Know what your AI can do, prove what it did
An AI assurance programme for businesses already using AI in real work. We map every AI system you run, test how it behaves, put controls around what it is allowed to do, and leave you with evidence of what it did. It is the AI Control Review applied across your whole estate, with the fixes built in.
Service 06
No 17040048
Programme
From £8,000
Fixed quote after scoping
The price depends on how many AI systems and workflows are in scope. A monthly assurance retainer is available afterwards. Legal advice is not included; we work alongside your advisers.
Who it is for
- Businesses with several AI tools, agents or automations in use and no single view of what each one can access or change.
- Teams whose AI sends messages, edits records or takes actions for customers and who need to show the safeguards.
- Companies preparing for client, investor or insurer questions about how their AI is governed.
What the work covers
Finding every AI system
The register starts with an inventory, and the inventory is usually larger than anyone expects. Beyond the assistant on the website there are automations built by one person, AI features switched on inside existing software, and agents with access to email or a database. For each one we record the owner, the model or supplier, the data it can reach, what it may do on its own and where a person must approve. A system nobody owns is the first finding, because no control survives without someone responsible for it.
Testing behaviour against real cases
We build a test set from your own redacted examples: the ordinary requests, the awkward ones and the cases where the right answer is to refuse or hand over to a person. Each system is run against that set and the results are written up case by case, including failures. Testing against your own material matters because general benchmarks say little about how an assistant handles your prices, your policies or your customers. The same set is kept so it can be rerun before every significant change to a prompt, model or supplier.
Controls on what AI may do
Where controls are missing we design them and put them in place. That means limits on which actions a system can take, approval steps before anything irreversible such as a refund, a deletion or a message to many customers, and clear handover rules when a request falls outside what the system should handle. We prefer controls enforced by the system itself over ones that rely on someone remembering a policy. Each control is documented with what it prevents and how to check it is still working.
Gates that survive change
Controls fail quietly when the thing they check changes and the check does not. We learned this ourselves when a gate in our own agent pipeline reported green for months while checking the wrong branch. So we tie approvals to what was actually inspected, and make gates refuse when the configuration, prompt or code no longer matches what was approved. Where that needs code, such as Slipmark in a deployment pipeline, we build it in your environment. The aim is that the next update cannot switch off a safeguard without someone noticing.
Evidence and data protection
Assurance is only useful if you can show it. The evidence pack records what each system did, who approved changes and how you would reconstruct a particular decision if a customer, insurer or regulator asked. Alongside it, we review the AI estate under UK GDPR as amended by the Data (Use and Access) Act 2025, including automated decision-making, transparency and supplier processing. We are not lawyers and do not give legal advice. We work alongside your advisers and hand them organised material, so their time goes on judgement rather than fact-finding.
What assurance looks like
These are illustrations of the kind of work, not client results.
A register of every AI system
Each assistant, agent and automation listed with its owner, the data it can reach, what it may do alone and what needs a person. Example only.
Gates that survive the next change
Checks that refuse when what was approved no longer matches what runs, so a control does not quietly stop working after an update. Example only.
Evaluation against your own cases
A test set built from your real, redacted examples, run before every significant change, including the cases that should fail. Example only.
What you get
We work with the AI tools and suppliers you already use. Where a control needs code, such as a verification gate in a deployment pipeline, we build it in your environment and document it.
- 01An AI register covering every system in scope: owner, model or supplier, data, permitted actions and approval points.
- 02Testing of each system against named cases, including failures, with the results written up.
- 03Controls designed and put in place where they are missing: approval steps, limits on actions, logging and change gates.
- 04A data protection review of the AI estate under UK GDPR as amended by the Data (Use and Access) Act 2025, prepared for your legal adviser.
- 05An evidence pack: what each system did, who approved it and how you would reconstruct a decision.
- 06A handover session and a written plan for keeping the controls in force.
What changes
- One register showing every AI system in scope, its owner and what it may do.
- Test results for each system, including where it failed and what was changed.
- Approval steps and limits enforced by the systems, not left to memory.
- An evidence pack you can show to clients, investors, insurers or advisers.
- A written plan for keeping the controls in force as things change.
How it runs
- A 20-minute fit call and a short inventory, then a fixed quote for the systems in scope.
- Working sessions with the owner of each AI system to map and test it.
- Regular check-ins with your programme owner on findings and fixes as they land.
- A handover session with the register, test results, controls and evidence pack.
- An optional monthly assurance retainer afterwards.
What you supply
- A named owner for the programme and access to the people who run each AI system.
- Access to configuration, logs and supplier terms for the systems in scope.
- Real, redacted examples of each system's work, including some that went wrong.
Limits
- Not a certification and not legal advice. We prepare the evidence your legal or compliance advisers need; they advise on the law.
- Systems outside the agreed scope are listed but not reviewed.
- Controls only work if they stay in place. We hand over a plan for that; ongoing monitoring is a separate retainer.
How it works
- 01
Scope
A 20-minute fit call, then a short inventory of the AI in use. You receive a fixed quote covering the systems in scope.
- 02
Map
Each system registered: what it can reach, what it may do, where a person approves.
- 03
Test and control
Named test cases run against each system, and missing controls built in where they matter most.
- 04
Evidence
The register, test results, controls and data protection review handed over, with a plan to keep them in force.
Why we take controls seriously
Evidence from our own work
A security gate in our own agent pipeline once reported green for months while checking the wrong branch. The fix became Slipmark, now open source. Our Insights cover documented cases from Air Canada to Replit where the missing piece was an ordinary control. This is our own engineering experience, not client results.
Five real cases and the control each one lackedWhere it fits
A one-day Control Review on a single workflow often shows the need; assurance extends it across the estate, and monitoring keeps the controls working afterwards.
Questions before you start
How much does AI assurance cost?
Programmes start from £8,000 as a fixed quote. The price depends on how many AI systems and workflows are in scope, which we agree after a short inventory.
Is this an AI audit or a certification?
It is an engineering assurance programme: we map, test and put controls in place, and leave you the evidence. It is not a certification and not legal advice. We work alongside your legal advisers.
Does it cover UK data protection?
Yes. We review the AI estate against UK GDPR as amended by the Data (Use and Access) Act 2025, including automated decision-making, and prepare the material your adviser needs.
How is this different from the AI Control Review?
The Control Review is one day on one workflow, with findings. The assurance programme covers every AI system in scope and includes building the controls, testing and an evidence pack.
What is the next step?
Tell us which AI tools and agents you run in the contact form. We arrange a 20-minute call and a short inventory, then send a fixed quote.
Do you need access to our live systems?
We need read access to configuration, logs and supplier terms for the systems in scope, and a safe place to run tests. Where we build controls, we agree the access in writing and prefer a test environment first. We never need more access than the work requires.
What happens if testing finds a serious problem?
We tell your programme owner straight away rather than waiting for the report. Together we decide whether to pause the system, add a temporary control or fix it first. The finding and what was done about it go into the evidence pack.
Can our own team keep the controls running afterwards?
Yes, that is the intention. Every control is documented with what it prevents and how to check it, and the handover plan names who does what. If you would rather we keep watch, the monthly retainer covers that.
Tell us what needs to work better.
Describe the process and the tools involved. If it looks like a fit we will arrange a 20-minute call.