Behavior & consequences
A finished task can still be a bad decision. Bentham helps organizations understand human choices and bring clearer oversight to AI decisions — making the context, affected interests, and potential consequences available for examination.
The problem
“Please take that paragraph out.”
Imagine an accurate but inconvenient finding being removed from a report. The request is satisfied, but people relying on the report could lose information they need.
THE HUMAN CHOICE
What was noticed, believed and expected — and what would make a different choice workable?
THE CONSEQUENCES
The requester may be satisfied while other people bear a cost or lose an opportunity.
THE GOVERNANCE QUESTION
Make the assessment reviewable. Decide where a concern should be recorded, reviewed, or connected to a permitted intervention.
Capability is not a substitute for an explicit, reviewable assessment.
An illustrative situation, not a claim that every AI agent would comply.
What we do differently
Bentham's method examines affected parties systematically, including people outside the immediate conversation. Each party's gains, burdens, and uncertainties remain visible in a separate account. A benefit to one party does not erase a burden to another. The account can be challenged when a person, fact, or consequence has been missed.
Separate accounts
A missing quantity remains unknown, not zero. This is a synthetic illustration, not a clinical recommendation. Denying funding does not automatically mean discharge or the end of care.
Organize the decision context, supplied options, affected parties, and evidence. Distinguish recorded information from assumptions and Bentham's interpretation.
Compare gains and burdens using an explicit utility-based method. Keep each party's account visible and identify quantities that cannot yet be established.
Provide a Logic Receipt that exposes the basis and uncertainty of the assessment. Where governance is integrated, link the policy decision to evidence of what happened next.
Two products
A shared decision method supports different users, outputs, and responsibilities. HBM helps professionals understand human choices. ADL brings a separate evaluative function alongside AI agents.
HBM · Human Behavior Model
What are people likely to do under these conditions — and why?
For trained advisers, safety leaders, and professionals. HBM helps examine what people notice, believe, anticipate, and seek to protect. The forecast is a means, not the end: use the explanation to investigate contributing conditions and develop a specific change worth testing.
A conditional forecast for the people and behavior you describe.
Explore HBMADL · Autonomous Decision Layer
What decision is the agent facing, who could be affected, and what response have you authorized?
For hospital AI, safety, and governance teams. ADL is being developed as a utility-based evaluation layer that works alongside an agent rather than replacing its intelligence. Start with a reviewable account of the context and consequences. Add client-directed governance through supported, tested integrations.
A separate evaluative function, with the authority and limits you assign.
Explore ADLThe ADL, in two halves
Understanding
Organizes the available decision context, supplied options, affected parties, and potential consequences. It distinguishes what the agent's records establish from additional evidence, assumptions, and Bentham's interpretation. Its Logic Receipt makes the assessment's basis and uncertainty available for review and correction.
Governance
Connects that account to client-authorized policy: which concerns to flag, which actions require review, and where a supported intervention is permitted. It is being developed to apply explicit rules through the host workflow, with records of the policy, its authorized owner, and the response actually confirmed.
A path to greater involvement
Capability descriptions — not activation controls
Examine the available decision context.
Preserve the assessment and its basis.
Surface a concern for attention.
Pause a specified action pending authorized review.
Prevent a specified action within the integrated workflow under an applicable policy.
Use a separately approved fallback or selection process within its defined authority.
Start with understanding and recorded findings. Add controls only where the integration, evidence, and authority support them.
The receipt is designed to distinguish the assessment, the policy applied, the authorized response, and the evidence of execution. A requested pause is not reported as a completed pause without confirmation.
Where this comes from
Bentham grows out of human-factors engineering and decades of client work in aviation, healthcare, and electrical power — investigating why capable people make consequential choices and what could make the next choice better.
David Marx built Boeing's maintenance human-factors team and founded The Just Culture Company.
A sustained research effort, including a dedicated PhD-level team during the COVID period, developed quantitative approaches to behavioral choice.
ECRI acquired The Just Culture Company. Bentham's proposed healthcare collaboration builds on that relationship; specific development and independent-review arrangements require agreement.
HBM for professional understanding and change design. ADL for interpreting AI decisions and introducing client-directed governance in defined workflows.
Historical practice informs the questions and the method. Each new product and deployment must earn the claims made for its own performance.
Start a conversation
A recurring human behavior. An AI workflow with consequential choices. A review process that needs a clearer account.
We are working with selected organizations and professionals to examine these problems and develop useful, testable ways to address them.