less than people fear. The bigger, dearer models justify themselves when the work is genuinely hard: a hundred-page submission held in mind at once, a piece of writing where tone carries the argument, a chain of reasoning whose steps depend on one another. For the daily traffic of quick drafts, summaries and reformatting, smaller models do the job at a fraction of the price. There is no permanent winner, and the leaderboard will have churned again before your subscription renews.

What compounds is not the subscription but the practice. A team that writes clear briefs, keeps a person reviewing everything consequential and treats each output as a first draft will beat a better-funded team with lazier habits on any model either can afford. The four categories are worth learning precisely because they tell you where the review belongs. The review — not the model — is the part that was always yours.