Your Decision Support System for
Critical Decisions
Independent decision research for leaders who use AI in consequential decisions
Your AI agrees with you. Who checks the decision?
We build guardrails for AI agents by starting with the people they serve. First we study how your leaders and experts actually decide: their knowledge, skills, biases and outcomes. Then we shape the prompts, agents, guardrails and governance around that.
Built for CEOs, CTOs and the risk and AI leaders who answer for decisions made with AI.
- human decision research
- 70%human decision research
- AI guardrails and governance
- 30%AI guardrails and governance
- working days
- 15working days
- decision, fixed scope
- 1decision, fixed scope
In the debate
You make your decisions with AI, without guardrails and checks, amplifying both biases.
The harder ones are answers that are plausible, confident and slightly wrong, and review steps that no longer catch them.
Human in the loop is turning into a rubber stamp.
Wharton researchers Steven Shaw and Gideon Nave describe "cognitive surrender": adopting AI answers with little scrutiny. In their experiments, accuracy fell when the AI was wrong, while confidence rose.
Our view: Real oversight needs time, context, expertise and the authority to override. We test whether your review step has all four.
AI tells you what you want to hear.
A study in Science compared 11 leading models with human responses to people asking for personal advice. The models endorsed the user 49% more often than people did, including in cases where the crowd had judged the user to be in the wrong.
Our view: An AI review is not an independent check when the prompt already contains the preferred answer.
AI now acts. Oversight has not caught up.
In an October 2026 vendor survey of 417 governance, risk, compliance and audit leaders in North America and Europe, 34% said their organisation had acted on an inaccurate AI agent output. Only 9% had redesigned workflows for agents.
Our view: The authority you give a system should match your ability to govern it. We map who can approve, override and stop.
Who answers when the model is wrong?
Our view: Every decision shaped by AI needs one named person who answers for the outcome and can explain why the recommendation was accepted.
Self check
Is your human in the loop a rubber stamp?
If two or more of these sound familiar, your oversight may exist on paper only. These are the patterns we look for first.
Our method
Guardrails only work when they fit the people they protect.
AI agents need guardrails. But a guardrail is only as good as its understanding of the decision and the person making it. That is why we start with the human.
Understand how your people decide.
Most of the work happens before anyone touches a prompt. We interview the people who own the decision and map how it really gets made.
- Interviews with the decision makers and the experts around them
- Maps of the real decision process, including informal handovers
- The specific knowledge and skills each decision depends on
- The biases and pressures that shape judgement
- How past decisions connect to their outcomes
Shape the AI around that decision.
Then we tune the AI to support the person who decides, using what the interviews revealed. We work on your premises or externally, whichever you prefer.
- Prompts and context built from your experts' knowledge
- Language checks: does the model read your terms and phrasing the way your experts mean them?
- Agent instructions and clear decision rights
- Guardrails for when the AI must stop, flag or hand back to a person
- Escalation and override rules people can actually use
- Governance: named owners, review records and oversight