AI tools for marketing agencies: how to choose them by job

AI is already in almost every agency. In HubSpot's 2026 State of Marketing report, which surveyed more than 1,500 marketers, 86.4% of marketing teams said they use AI in at least a few areas. HubSpot's Spanish-language blog also cites its 2024 report, in which 64% of marketers already used generative AI. The question is no longer whether to use it, but for what.

The usual mistake is buying by the tool: someone tries a new assistant, likes it, and the whole team pays for it. Six months later there are five subscriptions doing the same thing and none that solves the job that costs the most hours. Work the other way round: list the jobs that repeat every week and find the kind of tool that solves each one.

A quick way to build that list: ask each person to write down, for one week, every task they repeated more than twice and roughly how long it took. Sort the list by hours. The top five are where an AI tool can earn its price; anything below is a nice-to-have, and nice-to-haves are how agencies end up with subscriptions nobody can explain at renewal.

These are the jobs where AI saves real hours at an agency, and what to demand from a tool in each:

• Desk research: an assistant with search that cites its sources. Check every citation before it reaches a deck; a source that doesn't exist is worse than none.
• Briefs: a tool that works in fields (objective, audience, constraint, deliverables) and flags what's vague, not one that writes a polished brief that still decides nothing.
• Meetings: transcription that tells speakers apart and pulls out action items. Ask the client's permission before recording.
• Reports: AI to summarize and point out what changed, never to supply numbers. The number comes from the platform; the AI explains it.
• Images and moodboards: generative models to explore directions. Check each model's commercial-use terms before an image reaches final work.
• AI visibility: knowing what ChatGPT, Gemini, Claude or Perplexity answer when someone asks about your client's category. It's the newest job and the one fewest agencies measure.

Before adding a subscription, ask these five questions. If a tool can't answer them, it isn't ready for client data.

1. Does it say where each claim comes from, and can you open the source?
2. What happens to client data? Read the vendor's policy for its business plan. Anthropic, for example, says that by default it doesn't use inputs or outputs from its commercial products to train models. And not used for training doesn't mean not stored.
3. Who approves before anything reaches the client? A tool that sends on its own, with no review step, shouldn't touch an account.
4. What does it cost at real usage? Many tools charge by credits or have daily limits; work it out at a normal month's volume, not the demo's.
5. Does it work well in the languages your clients use? Test it with a real brief, with the client's own terms and local brands.

Test a new tool the way you'd test a new hire: on real work, for a short time, with someone checking. A month is enough. In week one, note how many hours the job you want to solve takes, so you have something to compare against. In week two, use the tool with one client and one person responsible. In week three, measure the hours again and count the mistakes that had to be fixed, including citations that didn't exist.

Week four is the decision. If it saves hours and its mistakes can be caught, it stays, and three rules go on one page: which client data may be pasted in and which may not, who reviews before anything goes out and who owns the subscription. If it doesn't save hours, cancel it before the renewal, even if someone likes it.

More people ask an AI assistant before they search Google, and those visits can now be measured: Google Analytics 4's default channel group includes an AI Assistant channel for traffic from sources such as ChatGPT. What GA4 can't tell you is whether the assistant names your client when someone asks about the category, or names a competitor instead.

For a first answer, the free brand scan runs one real check on any domain: whether an assistant knows the brand, and whether it names it unprompted next to similar companies. To measure it properly there's , which asks ChatGPT, Claude, Gemini and Perplexity a buyer's questions on a schedule, scores how often the brand is named and where it's placed, and shows which sources the assistants cite. ChatGPT is on every plan; Claude, Gemini and Perplexity come with Agency and Enterprise or as an add-on. Two honest limits: it doesn't track Google AI Overviews yet, and it doesn't read Reddit or X.