Written by: Luke Sobieraj, Founder & COO, Digital Colliers
The data model you pick in year one quietly decides what products you can ship in year three. Most custom platform teams don't feel this until they try to add something like predictive maintenance or a recommendation engine and realise the shape of their tables won't allow it. By then the retrofit cost is enormous and half the team is arguing about migrations instead of features.
The pattern I keep seeing in industrial and field-service platforms is the same choice, made quickly, that either pays off or punishes you later. You either model around the transaction (the order, the ticket, the invoice) or around the asset (the machine, the vehicle, the site, the device). Both look fine on day one. They look very different on day 700.
Order-shaped data is a dead end for prediction
If your core entity is the order, your history looks like a stream of discrete events attached to customers. That's great for billing, dashboards, and quarterly reporting. It's close to useless if you later want to ask, "is this specific machine about to fail?"
Because to answer that, you need a continuous, ordered timeline per asset. Sensor readings, service visits, part replacements, firmware versions, operator notes, warranty status, all keyed to the same physical thing. If your schema treats each service visit as a line item on an invoice, you can reconstruct that timeline, but it's slow, lossy, and painful. Every ML feature engineer who touches it quits a little inside.
This is one reason so many AI initiatives stall before they ship anything useful. IDC found that 88% of AI proofs of concept never reach widescale deployment, and S&P Global reported that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The models aren't usually the problem. The data shape is.
Asset-shaped data compounds
When the machine is the primary entity, every event you record gets attached to it for its entire lifetime. Ownership can change. Contracts can churn. Customers can leave and come back. The asset's history stays intact and keeps getting richer.
Two years of that, and you have something rare:
- A per-asset failure history with real labels
- Maintenance intervals you can compare across fleets
- Usage patterns tied to specific components
- A clean input for anomaly detection or remaining-useful-life models
That's the point where predictive maintenance becomes possible without heroics. And once you have predictive maintenance working, the next AI product (a service recommendation engine, an automated dispatch prioritiser, a parts-demand forecaster) is mostly a variation on the same data spine. You don't rebuild. You reuse.
The left-behind risk is a schema problem, not an AI problem
Here's the uncomfortable part. MIT's 2025 study found that 95% of enterprise GenAI pilots delivered zero measurable P&L impact. Most write-ups blame the models, the vendors, or change management. In custom platform work, the more honest answer is usually that the underlying system was never shaped to answer the questions the AI needed to ask.
If you're the operator running a platform that was built order-first five years ago, you're not alone, and you're not doomed. But you should be honest about which of these applies:
- You can't easily produce a full lifetime history for any single asset
- Sensor or telemetry data lives in a different system than service records
- Every new analytics request requires a bespoke join across three services
- Your ML team spends more time on ETL than on modelling
That's the left-behind shape. It doesn't kill you this quarter. It kills the AI roadmap two years out, quietly, by making every idea 6x more expensive than it looked in the pitch deck.
What the teams shipping in 2026 tend to do
The operators I see actually landing AI features in custom platforms have usually done one of two things. Either they picked asset-shaped modelling early and it aged well. Or they took a deliberate 6 to 12 month bet on restructuring their core entities before touching any ML work.
The second path is unglamorous. It doesn't demo well to the board. But it's the reason their predictive feature ships in a sprint instead of a year, and it's the reason their next AI product after that ships in weeks. The data model isn't infrastructure. It's the product roadmap in disguise.