Where AI actually helps marketing operations

I have run the same scoring exercise across 23 teams inside one organisation, as an internal program — not a pilot. The output was a ranked list rather than a set of opinions, and the most useful thing about it was how much it removed from the roadmap.

Here is the scoring model, because it is more useful to you than the conclusions from someone else’s organisation.

The four axes

Score every recurring workflow — not every tool, every workflow — on four things.

1. Time consumed today. Hours per month, across everyone who touches it. Measure it rather than estimating it; the estimates are wrong in both directions and the error is not random.
2. Tolerance for a wrong output. If the model is wrong one time in twenty, what happens? A draft that a human edits: nothing. A price shown to a customer: a very bad afternoon. This axis eliminates more candidates than the other three combined.
3. Availability of the inputs. Does the model have access to what it needs, in a system, today? If the context that makes the decision correct lives in someone’s head or in a Slack thread from March, the workflow is not automatable yet — it is undocumented, which is a different and cheaper problem to fix.
4. Cost to change. What has to be rebuilt, re-approved or re-integrated? A workflow inside one tool is cheap. A workflow that spans your CDP, your messaging platform and a legal review is not.

As a rubric, on one page:

Axis	The question	Scores high when	Scores low when
Time consumed	Hours a month, across everyone who touches it	It is a standing job for several people every week	It happens twice a year
Error tolerance	What happens when it is wrong one time in twenty?	A human edits the draft before anyone sees it	A customer acts on it, or a regulator reads it
Input availability	Is what the model needs in a system today?	The inputs are fields in a table	The context lives in someone’s head
Cost to change	What has to be rebuilt, re-approved or re-integrated?	It sits inside one tool	It spans the CDP, the messaging platform and legal

Score each axis on the same scale — one to five is plenty, provided one person calibrates all of them in one sitting — and rank on the total. Treat error tolerance as a veto rather than a weight: a workflow that scores one there does not belong on the list at any total.

The interesting quadrant is high time, high error-tolerance, available inputs, cheap to change. In most organisations it holds a handful of workflows, and rarely the ones with the most enthusiasm behind them.

Illustrative. The rows below are invented — a real scored list is a client artefact — but the shape of the output is exactly this: four scores, a total, and a call you can argue with.

Workflow	Time	Error tol.	Inputs	Cost to change	Call
Localising campaign copy into six markets	5	4	4	4	Do first
Weekly performance commentary	4	4	5	3	Do first
Tagging and summarising support tickets	5	4	4	2	Do next
Segment eligibility rules	3	1	3	2	Do not
Annual brand tracker write-up	2	3	2	2	Leave it

Note the fourth row. It has a respectable total and it is still a no, because the error-tolerance veto applies: an eligibility rule that is wrong one time in twenty is a customer being told something untrue about their own account.

Where the leverage genuinely is

Consistently, across teams and companies, four categories score well.

• Draft generation. First versions of anything a human will edit: copy, briefs, summaries, test plans. High time, high error-tolerance, inputs usually available. This is unglamorous and it is where most of the real hours are.
• Variant production. Twelve versions of a subject line, localised into six markets, is a task where the marginal cost per variant was the actual constraint on doing it properly. Removing that constraint changes the strategy, not just the workload.
• Classification and summarisation of unstructured text. Support tickets, reviews, survey free-text, sales call notes. Enormous time cost, high tolerance for individual errors, and the aggregate is what you act on anyway.
• First-pass QA. Checking links, tokens, localisation completeness and rendering before a human review rather than instead of one. It never catches everything, and it does not have to.

Notice what these have in common: a human remains in the loop, and the model is compressing effort rather than making a decision.

Where it is theatre

Three categories come back consistently as bad candidates, and naming them is worth as much as the roadmap.

Anything requiring a guarantee. Pricing, eligibility, regulatory claims, anything with a number a customer will act on. Error tolerance is effectively zero, and the cost of the review process you would need to bolt on exceeds whatever you saved.

Anything whose inputs are undocumented. “Why do we always exclude this segment in Norway?” If the answer is institutional memory, the model cannot have it, and no amount of prompt engineering substitutes for writing it down. Write it down first; you may find that is the whole improvement.

Anything measured by adoption. Seat usage and prompts per employee are counts of activity, and an initiative reporting them has usually skipped the question of what it was meant to change — the same failure as a churn program with no number attached to it. This is the clearest tell there is, and it survives contact with any amount of executive enthusiasm.

Build or buy

The build-versus-buy call is where most of the money is won or lost, and the honest version is uncomfortable for both sides of the market.

Buy when the workflow is generic, the vendor’s data advantage is real, and the integration is the product. Build when the workflow encodes something specific about your business — your segmentation logic, your tone, your definition of lifetime value — because that specificity is exactly what a general tool will flatten.

And a third option that gets forgotten: write twenty lines of code. A surprising share of what platforms are sold for is a scheduled script with a model call in the middle. I have built and shipped an AI tool in this category myself, which is mostly useful as a calibration device: it tells you how much of the price is capability and how much is packaging.

How to run this yourself in a week

You do not need help for the first pass.

1. List every recurring workflow in the marketing organisation. Not tools — workflows. Aim for forty to eighty for a mid-sized team.
2. Interview rather than survey. Forty-five minutes per team, and ask what actually happens, not what the process document says. Survey answers describe the aspirational version.
3. Score all four axes, on the same scale, in one sitting so the calibration holds.
4. Take the top item and actually build it, badly, in a day. The prototype is what tells you whether the ranking is right.
5. Publish the “not doing this” list as prominently as the roadmap. It is the half that saves money.

If that takes you three weeks rather than one, that is the correct answer and not a failure. If it takes a quarter, the organisation is answering a different question — usually whether it can agree on anything — and that is worth knowing too.