---
title: AI Assurance
seoTitle: AI Governance and Testing for Compliance Monitoring
description: Test AI review rules on your own past calls before they go live, and keep every version, change note and test run on record for audit.
subnav: true
finished: true
---

> [!herorow]
> AI Assurance
>
> # Test every AI rule change *on your own calls*
>
> Try each change on past calls from your archive and on test conversations the AI drafts for you, before it goes live. Every version, change note and test run stays on record, ready for your auditors.
>
> [Get a demo](/request-a-demo/) [See pricing](/pricing/)
>
> ![The Argus Archive Evaluate screen with a policy tested on a set of conversations](/media/uploads/2026/10/20-policy-test.jpg)

> [!statement]
> Prove
>
> ## Every rule change tested on your own calls, *and kept on record.*
>
> ### Tested on real calls
>
> Past calls from your archive, plus rare cases the AI drafts for you.
>
> ### Every change on record
>
> Draft and live versions, a note on each change, and a way back.
>
> ### Proof for audit
>
> Each test run kept with its date, its version and who ran it.

> [!table]
> Frameworks
>
> ## What the frameworks ask
>
> AI rules and guidance agree on one thing: trust in an AI system has to be shown, not assumed. Which of them apply depends on your systems and where you work.
>
> | Framework | What it asks | How Argus Archive helps |
> |---|---|---|
> | UK AI assurance (DSIT) | Measure, evaluate and show that an AI system can be trusted | Every rule tested on your own calls, with each run kept |
> | EU AI Act | Testing, documentation and human oversight, more as the risk rises | Tests before publishing, versions with change notes, people decide what goes live |
> | NIST AI RMF | Govern, map, measure and manage AI risk, with documented performance | The AI's answer to each rule and the alerts that would fire, kept for every run |
> | ISO/IEC 42001 | A managed, auditable process for how AI is used | Version history, publishing records and publishing by role |

> [!highlights]
> Record
>
> ## What stays on record
>
> Everything an auditor asks about a rule, in one place.
>
> ### Test sets
>
> The conversations each rule was tested on, and where each one came from.
>
> ### Test runs
>
> The date, the policy version and who ran it, for every run.
>
> ### Answers per rule
>
> What the AI answered for each rule on each conversation.
>
> ### Alerts that would fire
>
> Each alert's condition, and whether it was triggered.
>
> ### Change notes
>
> What changed in each version, listed for you, with your own note.
>
> ### Publishing
>
> Who made each version live, and when.

> [!image]
> Test
>
> ## Try a change before your team sees it
>
> Run a test set against your draft and check every result. Nothing reaches reviewers, Teams channels or your own systems until you publish.
>
> - **Everyday calls and rare ones**, including past misses, in one test set
> - **A whole set or one conversation**, run in one go
> - **The AI's answer** to each rule, for each conversation
> - **Which alerts would fire**, with their conditions
> - **Stop at any time**, then change the rule and run again
>
> ![One Argus Archive test result with the AI's answer to each rule and the alerts that would fire](/media/uploads/2026/10/30-test-result-drawer.jpg)

> [!grid]
> Catch
>
> ## What testing catches
>
> Run the same test set after every change, and these show up before a customer, an auditor or a regulator finds them.
>
> ### A fix that breaks something else
>
> A change helps one call but changes the answer on another.
>
> ### Alerts that fire, or don't, when they should
>
> Cases you've already found once stay in the test set.
>
> ### Slow drift
>
> Small changes that add up over many versions.
>
> ### Different answers on similar calls
>
> Near-identical conversations that get different results.

> [!faq]
> FAQ
>
> ## Frequently asked questions
>
> Still have questions? Ask the team that built it.
>
> [Talk to us](/contact/)
>
> ### What is AI assurance?
> Showing, with evidence, that an AI system works as intended. For AI review that means testing each rule before it goes live and keeping the record of every test and version, so you can answer an auditor without guessing.
>
> [AI assurance explained](/blog/ai-assurance-explained/)
>
> ### How is a test set different from a benchmark?
> A benchmark tests how capable an AI model is in general. A test set is your own: the calls, edge cases, past misses and rules that matter to your business.
>
> ### How often should we test?
> After every rule change and before each publish. Run the same test set each time, so you can compare the results with earlier runs.
>
> ### Can I test on real calls?
> Yes. Import past calls from your archive into a test set. You can also paste or type a conversation, or describe one and let the AI draft it.
>
> ### Do test runs send alerts?
> No. A test shows which alerts would fire and why, but raises no alert and sends nothing to Teams or your own systems.
>
> ### Can I go back to an earlier version?
> Yes. Turn any earlier version into a new draft, test it and publish it again. Every version that was live stays on record.
>
> ### Who can publish a change?
> Only people whose role allows it. Each publish records who made the version live and when.
>
> ### How does it relate to AI Workflows?
> AI Workflows is where you write your policies and rules. AI Assurance is how you test each change and keep the record, for rules that run after the call or live.
>
> [AI Workflows](/product/ai-workflows/)

> [!articles]
> Articles
>
> ## Worth reading
>
> How AI review is tested, and what regulators expect from AI in compliance.
>
> - /blog/ai-assurance-explained/
> - /blog/the-fcas-mills-review-why-ai-makes-communications-archives-more-important-than-ever/
> - /blog/regtech-summit-london-2026/