How do you trust what an AI says about your business?
You constrain it, and you publish the constraints. An AI system that analyses your operation is only trustworthy if it can tell you what it does not know, and if inventing an answer is structurally harder than admitting a gap. Anyone can claim their AI is reliable; the claim is cheap. What follows is the mechanism: the rules Omni enforces, stated plainly, including the caveats.
An AI that tells you what it does not know
Absence is never coerced into a number.
A missing price is never rendered as zero, in the interface or in analysis. A missing figure makes the caller ask for it. Zero and unknown are different facts, and the system refuses to conflate them.
Fabricated precision is banned.
No scores out of 100, no star ratings, no "percentage compliant" figure. The underlying evidence does not support that precision, so the product does not display it.
Model output is checked against real records.
When the AI refers to a person, a system, or a data object, the reference is verified against the organisation’s actual catalogue. A hallucinated reference to something that does not exist cannot be saved.
Process generation is gated behind a gap report.
When the AI builds a process from a document, it reports what the source failed to say instead of inventing detail to cover the silence. The gaps are the deliverable alongside the process, and a person resolves them.
A redesign cannot silently drop a requirement.
Every step, system, actor, and data object in a replaced process must be accounted for in the future state: carried through, migrated, or consciously dropped with a recorded reason. The tool blocks progress past an unexplained omission.
Every instruction given to the AI is version-controlled and checksummed.
It can be established afterwards exactly which wording produced a given result. When an output is questioned, the answer is a record, not a reconstruction.
Financial figures are structurally withheld from the AI model.
Your financial data is not in the prompt. Where money enters an analysis, it is joined to the AI’s output by conventional code, not generated by the model.
A quotation that is not in the source is not stored.
When the AI proposes a requirement, it must quote the words it read. Those words are searched for in the stored material before the record is written, and a quotation that cannot be found is dropped rather than saved. The check runs at the moment of writing, not when somebody happens to click into it, so a citation to nothing never reaches the register.
The AI never claims to have done something it cannot do.
Nothing is written to your records by a conversation. Every change an AI proposes arrives as a list you accept row by row, what it could not do is shown with the reason rather than quietly dropped, and what the assistant is unable to do is stated along with where the work is done by hand.
How any figure we quote is derived
Tools in this market publish confident returns (a percentage saved, a payback period) generated by a language model with no stated derivation. We do not. Any figure Valstra puts in front of you follows four rules:
- The assumptions are stated with the figure, so you can challenge them.
- Estimates are ranges, not point estimates dressed as certainty.
- A named person on your side is accountable for each expected benefit.
- An estimate and a realised result are never conflated. Realised means measured against a baseline set before the work started, and we only use the word when that measurement exists.
Where the enforcement is not finished
Two caveats, stated because this page would be worth less without them. Audit-log immutability is currently enforced in application code, with a database-level hardening step still outstanding. And the automated code-review gate that checks these principles in our own development advises rather than blocks: a human decides. We would rather you knew.
Put it to the test
A demo is not a slide deck: we show you a living model of your own operation, and you are welcome to ask it something it cannot know. What it does next is the product. The full list of what is and is not built today is on the capabilities page.
Book a demo