Case studies
What assurance looks like in practice.
Accounts of recent engagements: the system, what we found, what changed and how we keep testing it. Client identities and identifying details are withheld under confidentiality agreements.
Engagements
Four systems. Four decisions that needed evidence.
Two are the AI products most enterprises are shipping now. Two are industry-specific systems that act in the real world - approving credit, taking payment, holding a table - where the cost of being wrong is immediate. In each, the team knew what the system was meant to do, and needed to know what it would actually do.
Enterprise software
01Customer support assistant
Giving a support assistant real authority, with evidence it won't be misused.
A global B2B software company wanted its support assistant to adjust subscriptions and issue credits. We showed how customers could turn those tools against the company, and how to close every path before launch.
Read the case studyProfessional services
02Internal knowledge copilot
Proving an internal copilot respects the permissions its documents already had.
A professional-services firm was rolling a knowledge copilot out to more than ten thousand staff across client files, mailboxes and HR records. We found where it retrieved more than the user could see, and fixed it at the source.
Read the case studyFinancial services
03Credit-decisioning assistant
A credit assistant that can answer a regulator's questions.
A lender in several regulated markets built an assistant to help underwriters work faster. We measured whether it stayed decision support in practice, and gave the risk function the evidence to say so.
Read the case studyHospitality
04Ordering and reservations chat assistant
A chat assistant trusted to take orders, bookings and payment across every site.
A multi-site restaurant group let an LLM-powered chat assistant take orders, bookings and payment with no staff in the loop. We found where it could be talked into giving away money and tables, or answering allergen questions it could not ground - and closed every path before launch.
Read the case studyBegin with evidence