DE|EN

Quality Assurance, before uncomfortable questions arrive.

I write the gate as a script. It runs before delivery and holds it back when an output misses the standard. What it checks is set by your product and your evidence obligations.

A Quality Gate is a check that runs as a script before delivery and tests a result against conditions fixed in advance. If one of them is not met, the delivery does not go out: not a notice someone could overlook, but a halt. It is needed wherever an AI produces something a client, an auditor or a court gets to see, and where you answer for every sentence in it. Not because AI is often wrong, but because nobody notices when it is: a system that was right yesterday and misses something today does not raise its hand, it keeps delivering. A gate moves the control from the spot check to the rule, automatically, traceably and every time.

Contact

How it works

PhaseWhat happensPrice
1First call 45 minutes. What the case is, what is at stake, whether this is worth doing at all. free
2Specification Sessions with your specialists, one visit on site, a record of the current state. The result is the specification of your gate and a fixed price for the build. 4,500 euros, 6,500 in regulated domains
3Build The gate gets built, installed and handed over, until it runs. fixed price, set by the specification
4Review Once a quarter the gate runs against everything that changed. Report and corrections. 2,500 euros per quarter, optional

You can stop after any phase and keep what you have. The specification from phase 2 holds its value even if somebody else builds it. All prices exclude VAT.

What you keep, even if you stop

  • After the design phase. The written specification: what “correct” means for each of your output types, which conditions are checked against it, which of those run automatically and which a person answers. Plus the fixed price for the build.
  • After the build. The running gate inside your environment. You can read the code, hand it on and run it without me. Plus evidence an auditor reads without running your model again.
  • After each review. A report on what changed and what failed because of it.

You end up with something you can show to someone. And something that can fail.

Contact

Have a short assessment written

Enter your company address. The run reads up to six of your public pages, looks for places where a statement goes out without any sign of who checked it, and writes a short assessment with five quality assurance recommendations, each with its reasoning and its source. Then the gate runs: 33 named checks, not against your website but against the document you are about to take away. That is the difference between a list of defects and an assessment.

We do not store the address and we do not keep the report. What is read is what is publicly available, as any visitor sees it. The report is built in your browser and stays there.

Four examples, four industries

A gate looks different in every industry, because the risk does. What stays the same: the damage comes not from the AI being wrong, but from nobody noticing.

Consulting and analysis

Where the AI sits
The AI researches and helps write documents that land on a client's desk.
What goes wrong quietly
In October 2025 Deloitte Australia repaid part of a AU$440,000 engagement to the Australian government. The report contained an invented quotation from a federal court judgment and references to studies that do not exist. It was not Deloitte who found it, but an outside researcher who read the footnotes.
What the gate checks
Every source resolved at the time of writing, with the quoted sentence found in it. Three pieces of evidence per headline claim, and independent ones: three outlets reprinting the same press release count as one. Every figure carries its formula.
Where it hooks in
The gate sits before the document goes to the client. It halts the delivery, not the draft.

Education and learning material

Where the AI sits
The AI produces and adapts material that reaches learners.
What goes wrong quietly
A school in Louisville sent AI-generated material full of errors home on the first day of term, and it became a news story. In 2025 Springer Nature published two titles carrying invented or untraceable references. Here the customer is a parent or a ministry, and the damage is not a discount. It is a photograph of the worksheet with your imprint on it.
What the gate checks
Every factual claim checked against a stored subject source. Curriculum reference stated per exercise and validated against the target year group. Reading level measured rather than estimated.
Where it hooks in
The gate sits before release for reproduction, not inside the editorial process. It does not replace an editor; it catches what a sample no longer reaches.

Medical software

Where the AI sits
The AI produces findings, triage ratings or risk scores that a clinician sees.
What goes wrong quietly
The model was validated once, at approval, against a population it no longer sees. Longitudinal work on deployed clinical systems shows the decline is gradual, appears first as calibration drift, and fails to trigger exactly the alarms teams watch. Only around 9% of FDA-registered AI medical devices ship with any post-market surveillance plan at all.
What the gate checks
No finding without a confidence value and model version in the log. Weekly comparison of the output distribution against the approval population, against a defined threshold. Any deviation from a prior finding is flagged rather than silently replaced.
Where it hooks in
The gate sits between inference and release to the hospital system. It does not hold up individual findings; it fires when the distribution moves.

Production and inspection

Where the AI sits
The AI decides pass or fail on the line, on real parts.
What goes wrong quietly
Good parts wrongly rejected are noticed at once, someone complains before lunch. Bad parts let through are not: they look exactly like parts that were never inspected, unless a confidence value is logged for every single inspection. Optics age, batches change, the decision drifts slowly, and escapes rise before any alarm fires. 75% of manufacturers had a recall in the past five years; 48% put a single one at $10M to $50M.
What the gate checks
Confidence value logged per inspection, not just the verdict. Daily run of a reference set of known bad parts against a defined minimum detection rate. Threshold tied to your economics: scrap against escape, not against a vendor default.
Where it hooks in
The gate sits beside the line, not in it. It does not stop production on suspicion; it reports when the reference set stops being recognised.

Why this is not a template. Writing down a check is easy. Setting it correctly is the work. Too tight, and the gate halts every morning until someone switches it off. Too loose, and it never fires. That threshold depends on your data, your costs, and on what you would have to justify to an auditor. This is exactly where the questions start that a template does not answer.

Contact