Marketing adopted AI faster than any other function, and it has less to show for it than almost any other function. Both statements are true, and the reason they can coexist is straightforward. Generation scales. Review does not.
Once a team can produce ten times more content, the constraint moves immediately to the people who have to approve it, the channels that have to carry it, and the audience that has to want it. AI for transforming businesses through marketing works when you fix the constraint rather than feeding the bottleneck. That means gating claims at the point of generation instead of catching them at approval, and being honest about what your attribution model is actually measuring.
At iSpark we work with marketing leaders on that distinction, which is usually less about tooling than about process and measurement. This article covers where AI marketing automation produces measurable results, the attribution problem nobody wants to name, a documented example from a US retailer with fresh numbers, the mistakes that create brand risk, and how to start.
Where AI Powered Marketing Produces Results
| Application | What AI contributes | The real constraint |
|---|---|---|
| Content production | Drafts, variants, localised versions at volume | Review and approval capacity |
| Personalisation | Message and offer selection per customer | Quality of first party data |
| Customer insight | Synthesis across reviews, tickets and calls | Acting on what it finds |
| Onsite assistance | Guided discovery instead of keyword search | Product data completeness |
| Campaign analysis | Surfacing what moved and what did not | Correlation is not incrementality |
| Ad creative testing | Generating and rotating variants quickly | Brand consistency across output |
The last row in that table is where most brand damage originates. A model producing hundreds of variants will eventually produce one that makes a claim your legal team would never have approved. The fix is not more approval stages, which simply recreates the bottleneck. It is constraining generation so regulated or sensitive claims cannot be produced in the first place. Our work on AI for marketing teams puts that gate at generation for exactly this reason.
The Attribution Problem Worth Naming
Marketing leaders are routinely asked to prove AI improved performance, and the tools available to answer make that harder than it sounds. Attribution models report correlation between touchpoints and outcomes. They do not report what would have happened otherwise. Incrementality is the honest word for the thing you actually want to know, and it requires holdout groups and controlled tests rather than a dashboard.
The practical version is simple. If you are going to claim an AI driven marketing strategy delivered a lift, design a holdout from the start. Retrofitting proof onto a full rollout is the most common reason these programmes get quietly defunded in year two.
A Real Example: Michaels and Ask Mike
The challenge. Michaels, the US arts and crafts retailer, had a discovery problem rather than a traffic problem. Customers arriving with a project in mind, rather than a product name, were poorly served by keyword search, which requires you to already know what the thing is called.
The solution and implementation. The retailer launched Ask Mike, its first customer facing AI shopping assistant, in May 2026 and announced it formally that July. Rather than replacing search, it sits alongside it, letting shoppers describe a project and get guided toward the products that fit.
The outcome. Michaels reported that shoppers using Ask Mike converted at more than double the rate of customers using traditional search, and that 27 percent of interactions led to a product click or an item added to a cart. The tool handled nearly 75,000 conversations in its first weeks.
The business impact. Higher conversion on a segment that was previously badly served, plus a stream of data about what customers are actually trying to make. Two honest caveats belong here. Shoppers who choose an assistant are already engaged, so part of that conversion gap is self selection rather than tool performance. And a few months is not a seasonal cycle. The direction is encouraging, the magnitude needs another year.
Common Mistakes in AI Marketing Analytics and Content
- Scaling generation without scaling review, which creates a backlog and a risk surface.
- Catching regulated claims at approval instead of preventing them at generation.
- Claiming lift from attribution data that cannot show incrementality.
- Personalising on data that is incomplete, so the personalisation is visibly wrong.
- Publishing volume that dilutes the brand voice because nobody defined it for the model.
- Measuring output produced rather than pipeline or revenue moved.
Best Practices Checklist
- Write down your brand voice and claim rules in a form a model can follow.
- Constrain generation so prohibited claims cannot be produced at all.
- Design a holdout group before launch if you intend to claim a lift.
- Audit first party data completeness before building personalisation on it.
- Keep a named human accountable for anything published externally.
- Report pipeline and revenue, not assets produced.
How to Get Started
- Pick one channel where volume is genuinely constrained by production capacity.
- Document voice, claim rules and prohibited language before any generation.
- Run with a holdout from day one so the result is defensible.
- Measure conversion and revenue, not clicks and output.
- Expand only after the review process has absorbed the new volume without slipping.
Future Trends in AI Marketing
Three shifts matter. Marketing operations work is becoming automatable in ways campaign work is not, and OpenAI’s published account of how Zapier’s enterprise marketing team automated its lead funnel quality checks is a good illustration of where that value sits. Discovery is moving away from keyword search toward conversational assistance, which changes how products need to be described. And AI answer surfaces are reshaping how brands get found, shifting the winning unit from a ranking position to a quotable, corroborated passage.
Key Takeaways
- Generation scales and review does not. Fix the constraint, do not feed the bottleneck.
- Gate regulated claims at generation. Approval stages cannot catch volume.
- Attribution shows correlation. Incrementality needs a holdout designed in advance.
- Michaels’ conversion gap is encouraging, and part of it is self selection.
Frequently Asked Questions
Does AI content hurt search rankings?
Not inherently. Quality, originality and usefulness matter more than production method. Thin, high volume content performs badly whether a human or a model wrote it.
How do we prove AI improved marketing performance?
Design a holdout group before launch. Attribution models show correlation between touchpoints and outcomes, not what would have happened without the change.
What is the biggest risk in AI marketing?
Publishing a claim nobody approved. Constrain generation so prohibited or regulated language cannot be produced, rather than relying on review to catch it.
Where should a marketing team start?
Where production capacity genuinely limits volume, with brand voice and claim rules documented first and a holdout designed in from day one.
Does personalisation need a large data set?
It needs complete data more than large data. Visibly wrong personalisation damages trust faster than generic messaging ever does.
Where to Take This Next
The marketing teams getting durable value here are not the ones producing the most content. They wrote down their voice and claim rules, moved the gate to the point of generation, designed measurement that could survive a finance review, and expanded only when approval kept pace. That is slower in the first quarter and considerably faster by the fourth.
If you want an independent view of where AI would improve your marketing performance and where it would only add volume, iSpark runs fixed scope assessments that end with a written recommendation either way. Start with the channel where production capacity is the real limit.
Published by iSpark.
