What AI can replace from your marketing agency, and what it can't

The agency retainer got cut by two-thirds in the last budget review. AI tools now draft the product descriptions, the email sequences, the ad copy variants, and most of the monthly report that used to take the agency a week to pull together. Three weeks later, a social post goes out congratulating a competitor's product launch by name, in a tone nobody who knows the brand would have signed off on. Nothing in the AI tooling was broken. Nobody who understood the brand looked at it before it published.

The question teams actually ask is "agency or AI". The more useful question is which specific deliverables move, one at a time, and which ones still need a person who owns the brand before they ship.

This isn't a hypothetical shift. Marketer surveys this year show a strong and growing expectation, on both the client and the agency side, that AI will take over some share of what an agency currently does. The open question for any specific team isn't whether that happens; it's which share, and which deliverables are safe to move first.

What AI already does reliably

A first draft of a product description, an email sequence, a set of ad copy variants, or a batch of social captions is squarely in reach: usable in minutes, and cheap to redo if the first attempt misses. These are templated, repetitive, and reversible, which is exactly the profile of work that doesn't need a person checking every output.

The same is true for prediction work run on data the company already owns: churn-risk scoring, flagging an account's likely next order date, tuning a PPC bid within budget rules already set by a person. None of this is new judgement each time; it's the same decision rule applied consistently, which is what makes it safe to run without a person re-approving every instance.

What still needs a person

Brand voice isn't a style guide a model can be handed once and trusted on forever. It's a set of judgement calls about what the brand would and wouldn't say in a situation nobody wrote a rule for yet, and that's exactly where a model trained to sound plausible will sound plausible and wrong at the same time.

Campaign strategy and positioning sit in the same place: deciding what to say, to whom, and why this quarter rather than last quarter is a judgement call about the business, not a drafting task. AI can generate options quickly; it can't tell you which option fits what the company is actually trying to do right now.

The clearest line is anything that's the brand's first public word on a topic: a new product launch, a response to a competitor, anything that reads as a reaction to current events, or a reply during a reputational problem. Getting a templated email variant wrong costs a redraft. Getting a brand's first public statement on something wrong is much harder to walk back.

The risk of running only the AI layer

Industry commentary this year has started flagging a related risk: when most brands run similar AI tools on similar prompts, outputs start to converge. A distinctive point of view is specifically the thing a model trained to produce plausible, safe output tends not to generate on its own. That distinctiveness is also usually the first thing to erode when nobody on the team is actively steering for it.

One approval gate, not a review of every draft

The fix isn't a person rereading every AI draft before it goes anywhere; that just recreates the agency's hourly cost inside the company instead of outside it. It's a single checklist gate before anything publishes: does this match the brand's voice, is every claim in it something the company can stand behind, and is this the kind of post that needs a strategic call rather than a drafting one. That's the same gate the action, not the analysis rule we've written about for finance approvals, applied to what goes out under the brand's name instead of what moves money.

Move one deliverable at a time rather than cutting the retainer in one go. Shift the most templated work first, watch the approval gate for a few cycles, and only move the next deliverable once that one is running cleanly. It's the same restart-one-workflow lesson that applies to any team adopting AI for work a person used to do end to end: moving everything at once makes it hard to tell which specific change caused a problem when one shows up.

What this does and doesn't solve

This replaces the execution hours an agency used to bill for: the drafts, the variants, the routine reporting. It doesn't replace a person who owns the brand's voice, makes the strategic calls, and signs off before anything genuinely new goes out, and it doesn't replace an outside perspective if that's what the agency relationship was actually for. If what's left on your retainer is mostly execution hours, that's worth a conversation about what moves and what doesn't.