The UK government defines AI assurance as "the process of measuring, evaluating and communicating the trustworthiness of AI systems". Put simply: knowing how your AI behaves, and being able to show it.
The term is used in two ways, and it helps to separate them from the start:
- Your own assurance. Your organisation does it: it sets the rules, evaluates its systems and keeps the evidence.
- Independent assurance. Someone from outside does it: the body that certifies your ISO/IEC 42001, a notified body or the market surveillance authority. They review what you show them and give their opinion or their certificate.
The second relies on the first. An outsider can only check what has been done and written down inside. This article is about your own assurance. It is quality assurance: what your organisation does to be confident that its AI meets what is asked of it.
AI assurance in six questions
What it is
Checking that what you decided about an AI system is actually happening, and having the evidence to hand. It takes three tasks: govern, evaluate and prove.
- Who
- Whoever builds the AI, and whoever uses it in their processes: management, product, the technical team, quality and the teams that use it. It is asked for by customers, auditors, notified bodies and the market surveillance authority.
- When
- Before the system goes live, and every time the version, the data or the use changes. For Annex III high-risk systems, the AI Act obligations apply from 2 December 2027.
- Where
- In your quality management system, the technical documentation and the system's logs. That is where an auditor looks.
- Why
- Because you will be asked for it. And without evidence, you can't be sure your AI does what you say it does.
- How
- By governing, evaluating and proving. Each task has its own section below.
What people mean by AI assurance
The term is used today for four different things:
- Certifying. An accredited body assures that an organisation meets a standard, such as ISO/IEC 42001.
- Testing. A laboratory assures that a system has passed specific tests.
- Monitoring. A tool assures that certain behaviour of the system is logged and watched.
- Getting ready. An organisation prepares so that someone else can review what it does.
All four rest on the same thing: controls that work, and evidence that they work.
From slide deck to infrastructure
Governance, risk and compliance (GRC) often lives in slide decks, spreadsheets and risk matrices. They describe the controls, but they don't run them or leave a trace.
AI assurance means moving that into infrastructure. Each control is applied in the systems where the AI is built and used, checked there, and leaves dated evidence without anyone having to write it up at the end. When someone from outside reviews it, the evidence already exists.
In the EU, this usually goes by the name of the law: AI Act compliance, AI governance and conformity assessment. Governance decides what to control. Assurance provides the proof that it is controlled, and that proof is what a conformity assessment reviews.
Govern
Governing means getting the people who decide on, build and use the AI working to the same rules. You set what each system may be used for, who is responsible for it, and what has to happen before you change it or start using a new one.
How you do it depends on your role:
- If you use AI in your processes, even if someone else built it, the reference standard is ISO/IEC 42001. It helps you know which AI you have, who is responsible for each system and how to monitor it. If the system is high-risk, Article 26 of the AI Act also asks you to use it according to its instructions, with human oversight, and to keep its logs.
- If you build AI into your products, the AI Act asks you for a quality management system (Article 17), and the standard that spells it out is EN 18286. If you already work with ISO 13485 or ISO/IEC 42001, you are already part of the way there: the differences between ISO 42001 and EN 18286.
Evaluate
Evaluating means checking what can go wrong, who it affects and whether the controls work. It is done with tests on the model and with the risk management required by Article 9. Here is how to identify the risks of your system.
An evaluation is useful when it says which version was tested, with what data, what the results were and what was left untested. Without that, nobody can repeat it or rely on it.
Prove
Proving means having the evidence of what has been done, up to date and ready to show. Under the AI Act, that is the technical documentation in Annex IV and the logs the system keeps by itself (Article 12).
The documentation has to be able to explain any decision: who made it, on which version and with what data. If it is written by hand at the end, it goes out of date as soon as the model changes. That is why it should come out of the team's daily work.
An example
This case is made up. A company sells software that sorts chest X-rays by urgency.
- Govern. It decides the software is only used on adults, that a radiologist always reviews the result, and that every new version goes through the quality team before release.
- Evaluate. It tests version 2.3 on 5,000 X-rays from three hospitals. It measures how many urgent cases are missed, by age and by sex. It notes that it has not been tested with portable equipment.
- Prove. It stores the results with the version, the date and the data used.
Six months later, a new hospital wants to use it with portable equipment. Because the limit was written down, the company knows it has to test it before saying yes. And when the notified body asks about version 2.3, the answer is already there.
Assurance and insurance
The two words sound alike but mean different things. Assurance gives evidence of how a system behaves. Insurance covers financial losses under a policy. An insurer may ask you for that evidence, but having it does not guarantee cover or a lower premium.
AI assurance consulting at Venturalítica
If you searched for AI assurance consulting, this is what we do. We implement AI governance and assurance systems. We work on your own assurance with you, so you reach independent assurance with everything ready. We don't certify: with any one client, we either implement or audit, never both. This is our approach to AI assurance consulting.
