Know where AI runs
Including shadow use — the tools people wired in themselves. Usually half the picture.
AI fails confidently. I keep the receipts
A model returns an answer that sounds right and has no source behind it. People believe it, because it reads well. Every time that happens in my projects I write it down, trace it to the cause and turn it into a rule. The log below is open.
What has piled up so far
These are not headlines about AI going wrong somewhere else. Each line is a case from a working project: what failed, why, and which rule came out of it. The figures are recomputed from the log when the page is built.
Open part of the log
observed 02 Jun 2026 — 05 Jul 202645failures logged and analysed
32rules derived from them
8systems under watch
5classes of failure
The nastiest case so far: a false “verified against the primary source” stamp on a figure that had no primary source at all.
AI governance comes down to four things
The name covers one question: can you trust what the system returns, and what do you do when you cannot.
Catch it before the client does
Rules for what gets checked always, what gets sampled, and what never goes to a model at all.
Trace it to the cause
Model, data, system, process or prompt. Each cause needs a different fix.
Know where machines stop
Decisions, outbound documents, figures in reports. Where a mistake is expensive, a person has the last word.
Verification by evidence
I look at what already happened. Failures are found in real output, not in policy documents.
01
Find where AI is used
Shadow use included — the tools people connected without telling anyone.
02
Trace the output
Take real reports, answers and documents and follow each claim back to its source. Where the source does not exist, the interesting part begins.
03
Classify what failed
Model, data, system, process or prompt.
04
Map the risk
Where a failure is expensive, where it is a legal problem, where it is merely annoying. With likelihood attached.
05
Write the rules
What a machine may decide, what a person reviews, and where a second source is mandatory.
What remains after the work
A risk map
Where the system can fail, what it would cost and how likely it is.
Verification rules
Plain limits: checked always, sampled, or never handed to a model.
The failures, analysed
Each with its cause and the rule that stops it happening twice.
What people usually ask
What is an AI hallucination?
An answer that reads as plausible and has no source behind it. The model is not lying on purpose — it completes text by probability, and invention comes out looking exactly as confident as fact. You catch it by tracing claims to a primary source, never by reading more carefully.
How do you reduce AI hallucinations?
Three things together: demand a source and check it, ground the model in your own data instead of general knowledge, and make verification a permanent step of the process. Any one of the three on its own fails.
We are small. Is this early for us?
The question is whether anything leaves the model and reaches a client, a document or a decision. If it does, the risk already exists.
What if we do not use AI at all?
You probably do. People connect tools themselves and rarely report it — shadow use turns up almost every time.
Will you show me your own system?
I will show what it produces: the figures and the analyses are open. The tool itself stays internal. The verification rules that grew out of it are yours.
Do you work outside Russia?
Yes. I am based in Thailand and work remotely across time zones. Contracts and invoicing go through a Russian sole proprietorship — details in the footer.
Where is AI already running in your company?
Tell me what is connected and where the output goes. I will say whether there is anything worth checking and where I would start.
Write to me on Telegram — it reaches me directly and I answer myself, usually the same working day.
Or email info@iterans.ru.