
AI Practice/AI Assurance
Assure · Independent testing
Prove it before anyone asks.
Independent evaluation and red teaming for AI systems, ours or anyone else's. We test them against cases from your own world, score them against a simple baseline, and report plainly what passed, what failed and why.
- The method behind our public leaderboard
- Any vendor, any model
- Tested inside your walls
Why test AI
Confident is not the same as correct.
AI systems fail quietly. Testing them on your own cases is the only way to know how.
Confident and wrong
Fluent answers hide stale, partial and conflicting evidence. Users rarely notice until it matters.
New ways to attack
Chatbots and agents can be tricked into leaking data or taking actions nobody approved.
Buyers now ask for evidence
Boards, regulators and public buyers increasingly ask how an AI system was tested before they rely on it.
Services
Independent testing, before and after go-live.
Built from the method we use on our own products, and run on your cases, inside your environment.
AI evaluations
Test cases built with your own experts, scored against a simple baseline, so you see what the AI adds and where it fails.
- Cases from your real work, not public benchmarks
- Behaviours scored one by one
- A plain report your board can read
Red teaming
We attack chatbots and agents the way a determined user would: data leakage, prompt injection, unsafe actions and harmful answers.
- Text, document and agent attacks
- Each finding rated and reproducible
- Fixes re-tested before sign-off
Product claim review
For vendors: we check that what your product sheets and bids say about AI is true, and help you evidence it.
- Claims tested against real cases
- Evidence packs for bids and due diligence
- Gaps named before a buyer finds them
Buyer-side assurance
For buyers: we help write the AI part of a tender and test the shortlisted vendors on your own cases before you sign.
- AI requirements you can actually test
- Shortlisted vendors scored side by side
- A recommendation backed by evidence
Ongoing monitoring
Re-tests after every model or data change, so quality does not drift silently after go-live.
- Your evaluation set, re-run on a schedule
- Alerts when a behaviour gets worse
- A short report every quarter
AI governance as a service
We keep your register of AI systems current, re-test them on a schedule, and report to the board each quarter.
- One register of every AI system you run
- Risk, owner and last test for each
- Board reporting without the busywork
How we test
Built so the AI can fail.
The same method we use on Kybern Prime in public. A test that cannot fail proves nothing.
- 01Build the cases
With your experts, from your real work, including the traps: missing facts, conflicting sources, outdated rules.
- 02Set a baseline
A simple, honest comparison, so you can see what the AI actually adds over doing nothing clever.
- 03Score behaviour by behaviour
Not one accuracy number. Each behaviour that matters is scored on its own.
- 04Report plainly
What passed, what failed and why, with every finding reproducible and mapped to the framework you use.
Behaviours we score, among others
- Missing information stays unknown
- Source conflicts are kept, not silently resolved
- Recognises when the model may not fit
- Revises when reality changes
- Abstains instead of guessing
- Decides when the evidence is enough
Frameworks
Findings your auditors already recognise.
NIST AI Risk Management Framework
Findings mapped to the framework's functions, so they slot into your existing risk process.
OWASP guidance for AI applications
Red-team findings organised against the most common weaknesses in AI and agent applications.
Your sector's rules
Banking, public sector and defence requirements, and your own internal policies.
Questions
Questions about assurance.
Something else on your mind? Ask us directly.
Can you test a system you did not build?
Yes. We test AI from any vendor, open or commercial models, chatbots, agents and document AI. Independence is the point.
Can testing happen inside our environment?
Yes. We can run evaluations and red teaming on your premises and fully offline, so your data and the system under test never leave.
Will you publish our results?
No, not without your written permission. We publish our own results; yours belong to you.
We are a vendor. Can you help us prepare for a buyer's test?
Yes. A product claim review checks what your materials say about AI against real cases and builds the evidence pack before a buyer asks.
AI Assurance
Tell us what you need to trust.
A system you run, a product you sell or a vendor you are choosing. We will suggest how to test it.
- A conversation with someone who does the work, not a script.
- A suggested first step, sized to your problem.
- A clear view of scope, cost and how success is measured.
Prefer to talk now? Call +91 98150 23001 or write to reach@requisitetech.com.
The AI Practice
Explore the AI Practice
Strategy & Advisory
Know where AI pays, and who owns it. Decision Diagnostic, AI strategy, organisation design, governance and decision exercises.
Explore Strategy & Advisory →BuildBuild & Deploy
AI inside the systems you already run. Data readiness, solution development, document AI, agents, private AI and running it all.
Explore Build & Deploy →AI Assurance
Evaluations, red teaming, claim review, buyer-side testing and ongoing monitoring.
Requisite Academy
Build the capability, not a dependency. Programmes for boards, leaders, public servants, engineers and partner teams.
Explore Requisite Academy →