PixelPrint — AI CRM Integration for Hyper-Personalisation

PixelPrint is one of Europe's leading personalised photo product platforms. Founded in 2000 and operating across 12+ countries, it serves over 5 million active customers who turn their photos into prints, books, gifts, and wall art. With 154,000+ Trustpilot reviews and a 4-star rating, the product quality and core creation experience were strong.

The opportunity wasn't in the product — it was in how PixelPrint communicated with its customers. Every touchpoint — email, homepage, product recommendations — was built on the same static template regardless of who the customer was, what they'd bought before, or what life moment had driven them there.

A first-time buyer uploading graduation photos got the same homepage and email flow as a grandmother ordering her fifth annual Christmas photo book. The platform knew nothing it didn't use.

PixelPrint had rich data sitting across three disconnected systems: an order history database going back 8 years, a behavioural analytics platform tracking upload events and editing sessions, and a legacy CRM holding email preferences. None of these talked to each other in real time.

The consequence: a customer who had just uploaded 80 baby photos received the same "Don't miss our summer sale" campaign as someone who ordered canvas prints every year. Intent was visible in the data. The platform was blind to it.

Three signals surfaced this during discovery:

• Seasonal dependency: 61% of annual revenue landed in the 8-week Christmas window. Outside of that period, retention collapsed — because there was no mechanism to bring customers back to the product at the right moment.
• Recommendation blindness: The "Customers also bought" logic on the product page used global bestsellers, not user history. A customer who had purchased 3 photo books was being shown mugs as a next step.
• Email irrelevance: The email open rate sat at 18%, with a click-through rate of 2.4% — driven almost entirely by the 3 mass campaigns per year. Between campaigns, email sent zero personalised communications.

Before defining the solution, I ran a 4-week discovery sprint to understand what the data could actually tell us, and where the gaps were. The goal wasn't to prove a hypothesis — it was to surface what the platform already knew about its customers that it wasn't using.

We ran three parallel tracks:

• Behavioural data audit: Mapped 24 distinct event types in the analytics platform — from photo upload sessions to creator tool interactions. Found that upload event clusters reliably predicted product category intent within 3 days with 74% accuracy.
• Customer interviews (n=28): Recruited across 4 segments — first-time buyers, lapsed customers (no purchase in 12 months), high-LTV repeat buyers, and seasonal-only (Christmas). The finding that changed the framing: lapsed customers didn't leave because of price or quality — they left because "nothing reminded me I had photos waiting."
• CRM capability audit: The existing CRM could send triggered emails based on last-purchase date, but had no access to real-time behavioural data or upload events. The gap wasn't in the email tool — it was in the data feed.

The core idea: replace calendar-driven mass campaigns with an AI layer that listens to behavioural signals and triggers personalised communications, recommendations, and homepage experiences in response to what customers are actually doing — not what month it is.

The north star metric was repeat purchase rate. Everything else — email open rate, AOV, NPS — would follow if we got personalisation right. We defined three pillars:

• Moment detection: Identify life events and purchase intent signals from upload behaviour, album metadata, and session patterns in real time.
• Contextual surfacing: Route detected intent through the CRM to trigger personalised email, homepage modules, and product recommendations — aligned to the detected moment.
• Churn prediction: Score every customer weekly against a churn risk model and intervene proactively at 30, 60, and 90-day inactivity thresholds with contextual win-back content.

The architecture had four layers: data sources, AI/ML models, CRM integration, and customer touchpoints. The design principle was to make each layer independently testable — so model improvements didn't require redeployment of the CRM integration, and new touchpoints could be added without touching the model layer.

The build was structured into six sequential phases, each with a defined output, a go/no-go checkpoint, and clear rollback criteria. The constraint throughout was real customer data — every model was trained on PixelPrint history, not synthetic proxies.

The most visible change wasn't in the model — it was in the customer experience. Two surfaces shifted most dramatically: the homepage and the post-upload email.

This was a cross-functional product initiative that required close ownership across three domains that don't naturally talk to each other: ML engineering, CRM operations, and product design. My role spanned all three.

• Discovery and framing: Ran the 4-week data audit and 28 customer interviews. Reframed the brief from "improve our email performance" to "build a system that responds to customer moments." The framing shift determined the entire architecture — a campaign optimisation framing would have led to a much narrower, less impactful solution.
• Model specification: Wrote the ML model requirement documents for all 5 models — defining inputs, outputs, confidence thresholds, and evaluation criteria. Maintained the decision log for every go/no-go at model checkpoints. Pushed back on the ML team's initial recommendation to skip offline evaluation for the churn model — the subsequent offline test found a 14% false-positive rate that would have triggered unnecessary win-back emails to active buyers.
• Privacy and trust guardrails: Worked with legal to define what the AI was permitted to infer and act on. Explicit rule: if the only signal for a life event was album name, confidence was capped at 0.6 and no message was triggered without a supporting behavioural signal. This prevented, for example, a misclassified album triggering a baby product email to a customer who had titled their album "Baby shower gift for friend."
• Launch governance: Owned the phased rollout decision criteria. Defined rollback thresholds, monitored the real-time dashboards at each phase expansion, and made the advancement decisions at weeks 26, 28, 30, and 31.

The insight that changed everything: The lapsed customer interviews. Before those, the team believed the problem was marketing channel performance — open rates, send times, subject line testing. After hearing "nothing reminded me I had photos waiting" 19 times, it was clear the problem was structural: the platform had no mechanism for responding to customer behaviour in real time. That reframing was worth 4 weeks of discovery.

The decision I'm most glad I made: Requiring a 70% offline precision threshold before any model entered shadow mode. Two models failed this gate on first evaluation and required retraining. Shipping them would have generated a significant volume of incorrectly personalised communications — damaging trust in a context (personal photos) where trust is the entire product.

The mistake I made: I underestimated the change management required within the CRM team. They had operated on a campaign calendar for years — a system where they controlled every email sent. Shifting to an AI-triggered model meant relinquishing that control. I should have co-designed the trigger rule engine with the CRM lead from week 1, not introduced it at week 19. The late inclusion added 3 weeks to the integration phase and required rebuilding parts of the suppression logic to match their operational mental model.

What I'd carry forward: The shadow mode approach to AI launch is the most underused tool in product management. Generating predictions and logging them against actual outcomes — without acting on them — is the clearest possible test of a model's real-world performance. Every AI-adjacent product initiative I run now includes a mandatory shadow period before any live communications fire.