If you are comparing proposals for predictive maintenance, nearly every one of them contains a number. Some accuracy figure, usually above ninety percent, often with a decimal place. It looks measured. It is the reason the proposal gets taken seriously.
Ours does not contain that number. That is deliberate, and we think it is the more honest answer.
What such a figure actually describes
An accuracy figure is always a measurement on one particular installation: particular machines, particular sensors, particular load, particular failure modes. It describes how well a model performed there.
About your plant, it says roughly as much as another vehicle's fuel consumption says about your own route.
That is not a rhetorical objection. The factors behind such a number differ substantially between two plants in the same industry:
- Sampling rate and sensors. A signal captured ten times a second carries different information than one that lands in a database once a minute.
- History. A model needs failure cases to recognise failures. If a bearing failure was never recorded, nothing can predict one.
- How the plant is run. A machine running evenly across three shifts produces a clearer picture than one alternating between small batches and standstill.
- What counts as a hit. A detection two hours before failure and one two weeks before are both "correct" — the commercial difference is enormous. How the figure was formed is rarely stated.
A percentage compresses all of that into a single value and omits exactly the information you would need to judge it.
The practical reason: you cannot verify it
Suppose a vendor quotes you an accuracy figure from a reference project. What can you do with it?
You cannot ask which plant it was — that is under NDA. You cannot see the raw data. You cannot check how it was counted. You can only believe it or not.
That makes the number not an argument but an act of trust dressed as a measurement. We think trust should be earned differently.
What we commit to instead
Three things, and all three are verifiable before anything is signed:
A fixed price. Agreed in writing before the start. Not billed by effort, not by day rate. If it takes longer than planned, that is our problem, not your invoice.
A fixed duration. One to seven weeks depending on the module, named rather than estimated. You know the end date before day one.
An honest answer. Including the one that says: this is not worth it in your plant. Then it stops there, and you still know more than you did.
The difference from a percentage: these three commitments concern our own offer, not a third party's equipment. We can keep them or break them, and you can tell which.
And the numbers you do get in the end
None upfront — instead, ones that belong to you.
At the end of the pilot you receive an accuracy report for your own equipment, measured on your own data. With a statement of how it was counted, how many cases it rests on, and where the model was uncertain.
If that report comes out unfavourably, that is what it says. This is the point where it becomes clear whether a commitment is worth anything.
What we do quantify
So there is no misunderstanding: some figures we do state, because they are properties of the method rather than results from someone else's plant.
- Six to twelve operational cycles of lead time. That is how early degradation surfaces before it becomes mechanically measurable. It is the reason there is any time to act at all — and the only figure that answers that question.
- Under 40 milliseconds classification per cycle, entirely on embedded hardware, with no offload to a cloud.
- Parallel, independent detection skills, each specialised on its own class of signal and sharing no state — so that a false positive in one area does not distort the assessment in another.
These describe what the system is. Not what it achieved somewhere else.
How to recognise a serious vendor
When comparing proposals, three questions tell you more than any accuracy figure:
- What happens if it does not pay off? A vendor without an answer has only one direction in the offer.
- How is success measured, and by whom? If the vendor sets the yardstick alone, every number is theirs.
- What does the first step cost, and what is decided afterwards? An entry point without a price is not an entry point, it is a negotiation.
You can ask us the same. The answers are on this site before you talk to us.
If you want to know whether AI pays off in your operation at all: a bounded pilot examines exactly that on your real data in four phases — at a fixed price, with an open outcome. Talk to us. If it does not pay off, that is precisely what the report says.