Case studies
We know how AI fails because we have run it in production.
Our team has built and run AI for public services and enterprises. Each of these projects taught us something we now test for. We never assess systems we built, and we never sell fixes for what we find.
Scottish Government and NHS 24 Public sector and healthcare
Citizens describe what they need. AI finds the right service.
The challenge
Public services are spread across many organisations and websites. Citizens often don't know what a service is called, who runs it or where to look, so they give up or contact the wrong place, which pushes demand onto already busy front-line teams.
The work
- A conversational AI search engine where citizens describe their need in their own words, with no service names or jargon required
- Location-aware results that point people to the right service near them
- Analytics showing what citizens are asking for, where demand is growing and where services have gaps
- Safety and accuracy designed in from the start, and tested with our AI Test Harness before it reached the public
What we now test for
- People rarely ask the way a system expects, so we test with the questions real users actually ask
- Pointing someone to the wrong health service is a safety risk, so we check answers against a known-correct set
- Different groups describe the same need differently, so we compare outcomes across groups
Thoughtworks Technology
Employee issues solved instantly, and escalated intelligently when they can't be.
The challenge
Employees were losing time to everyday IT and workplace issues. The answers existed, but they were scattered across internal documents and people's heads, and complex problems reached the service management team with too little detail to act on quickly.
The work
- An AI knowledge base that answers employee questions from the company's own trusted content
- AI agents that take over when an issue is too complex to self-serve
- Automatic tickets for the service management team, with a clear description of the issue already written
What we now test for
- An assistant should only answer from trusted content, so we test for answers it has made up
- Agents that raise tickets act in real systems, so we test what they can do, not just what they say
- Hand-off to a human has to happen when it should, so we test where escalation fails
Crimson Enterprise
From scattered AI experiments to a strategy that scales across the enterprise.
The challenge
Like many organisations, Crimson had AI interest and activity across the business, but no shared direction. Without alignment between strategy, priorities and delivery, AI risked staying a set of disconnected pilots.
The work
- Alignment between AI and business strategy, agreed with the leadership team
- A prioritised set of use cases, chosen on value and readiness to deliver
- A practical roadmap to scale AI across the enterprise, with governance built in
What we now test for
- When AI is spread across a business, leaders need one view of the risk, so we report in terms a board can sign off
- Governance added at the end rarely holds, so our evidence is built to be re-run as systems change
“FireBreak told us plainly where our AI plans carried risk and what had to be in place before we scaled. That independent view gave our leadership team the confidence to move forward.”
Find out what your AI will get wrong, before anyone else does.
Tell us what your AI is meant to do, and we'll show you how we would test it.