AI Won't Fix a Broken Marketing Operation

I run AI empowerment programmes for marketing teams, so this may read oddly: the majority of AI disappointment I encounter has nothing to do with the tools.

The tools mostly work. The pattern is that AI compounds whatever operating model you already have. Sound model, real and large gains. Broken model, and you have bought a faster way to produce the wrong work.

That is an uncomfortable finding to deliver, because the AI budget has usually already been approved and the expectations already set.

The three questions before any tooling

Before I let a team pick tools, I ask three things. The answers predict the outcome better than any evaluation matrix.

1. Where does work currently queue? Every marketing operation has a bottleneck. Sometimes it is content production. Often it is approvals, or legal review, or a single overloaded designer, or waiting on data.

AI applied away from the bottleneck produces no throughput gain whatsoever. This is the single most common failure I see: a team whose real constraint is a three-week legal review cycle buys content generation tooling, triples draft output, and the queue in front of legal simply gets longer. Everyone is busier and nothing ships faster.

Find the constraint first. Sometimes the answer is that AI is not the intervention — the intervention is a delegated approval threshold, and it costs nothing.

2. What does your team do that is genuinely repetitive? Not "creative work we could accelerate". Repetitive work. Reformatting the same report weekly. Adapting one campaign across six markets. Writing forty product description variants. Summarising the same feedback channels.

This is where returns are immediate and undeniable, and it is where I always start — partly for the results, mostly because it builds trust with a sceptical team far faster than asking them to hand over the work they are proud of.

3. Who is accountable for output quality after automation? The question nobody asks in the business case, and the one that determines whether the programme survives its first bad quarter.

When a human writes something and it is wrong, ownership is obvious. When AI drafts something, a human skims it, and it goes out wrong, ownership evaporates — and the natural institutional response is to add a review layer, which reinstates the bottleneck you were trying to remove.

Name the accountable human per output type, up front, before anything is automated.

Where the returns are actually large

Being concrete about what has worked, repeatedly:

Localisation and market adaptation. The clearest win I have seen. Adapting a campaign across English, French and German markets used to mean agencies, weeks, and cost that made per-market tailoring uneconomic. Now it is a first draft in minutes plus native review. The quality is not the point — the economics are. Work that was previously not worth doing at all becomes routine.

Variant volume for testing. Testing programmes are usually constrained by how many variants a team can produce, not by traffic. Remove the production constraint and you can run genuinely meaningful tests. This compounds: more tests, faster learning, better baseline.

Qualitative data at scale. Support tickets, reviews, survey free-text, sales call notes. Most organisations have enormous qualitative reserves that go unread because reading them does not scale. It does now, and the customer language it surfaces feeds directly into positioning.

Analysis first drafts. Not conclusions — first passes. "Here are the notable movements in this dataset and plausible explanations" is a genuinely useful starting point for an analyst, and saves the boring half of the work.

Where it reliably disappoints

Strategy. It will produce a fluent, plausible, entirely generic strategy, and the fluency is the danger. It reads like competence. Strategy is fundamentally about what you decline to do, and that requires context, conviction and political capital that no model has.

Brand voice at volume without a strong existing voice. If your brand voice is well-documented and distinctive, AI reproduces it well. If it was always a bit vague, AI will average it toward the internet's mean, and the drift is slow enough that nobody notices for two quarters.

Anything requiring accountability for being wrong. Regulated claims, pricing, competitive comparisons, anything a customer might act on financially. The review cost exceeds the drafting saving.

Fixing bad data. Perhaps the most expensive misconception. If your customer data is fragmented and your events unreliable, AI will produce confident output built on it. Previously bad data produced obviously bad reports and someone investigated. Now it produces plausible narratives, which is considerably worse.

The sequence that works

The programmes that stick follow roughly the same order.

Start with the repetitive, low-risk, high-volume work — localisation, reformatting, variant generation. Build capability and trust where being wrong is cheap and correction is fast.

Then fix the data foundation, if it needs it. Unglamorous, always underestimated, and the ceiling on everything above it.

Then move to work that touches judgement, with named owners and explicit review.

And throughout, measure throughput against the bottleneck you identified, not activity. Drafts produced is not a metric. Campaigns shipped, time to market, tests completed — those are.

The part clients like least

Frequently, the honest recommendation after two weeks is: do not start with AI.

Fix the approval process. Consolidate the four overlapping tools. Assign an owner to the orphaned metric. Clean up the event tracking. Then come back, because the same investment will return several times more against a functioning operation.

That is a difficult message when an AI initiative has been announced internally. But the alternative is spending the budget to accelerate a process that was the actual problem — and arriving in six months with faster output, the same results, and a team that has concluded AI does not work here.

AI is a multiplier, not an addition. Multipliers are excellent news if the number you are multiplying is sound. Worth checking the number first.