Responsible AI / Last reviewed
AI people can check, and stop.
Responsible AI, for us, means a model only does work someone can check, a named person makes the decisions that matter, and you can see what the model did and why. It shapes how every AI feature is designed, not a review at the end.
Principles
Eight commitments in every AI feature.
They apply whether the work is reading documents, answering questions, automating a step or running an agent. How they apply to agents.
A person decides what matters
AI reads, sorts and drafts. Decisions about care, money, safety or someone’s rights stay with a named person, and the record shows who made them.
Tested on your cases first
Before launch, a feature is measured against real examples with agreed answers, including the awkward ones. It goes live only when it meets the bar you set.
Sources you can see
Answers link back to the document they came from, so staff can confirm them instead of taking them on trust.
Access rules come first
Assistants and agents can only retrieve what the signed-in person may see. That is enforced in the data layer, not left to the prompt.
Logged, with a manual route
Every model call, tool use and approval is recorded, and the process still works by hand when the model is unavailable.
Untrusted content gives no orders
Instructions hidden in emails, web pages or documents are treated as text to read, never as commands, and no single run mixes private data, untrusted content and a way to send data out.
Red lines and an off switch
Actions an AI must never take are blocked in code outside the model, every agent has a named owner, and any run can be stopped at once.
Honest about fit
Not every task should be automated. When ordinary software or a person is the better answer, we say so before anything is built.
Where AI should not decide
Our line, sector by sector.
Each sector has calls that need a qualified person. We design so the AI stops short of them.
- Health & social care ↗
AI should not decide on medication, change a care plan or judge whether someone is safe. Anything it drafts about a person’s care is checked and signed off by a named member of staff.
- Medical education ↗
AI should not set clinical values, rhythms or algorithm steps, and it should not decide whether a learner passes. Those stay with qualified educators.
- Property & field services ↗
AI should not decide whether a hazard is an emergency, sign off a safety check or close a complaint. A competent person makes those calls, and the record shows who did.
- Fintech & professional services ↗
AI should not decide whether to accept a client, clear a sanctions or PEP match, or give advice. Those calls need a named person, and the record has to show who made them.
Rules
The law AI features have to respect.
- UK GDPR, Articles 22A–22D
- Since 5 February 2026, significant decisions can be automated in more cases, provided people are told, can contest the outcome and can ask for a human review. Special category data stays tightly restricted.
- EU AI Act
- Transparency duties, such as telling people they are dealing with an AI system, apply from 2 August 2026. Obligations for high-risk uses in areas like employment, credit and education apply from 2 December 2027.
- UAE and Saudi data protection laws
- Personal data used in prompts, documents and logs falls under the UAE PDPL and Saudi PDPL, including rules on consent and transfers abroad.
We design software to support these obligations. Responsibility for meeting them stays with your organisation and its advisers.
What’s changing
July 2026
EU high-risk rules move to 2027
The AI Omnibus entered into force on 27 July 2026, moving obligations for Annex III high-risk systems, such as those used in employment, credit scoring and education, to 2 December 2027.
August 2026
EU AI transparency duties apply
Article 50 obligations, such as telling people when they are dealing with an AI system, apply from 2 August 2026.
February 2026
New UK rules on automated decisions
The Data (Use and Access) Act 2025 widened when significant decisions can be automated, provided people are told, can contest the outcome and can ask for a human review.
Sources
How it runs
Small, tested, then wider.
Where we run it
AI features are built with the same tools as the rest of the product, so they share its tests, logging and access rules.
Pick a bounded task
One job with clear inputs, a checkable output and a known cost when it goes wrong.
Build the evaluation set
Real past cases with agreed answers, collected with the people who do the work.
Run alongside people
The model suggests and staff decide, so accuracy is measured on live work before it counts.
Widen carefully, keep checking
Only what proves reliable moves to lighter review, and the evaluation runs again on every change.
What you receive
Evidence you can show an auditor.
Each AI feature is handed over with the records that show how it was tested and how it behaves.
Evaluation set and pass criteria
The examples and agreed answers the feature was tested against, and the level it had to reach.
Record of models and data
Which model each feature uses, who provides it, what data it sees and where that data is processed.
Review screens
Where people check, correct or reject what the model produced before it counts.
Decision and step logs
What the model did, what it saw and who approved the result, kept for the period you set.
Fallback procedure
How the work carries on by hand when the model is unavailable or unsure.
Repeatable re-test
A script that runs the evaluation again whenever a model, prompt or setting changes.
Frameworks
Built on published guidance, not our own rules.
We follow these in how we design and test. None of them is a certification we hold.
- NIST AI Risk Management Framework and AI 600-1 ↗
- The US framework for mapping, measuring and managing AI risk, with a profile for generative AI risks such as confabulation and data privacy.
- ISO/IEC 42001 ↗
- The international standard for AI management systems. We do not hold it, but we use its life-cycle and impact-assessment controls as a checklist.
- OWASP Top 10 for LLM and agentic applications ↗
- The security community’s lists of the main risks in AI applications and agents, from prompt injection to rogue agents.
- NCSC agentic AI guidance ↗
- The UK National Cyber Security Centre’s 2026 advice: start small, least privilege, named owners, monitoring and an emergency stop.
- ICO guidance on AI and agentic AI ↗
- The UK data protection regulator’s guidance on AI, its 2026 report on agentic AI, and its code of practice on AI and automated decision-making in development.
Common questions
Questions about responsible AI.
What does responsible AI mean in practice?
A model only does work someone can check, a named person makes the decisions that matter, and you can see what the model did and why. We build those controls into the feature itself rather than reviewing it at the end.
Is AI safe to use in regulated sectors like care or finance?
It can be, for bounded tasks such as reading documents, finding answers or drafting notes, with people reviewing the results. Decisions about care, safety, clients or someone’s rights should stay with a qualified person.
Do you train AI models on our data?
No. We use existing models and, where the provider allows it, choose settings that keep your data out of their training. Your examples are used to test the feature, not to train a model.
Can AI make automated decisions about people?
UK law allows it in more cases since February 2026, but only with transparency, a way to contest the outcome and access to human review. We design those routes in, and usually recommend a person makes the final call.
Are you certified for AI governance?
No. We do not hold an AI management certification such as ISO/IEC 42001. We design software to support your obligations, and responsibility for meeting them stays with your organisation and its advisers.
How do you keep AI agents safe?
We follow the UK NCSC’s 2026 guidance: start with low-risk tasks, give each agent its own identity and the least access it needs, use short-lived credentials, keep a named owner, monitor every step and keep an emergency stop. Red lines are enforced outside the model.
What happens when the model gets something wrong?
Review screens catch errors before they count, the log shows what happened, and the failing case is added to the evaluation set so the fix is tested and stays fixed.