AI agents in customer analytics: what a year of builds taught us

We've spent the past year building customer analytics with AI agents. The work covered segmentation, single customer views, value prediction and funnel diagnostics across several sectors. What decided the value each time was speed: how fast we got from a question to a working answer, and how quickly the team could use it.

A unified customer view and predictive analytics are the two data problems marketers name most often, both cited by 34% of teams in Supermetrics' 2026 Marketing Data Report. They take time. Half of marketers wait one to three business days just for data team support.

Here's what that looked like in practice.

AI-powered segmentation, built in hours

A global beauty group runs 21 brands across EMEA and up to 500 campaigns a quarter. Its RFM models were static and manual, so building a new segment took weeks while the campaign calendar kept moving. RFM ranks customers by recency, frequency and monetary value: how recently someone bought, how often and how much they spend. It's most useful when you can re-cut it as the question develops. Some of their reactivation campaigns were returning under 1% response.

We built two agents inside their CRM environment, supervised by CRM specialists. The first created dynamic lifecycle and product segments and standardised data fields across markets. The second analysed purchase cycles, engagement and churn risk to find the highest-value reactivation opportunities, sized the revenue in each and fed the priority segments into their journeys.

Segment creation dropped from weeks to hours. Reactivation rose 24% among dormant high-value cohorts. The work identified more than £7m of incremental revenue potential across priority audiences over twelve months. The CRM team spent the recovered hours on testing and creative.

The speed changed how the team worked with the model. When they wanted to split a segment further or add a variable, the change was ready in the same session, so they could ask the next question straight away.

One customer view of 3.47 million people, in 7 days

A UK holiday parks operator had booking history, on-park spend, email engagement and loyalty records sitting in separate systems. Its CRM recognised a customer at their last booking, so no one could see a whole lifetime of value. We unified 8.9 million rows from four data extracts into one row per customer, covering 3.47 million people, and matched 81.7 million email events to a named customer for nine in ten of them.

From raw extracts to the first findings took seven days. A full strategic framework followed by day eleven. We mapped every active customer by the value they hold today and their room to grow, then let behavioural groups form from the data through clustering, so the groups came from what customers actually do rather than a rule written in advance.

The map put £204.1m of annual booking value in one place, sorted by where it sits. Half the active base holds 82.5% of that value, so the marketing team can rank spend by the value each group holds. We reproduced both of the client's contracted growth KPIs on their own definitions, so the numbers matched the board's.

Predicting customer value from early signals

A national retailer wanted to know which new customers would become valuable while there was still time to act on it. We scored who was likely to become high value from their earliest signals: first-purchase behaviour, the mix of products bought, browsing depth and whether they looked like a trade or DIY buyer. Prototype in 10 days. Live scoring in three weeks.

It moved at that pace because the question was specific and the agent could build, test and extend within a week. The work that decides whether a model earns anything is defining what actually predicts value in that business, with that customer base, at that price point. That's where the time should go.

Verifying the figures that carry the decision

The same holiday parks operator came back for a second engagement, on their holiday home sales funnel. Sales and marketing were quoting different show rates from the same CRM, and 87% of leads were going uncontacted. We rebuilt the funnel end to end on one reconciled base of 181,604 lead records across 24 months, then sized each point of loss in pounds.

Faster building puts more weight on the checking. Every input was pinned to a versioned record and the load-bearing figures ran through an independent acceptance gate before anything reached the client, tested against eight known ways a number can mislead.

That checking earned its place. An apparent paid-social advantage turned out to be park-created leads mislabelled in the CRM, caught before it shaped media spend. The work sized about £20.8m of recoverable new-customer sales across six levers, an indicative central figure, and ranked 12,366 unworked leads warmest first. Contact was the lever that moved most. Contacted leads converted at 7.26%. Uncontacted leads at 0.55%.

Keeping first-party data where it already lives

One of the UK's largest casual-dining loyalty programmes was rebuilding its proposition to a fixed deadline. It had 4.3 million members, all the data sitting in Snowflake and no analyst free to interrogate it in time. We worked inside their own warehouse. Role-based access, no extract, nothing crossing their security perimeter.

We built one governed record per member across the 4.3 million base, segmented it, ranked the commercial levers by size and left an interactive dashboard and a plain-English query tool running in their account. From first access to readout took around two and a half weeks. The load-bearing figures were reconciled to their own reporting before they left the room.

Skipping the extract-and-anonymise cycle recovered about two weeks and made the data protection conversation simpler. Their team keeps the analytical layer and can read it, run it and extend it. The data never leaves their account.

What decided the value

Across every build the algorithm was the least interesting part. Three things decided the outcome.

The question. Speed pays when you've pointed it at something worth predicting. Deciding which question is worth asking is a strategic job and it stays with you.

The data the client owns. Agents are the delivery mechanism. Data is the moat. Your first-party behavioural data is the asset a competitor can't reach, so this work belongs where that data already lives.

The speed of the second question. A build that lands in days means the first answer arrives while you can still act on it and you have time to ask the next thing it raises. The model becomes something the team uses and keeps, rather than a report that arrives after the decision.

Only 41% of marketers can demonstrate a return on their AI investment in 2026, down from 49% a year earlier, according to Benchmarkit's State of AI in Marketing 2026. Teams see that return when a good question, data they own and a fast action come together. Agents shorten the time from question to action. Which questions to ask is still yours to decide.

Want to see how Data Agents can help you do more with less? Book a quick strategy call.