Build Practical AI Systems
How to Choose a Marketing Task for an AI Workflow
Compare recurring marketing tasks using repeatability, approved inputs, human review and the consequences of error before choosing a first AI workflow.
Choose a small marketing task that recurs, starts from information you can access and are allowed to use, produces an output a person can check, and has manageable consequences if the first draft is wrong.
A strong first candidate helps prepare an internal draft or decision. It does not publish, send or make a consequential decision automatically. The aim is to find one task worth mapping and testing, not the most impressive use of AI.
Start with the work, not the tool
It is easy to begin with “What can this AI tool do?” and then look for somewhere to use it. A better starting question is: “Which part of my existing marketing work is clear enough to examine?”
List tasks you already perform, such as turning approved notes into an outline, preparing interview questions or grouping feedback for review.
This article helps you choose the task. It does not map or automate the workflow. Anthropic's engineering guidance recommends starting with the simplest solution and adding complexity only where it is needed. Although its article is written for technical teams, that principle is useful here: a more complicated workflow is not automatically a better starting point. Read Anthropic's guidance on effective AI systems.
Four checks for a first AI marketing workflow task
These four checks are recommendations, not a universal standard. Write a short observation for each candidate instead of assigning points.
1. Does the task repeat?
A reusable workflow is most useful when the same type of work returns and its purpose and desired output remain similar. There is no universal frequency threshold. A monthly task may benefit from consistency, while a daily task may still be unsuitable if every case needs different judgement.
Ask: What stays the same when this task comes back?
2. Are the inputs available and authorised?
Name what the task needs, such as an approved brief, product facts or research notes. Check whether the inputs are current, accessible and permitted for the tool and purpose. Scattered, conflicting or private material may need preparation before the task is suitable. Convenience is not permission to upload information.
Ask: Can the input owner supply one approved starting set without guesswork?
3. Can a person review the output?
The output should be specific enough to assess. An outline can be checked against its notes, while “improve our marketing” is too vague. Name the reviewer and what they will inspect before the output moves on. If nobody can explain a good result, the task is not ready.
Ask: Who will check the result, against what evidence or criteria, and before which next action?
4. What happens if the output is wrong?
Consider the effect of an inaccurate or inappropriate answer. A checked internal outline usually offers more room to catch mistakes than an automatically sent customer email. Consequences still depend on the information, audience and context. UK government guidance recommends proportionate assurance, with stronger measures for higher-risk uses. Read the Introduction to AI assurance.
Ask: Can a person catch and correct a problem before it affects a customer, decision, commitment or public claim?
A worked comparison of three marketing tasks
The table below is illustrative. It does not report a client method, measured result or test of a particular AI tool.
| Candidate task | Repeatability | Inputs | Human review | Consequence if wrong | Decision |
|---|---|---|---|---|---|
| Turn approved subject notes into an internal article outline | Similar output recurs | A small approved note set can be prepared | An editor can trace each point to the notes before drafting | Mistakes can be stopped while the outline remains internal | Ready to explore |
| Create a complete campaign plan from scattered notes | Campaign work recurs, but this combines many decisions | Material is incomplete and may conflict | No single check covers audience, offer, evidence, channels and budget | A weak plan could shape several later assets | Narrow first |
| Automatically send webinar follow-up emails | The action may recur | Recap, links and recipient data would all need control | A pre-send check disappears if sending is automatic | Errors reach people and may create privacy, consent or reputational issues | Not a first workflow |
The first candidate is strongest because it is bounded and reviewable, not because outlines are universally best. The second could become several smaller tasks. For the third, drafting and sending are separate decisions. Support for a draft does not imply permission to choose recipients or send it.
Make a decision without fake precision
A score can look objective even when its numbers are arbitrary. For a first shortlist, use three plain decisions:
- Ready to explore: all four checks have clear, supportable answers.
- Narrow first: the outcome, input or review boundary is too broad, but a smaller task is visible.
- Not a first workflow: necessary information or ownership is missing, or an error would pass into a consequential action too easily.
Ready to explore does not mean proven. It means the task is defined well enough to map, test and reject if it does not help.
Write the selected task in one sentence
Before building anything, record the boundary. A useful structure is:
From [authorised input], prepare [reviewable output] for [named reviewer] to approve, revise or stop before [next action].
For the strongest illustrative candidate:
From one approved set of subject notes, prepare an internal article outline for the content owner to approve, revise or stop before drafting begins.
That sentence does not define the full workflow. It makes the chosen task precise enough for the next design step.
Common task-selection mistakes
- Choosing the broadest task: “Create a campaign” hides several decisions. Select one deliverable or handoff.
- Assuming repetition is enough: A recurring task still needs usable inputs and a reviewable result.
- Ignoring input ownership: A workflow cannot repair missing permission, contradictory facts or an unclear source of truth.
- Leaving review until after use: Put the human decision before publishing, sending or passing the output into another consequential step.
- Promising a time saving before testing: A suitable task is only a candidate. Review and correction time count when the workflow is evaluated later.
What happens after you choose?
Keep human responsibility visible. What Is AI-Assisted Marketing? explains what AI may support, what a person must decide and what should remain private.
Then map only the selected task. How to Map a Simple AI Marketing Workflow uses four boxes: input, AI-supported action, human check and output. Testing, measurement and any decision to automate come later.
Further reading
- Anthropic: Building effective agents, for the simplest-sufficient-solution principle and its current tooling caveat.
- UK Government: Introduction to AI assurance, for context-dependent and proportionate assurance.
- NIST: Generative Artificial Intelligence Profile, for a broader view of use-case-specific risk.
Your next step
List three marketing tasks you perform more than once. For each, write one sentence about repeatability, inputs, human review and the consequence of an error.
Mark each task ready to explore, narrow first or not a first workflow. Choose one ready task, or narrow one mixed candidate until its input, output and reviewer are clear. Do not choose by adding unsupported scores.
You Might Still Be Wondering...
Frequently asked questions
Choose a small recurring task with approved inputs, a reviewable output and a person who can catch problems before the work affects a customer, commitment or public claim. An outline is one illustration, not a universal answer.
No. The task should recur enough for a reusable method to be worthwhile, but clarity and reviewability matter as much as frequency.
No. Selection only identifies a candidate. Map the manual, human-reviewed sequence and test it before deciding whether automation is useful or appropriate.
Not necessarily. High impact can mean mistakes carry greater consequences. A bounded task with a clear review point may produce more useful first-workflow learning.
Do not assume that a useful task authorises the information. Check whether each input is necessary, current and permitted for the intended tool and purpose. Follow the relevant organisational, privacy and specialist guidance before proceeding.