People stay accountable
Every agent has a named human owner. Decisions that affect a person (credit, underwriting, claims, benefits, clinical care, employment) are made or approved by people, never by an agent alone.
Our policy for every model and agent we design, build or run, across all industries.
Every agent has a named human owner. Decisions that affect a person (credit, underwriting, claims, benefits, clinical care, employment) are made or approved by people, never by an agent alone.
Agents and models are tested against agreed evaluation sets before release; overrides, errors and drift are tracked in production and feed back into evaluation.
Answers are grounded in governed data and documents, show their sources or the query used, and figures are checked against the data. If the data does not support an answer, the system says so.
In credit, underwriting, claims, benefits and hiring we test for bias across relevant groups, document the results and agree mitigations with your risk and compliance teams.
People are told when they are interacting with AI or when AI drafted something they receive, and how to reach a person.
Personal data is minimised and masked before it reaches a model, in line with the Digital Personal Data Protection Act, 2023. See our security & data protection practices.
AI incidents are logged, triaged and reported to the owner. One switch halts all agents; cases fall back to people until the issue is fixed.
Plans, tool calls, guardrail checks, approvals and decisions are logged so any AI action can be audited and explained.
Each agent runs at one level, agreed with its owner and your risk team. Most start at levels 1-2; higher levels apply only inside policy limits.
Monitors and reports; takes no action.
Drafts a recommendation; a person does the work.
Prepares the full action; a person approves every case.
Acts on its own inside policy limits; everything else is escalated.
Acts end to end; people review samples after the fact.
This states alignment with these frameworks; it is not a certification.
Policy, approvals, autonomy limits and the kill switch in our live agent demo (banking sample data).
Agent governance demo Discuss AI governance