The most impactful AI use cases often sit in operational data that organisations already collect but rarely analyse — returns patterns, test data bottlenecks, and process friction hiding in plain sight.
We spend a lot of time in discovery workshops with engineering and operations leaders. One pattern keeps showing up — and it is not about the algorithm.
The data is already there
When we ask "what data do you have?", the answer is usually "a lot." Transaction logs, order histories, return records, deployment metrics, test results, support tickets. The problem is rarely data availability. It is data activation.
Most organisations start their AI journey by looking outward — customer-facing chatbots, recommendation engines, personalisation layers. These are valid use cases. But they are also crowded, complex to measure, and slow to deliver incremental ROI.
The faster wins tend to sit in operational processes that nobody has examined through an AI lens. The data is already being collected. The process cost is already known. The feedback loop is already tight.
Two examples from this week
Returns prediction in retail. A fashion e-commerce business processes tens of thousands of returns per month. They track every return reason, every product category, every customer's purchase-to-return ratio. That data sits in a warehouse, powering dashboards that people glance at monthly. A returns propensity model trained on that same data could flag high-risk orders before dispatch, trigger sizing guidance, and route likely returns to fulfilment centres optimised for reverse logistics. The data exists. The model is straightforward. The impact is measurable in weeks.
Synthetic test data in fintech. A payments company has 15 product squads, all waiting for sanitised test data that takes days to provision. Their production database has everything a generative model needs to learn the schema, distributions, and edge cases — without exposing a single real customer record. Synthetic data generation turns a week-long bottleneck into a self-service API call. Engineering velocity improves, compliance risk drops, and the test suite actually covers the scenarios that matter.
Neither of these is a moonshot. Both use well-understood ML techniques. The differentiator is not the algorithm — it is the willingness to look at an operational process and ask: "what would change if we could predict this?"
Why the operational angle tends to win first
There are three reasons operational AI use cases tend to outperform customer-facing ones in the first 12 months:
- The feedback loop is tighter. You can measure returns reduction or test data provisioning time in weeks. Customer lifetime value takes quarters to shift.
- The data is cleaner. Operational data is structured, timestamped, and generated by systems you control. Customer behavioural data is noisy, sparse, and consent-dependent.
- The stakeholder is motivated. The head of fulfilment who processes 10,000 returns a month wants a solution now. The head of marketing exploring AI for personalisation is still building a business case.
What we would ask you to look at
If you are a CTO or VP of Engineering evaluating where to invest in AI this year, try a simple exercise. Pick three operational processes that are:
- High-volume and repetitive
- Supported by 12+ months of historical data
- Currently managed by rules, thresholds, or manual review
Those are your candidates. Not because they are exciting, but because they are tractable — and tractable is what gets to production.
Start with one. Build a proof of value in four to six weeks. Measure the outcome against the process cost you already know. That is how AI projects earn their second phase — not with a strategy deck, but with a number the finance team can verify.
We have written up two detailed examples this week. Worth a read if either domain is relevant to your business.