Trust center
How we build AI responsibly
We use AI where it helps, test it before anyone relies on it, keep people in charge of consequential decisions, and say so when AI is the wrong tool.
In short
Responsible AI starts with asking whether AI is needed at all. When it is, we classify the risk of the use case, bound what agents may do, keep a person in the loop for consequential actions, and evaluate quality, safety and cost before and after launch.
Last updated
01 When to use AI
When do we recommend AI, and when not?
Before building, we ask whether AI is necessary, or whether conventional software or automation would be more reliable. Sometimes the honest advice is that AI could help, but conventional automation may be more appropriate.
We do not force AI into every problem or promise outcomes that cannot be substantiated.
02 Controls
Which controls go into the AI systems we build?
AI-related controls can include:
- Use-case risk classification
- An inventory of approved models
- Privacy and intellectual property reviews
- Dataset provenance
- Human oversight
- Model evaluation and performance monitoring
- Incident reporting
- Appropriate transparency with the people who use the system
03 Agents
How are AI agents kept within bounds?
Agents should work through explicitly defined responsibilities and permitted tools. Autonomy should increase only when testing, business requirements and risk assessments justify it. Their controls can include:
- Least-privilege access
- Input and output validation
- Explicit execution states and termination conditions
- Bounded retries, timeouts and budgets
- Human approval for consequential or irreversible actions
- Protection against prompt injection and unauthorized tool use
- Audit trails and execution traceability
- Safe failure, fallback and escalation
04 Evaluation
How is AI tested before and after launch?
Acceptance thresholds should be defined before deployment, and testing should use representative cases and failure scenarios. Evaluation should continue as models, prompts, data and requirements change. It can cover:
- Answer correctness and groundedness
- Retrieval precision and citation accuracy
- Tool selection and execution correctness
- Resistance to prompt injection and unauthorized disclosure
- Latency, reliability and cost per successful task
- Human acceptance and escalation rates
05 Sensitive domains
What about healthcare and other sensitive domains?
In healthcare, AI is never presented as a substitute for qualified clinical judgment without appropriate validation, oversight and regulatory authorization. Sensitive information, consent and data minimization shape the design.