Useful AI needs a record that was true at the moment it happened.
Almost every industrial AI disappointment traces back to the same thing: the model was trained on data that had already been averaged, reconciled and stripped of who, when and against what. You cannot recover that afterwards. It has to be captured at the point of work, which is the layer we operate.
The context is aggregated away before anyone sees it.
| What an ERP extract gives you | What the transaction record gives you | |
|---|---|---|
| Granularity | Monthly movement by stock code and cost centre | One event, to the second |
| Attribution | A cost centre, sometimes a job | A person, a crew, a shift, a work order, a location |
| Condition | Not held | In date, out of date, held, released under override |
| Exceptions | Inferred from variance at reconciliation | Recorded as they happen, with the name of whoever resolved them |
| Latency | Period close | Immediate |
Every transaction carries a story, and it is rarely just the item. An issue at 6:04 against a work order says something about how that job ran. A return says something about how it finished. A hold, and the name on the override that released it, says something about the state of the equipment and the call somebody made at the time. One at a time they are records. Together they are context — what gets drawn, by whom, against what work, and in what condition — and that context is what feeds a better decision about what to hold, what to reorder, where the next point belongs and which items need a regime. None of it has to be gathered specially. It is written as the store is used.
What is real now, what is coming, and what we will not claim.
| Attribution | Consumption resolved to person, crew, shift, job and cost centre, without a reconciliation step. |
| Exception detection | Overrides, out-of-date issues, unattributed movements and returns that never came back, on the same report as the successes. |
| Replenishment from consumption | Reorder driven by what actually left the shelf rather than by a static reorder point. The signal goes to your distributor, not through us — we track and distribute items, we do not supply them. |
| Demand shaping | Consumption forecast from the maintenance and shutdown schedule rather than from last quarter’s average. |
| Anomaly models | Consumption that is out of pattern for that crew, that job type or that time of shift — surfaced for a human to judge. |
| Estimating feedback | What a job class actually consumes, fed back into how the next one is quoted and planned. |
None of the second group is switched on by default and none of it works on a thin dataset. An anomaly model needs a baseline it does not have in month one, and a forecast built on a single site’s first year will be confidently wrong in a way that is worse than no forecast at all. Anything predictive is proposed, evidenced against your own data, and turned on deliberately — or it is not turned on. We would rather sell you the record now and the inference when it is earned.
The controls that protect the record protect everything derived from it.
| Position | What it means |
|---|---|
| One control set, no lighter regime for analytics | Reporting, models, extracts and exports run inside the same access control, logging, retention and residency as the transaction record itself. Deriving something from the record does not move it into a softer category. |
| Independently examined, continuously monitored | Our control environment holds a current SOC 2 Type 2 attestation — controls independently tested as operating across an observation period, not designed on a single day. Cybersecurity risk is monitored continuously between examinations rather than assembled for the auditor. The control set → |
| Your data stays yours | We do not train our models on your data. We do not sell your data. It is yours to export in a machine-readable format at any time. |
| A human stays on any decision that stops work | A gate that holds an issue is a control, not a verdict. Every override is available to a named, authorised person and every override is logged. Nothing here removes a supervisor from a safety decision. |
| Automated decisions are documented and disclosable | Where the system decides something about a person — a credential check, an in-date gate — that decision is documented, auditable and disclosable, and the documentation is maintained as expectations change rather than written once. |
| Attribution stops at three things | A transaction attributes the person, the item and the point it was drawn from. No pace, no duration, no output, no rating — none of it is collected, so none of it can be inferred from. The boundary, in full → |
| The detail goes to your reviewer, not onto this page | Governance and architecture documentation — control listings, data flows, the automated-decision inventory, the regulatory position and the assurance calendar — is provided under NDA as part of the security pack. It is written for assurance teams and it is kept current, which is why it does not live on a marketing page. |
Type 2 is the distinction worth asking any vendor for by name, and we hold it currently. A Type 1 report says the controls were designed properly on one day. A Type 2 says they were independently tested as operating across a period — which is only achievable where monitoring runs continuously rather than being reconstructed before an audit. That is the mechanism by which the assurance stays true between reports, and it is the same mechanism whether the thing being protected is a transaction record or something inferred from it.
The platform is only as good as the master data underneath it, and AI makes that worse rather than better: poor unit weights, duplicate SKUs and stale credentials produce answers that are confident, specific and wrong. Building that master is real work, we do it with you, and we would rather quote it than discover it.
Tell us what you would like your consumption data to answer.
If the question needs who, when, against which job and in what condition, that is exactly the layer we operate — and we can show you what the record looks like before you commit to anything.