AI marketing strategy: how to choose what to automate first

Most marketing teams do not lack AI ideas. They have a list of twenty, a few subscriptions and no agreement on which one deserves the next month of effort. The choice matters more than the tool, because the first project sets the team’s expectations for everything that follows.

My short answer: start with a task that happens often, whose inputs you already have, whose output a person can check quickly, and where a mistake is cheap to catch. Measure how the task performs today before you change it. Pick one, pilot it on real cases for a fixed period, and decide in advance what result would make you stop.

What an AI marketing strategy actually has to decide

An AI marketing strategy is not a list of tools or a slide about the future of the function. For a working team it has to answer four practical questions:

1. Which business problem are we trying to move? Pipeline, speed to market, cost per qualified opportunity, content quality, reporting trust. Pick one.
2. Where is the bottleneck in that problem? The step where work waits, gets redone or gets skipped.
3. Who owns the change? One named person who can decide what “good” looks like and can stop the pilot.
4. How will we know it worked? A measure of today’s performance and a threshold agreed before the pilot starts.

If you cannot answer the first two, AI will not fix the gap. It will make the unclear parts of the process run faster.

Start with the bottleneck, not the tool

A common mistake is starting from a capability (“it can write ad copy”) and looking for somewhere to use it. That produces activity, not improvement. A team whose campaigns launch late because approvals stall does not gain much from producing more ad variations faster.

So begin by naming the constraint. Ask the team where last month’s work waited. Look at the calendar, the ticket history or the shared drive, not only at what people remember. Then ask a harder question: is this slow because the task is genuinely laborious, or because the process around it is broken?

Some problems look like AI problems and are really process problems. If briefs arrive incomplete because nobody agreed what a brief must contain, a checklist and a template will do more than a model. Fix the process first. Automating a poor process mostly makes the poor output arrive sooner.

List the recurring work before you buy anything

Once you know the problem, list the recurring tasks around it. Recurring matters: a task done once a quarter rarely repays the effort of designing, testing and maintaining a workflow. For each task, note:

• what triggers it and how often it happens
• who does it now and roughly how long it takes
• what inputs it needs and where they live
• what the output is and who uses it
• what goes wrong, and how you find out

AI can help with this stage. Give a model your rough notes, meeting transcripts or task descriptions and ask it to organise them into this structure and flag missing information. It is good at noticing that you described an output but never said who checks it. It should not decide the priorities. That judgement depends on your commercial situation, which it does not know.

Five questions that sort the list

I sort candidate tasks with five questions. I use a simple 1 to 3 rating for each, as a way to structure the discussion, not as a validated scoring model. The value is in the conversation the ratings force, not in the total.

One rule sits above the ratings. A task whose mistakes are expensive and hard to review fails the first screen, whatever its other ratings. An AI step that can spend budget, email customers or overwrite CRM records without a person checking first is not a first project. It may become a later one, once you have evidence, controls and a way to undo the damage.

Reviewability is the question teams underrate. A faster draft only helps if checking it does not take longer than writing it. If your reviewer needs to redo the research to trust the output, you have moved the work, not removed it.

A worked example: same tasks, different answers

This example is illustrative. The team, tasks and timings are invented to show the method, not taken from a client.

A three-person marketing team at a B2B software company shortlists three tasks:

• Weekly performance report. Every Monday, the marketing manager pulls numbers from the ad platforms, analytics and the CRM into a summary for leadership. About three hours a week.
• Campaign brief checking. Briefs arrive from product and sales. Most are missing the audience, the offer or the success measure, so each one needs a round of questions. Perhaps 45 minutes per brief, four or five briefs a month.
• Automatic budget shifts. An idea from a vendor demo: let an AI system move paid budget between campaigns daily based on performance.

The budget idea fails before we discuss value. It acts on money, and a person cannot easily review each change before it takes effect. It is not a bad idea forever. It is a bad first project.

The other two both pass, and the value rating is where the team’s situation decides:

• If the bottleneck is leadership trust in the numbers, the report wins. The draft saves a few hours, but the bigger gain is a consistent structure that separates what changed from what might explain it. The detailed method is in how to write a weekly marketing report with AI.
• If the bottleneck is campaigns launching late, brief checking wins, even though it happens less often. An AI step that compares each brief with a required checklist and returns specific questions to the requester attacks the actual delay.

Same team, same tools, different first project. That is why I do not believe in a universal list of “the best AI use cases for marketing”. The ranking depends on what is slowing you down.

Turn the choice into a pilot you can judge

A pilot is not “let’s try it for a while”. Write down, before you start:

• Owner: the person who judges quality and can stop the pilot.
• Scope: exactly which inputs go in and what comes out. Keep it narrow: one report, one brief type, one market.
• Baseline: how long the task takes today and what a good output looks like. If nobody has measured it, measure the next few cycles manually before changing anything.
• Test cases: real examples, including the awkward ones. For brief checking, include a brief that is complete, one that is missing half its fields and one that is complete but contradicts itself.
• Quality criteria: what the reviewer checks, written as questions they can answer yes or no.
• Review route: who checks each output before it is used, and how their corrections are recorded.
• Rollback: how the team returns to the manual process if the pilot stops.
• Stop condition and time box: for example, stop if review plus correction takes longer than the manual task after the first two weeks, or if an unsupported claim reaches leadership.

On test cases, the evaluation advice from agent builders transfers well to smaller projects. Anthropic’s engineering team suggests that for AI agents, 20 to 50 simple tasks drawn from real failures is a great start. A marketing pilot may need fewer cases, but the principle holds: test on the work you actually do, including the cases that went wrong before.

Continue, revise or stop

At the end of the time box, the owner makes one of three decisions:

• Continue when the output meets the quality criteria and the total time, including review and correction, is lower than the baseline. Extend the scope one step at a time.
• Revise when the output is useful but review takes too long, or one type of case keeps failing. Change the inputs, the instructions or the scope, and run another short cycle.
• Stop when the task turned out to be a process problem, the inputs are not good enough, or the review cost exceeds the saving. Stopping is a result, not a failure. Write down why, so the same idea does not come back in six months without new evidence.

Time saved is only part of the value, and not always the most important part. Measuring the full cost and benefit of a pilot is a subject of its own: I cover it in AI marketing ROI: measure value beyond time saved. If the pilot works and you are tempted to give the system more autonomy, read AI agents vs marketing automation first. Anthropic’s advice for builders is a good default for marketers too: find the simplest solution possible, and only increase complexity when needed.

The prioritisation worksheet

Copy this table into a spreadsheet or document and fill one row per candidate task. Rate the four scored columns 1 to 3 using the anchors above, then discuss the result rather than adding it up.

What this method does not tell you

This worksheet helps a team choose a sensible first project. It does not predict the return on that project, and the ratings are judgements, not measurements. Two people can rate the same task differently; that disagreement is worth discussing before anything is built.

It also assumes you know your customer well enough to know which problems matter. If you are not sure who your best-fit customers are, that question comes before any automation. Ideal customer profile: build it from evidence, not AI guesses is a good place to start.