Human decisions. AI decisions. Consequences that matter.
Two tools for consequential decisions. HBM predicts human choices and helps professionals design behavior change. ADL independently examines the consequences of AI actions.
Start with a saved example. No sign-in required.
A medication error has occurred. What happens next?
100 synthetic profiles. A saved, conditional calculation — not a measured population rate.
Scenario-specific estimates. No action is selected or authorized by this analysis.
Two products. Distinct jobs.
HBM models the human chooser. ADL evaluates the AI agent’s options. Both make the basis of an analysis visible.
Understand what people notice, believe, value and anticipate. Model likely choices across a defined group, then compare how changes in circumstances could change behavior.
Examine an agent’s proposed actions and the consequences for each affected party. Inspect the evidence, estimates and unknowns in a reviewable decision record — a Logic Receipt.
A closer look · Narrated introduction
One clinical event. Two separately authored examples. See how HBM models the human chooser and ADL examines the consequences of an AI action.
User-initiated playback with audio, captions and a transcript. HBM 1.8.7: saved run, 5 October 2026. ADL 1.6.0: authored forecast, 5 October 2026.
The narrated introduction uses edited passages from the supplied recording. The updated examples are shown on screen and in the linked reports below.
I'm David Marx, founder of Bentham AI. We're bringing together decades of human-factors practice and years of research into human decision-making. Inspired by English philosopher Jeremy Bentham's work on social utility, we're combining our expertise with the power of today's frontier AI models to build two groundbreaking tools: the Human Behavior Model and the Autonomous Decision Layer.
The Human Behavior Model helps professionals explore what people in defined groups are likely to do under particular conditions, and why. What do they notice? What do they expect to gain, avoid, or protect? The point isn't just prediction. It's an explanation you can investigate and use to develop a change worth testing.
The Autonomous Decision Layer brings that concern for human consequences to the AI agents working on our behalf. Think of the ADL as a second, utility-based evaluator alongside your AI agents. The ADL Interpreter examines the available context, the options, and the gains and burdens for those affected. Its Logic Receipt makes evidence, assumptions, and uncertainty visible, so you can question the assessment rather than simply accept an answer.
We're also developing the ADL Governor: connecting that understanding to decisions your organization can support. Start with visibility. Add flags for review. Then, through tested integrations, hold or prevent specified actions where appropriate. You decide the scope of authority. We build the evaluation and controls to support it. Understanding first. Then governance at your direction, one decision at a time.
Watch the HBM and ADL demos here on our site. Then use your invitation to try the analysis tools for yourself. Bring a fictional or non-sensitive case. Examine the findings. Challenge the assumptions. Tell us what's useful, what's missing, and where you'd like to take the conversation. Bentham AI.
Shared setup: A known injection error has occurred. Urgent care is identical under both choices; the final clinical outcome is unknown. Compare accurate disclosure and documentation with an unsupported explanation and omission of the known cause. The human and robotic versions are separately authored cases.
HBM: 62 of 100 synthetic profiles select disclosure; 38 select the alternative explanation and omission; 0 exact ties. The result is conditional on the report's beliefs and assumptions, not a measured population rate. The human report retains unestimated patient clinical remainders.
ADL: Expected Net Utility forecasts for A / B: patient +3.9e-5 / -9.0e-5; dental practice -1.0e-4 / -3.3e-5; treating clinical team +3.7e-6 / -6.7e-6; future service users +1.3e-5 / +1.2e-7. These are independent accounts, not a cross-target total. All 64 expected components are calculated within the declared scope. No Safety Check threshold crossing is identified; that is not clearance.
Development status: Customer-authorized production controls are being engineered. The demonstrated reports do not select or execute actions.
Explore the work
The endodontist case
An injection error causes a serious patient reaction. Urgent care is underway. The decision is what to explain and record.
Option A
Explain the known error, acknowledge uncertainty and make an accurate clinical record.
Option B
Give an unsupported explanation and prognosis, omitting the known causal details while recording the reaction and care.
HBM · Human-clinician example
A modeled choice across 100 synthetic clinician profiles.
HBM 1.8.7 · 5 October 2026
Conditional calculation, not a measured population rate. The saved run assumes 1% hidden-error exposure and below 1% lawsuit risk conditional on revelation.
ADL · Robotic example
Separate forecasts for the patient, practice and other affected parties.
A / Disclose accurately. Accurate information and informed participation coexist with apprehension, care costs and modeled follow-up risk.
B / Alternative explanation. Brief reassurance coexists with misinformation, impaired follow-up choice and greater modeled follow-up burden.
ADL 1.6.0 · 5 October 2026
All 64 expected components calculated within the stated scope. Scenario-specific forecasts, not observed outcomes. No action selected or authorized.
These are separately authored synthetic analyses, not observed outcomes, clinical advice or the selected first pilot workflow. The original injury and urgent care are common to both supplied choices. The previews above are excerpts, not complete reports.
Request analysis access for you or your team. Start with a fictional or approved non-sensitive case.
Saved examples are open. Running a new analysis requires access to the analysis workspace.
Inside ADL
The action, the evidence and the consequences need their own examination. ADL separates interpretation, calculation and customer-authorized control.
Map the available evidence to each affected party across 28 defined consequence factors. Keep assumptions and missing information in view.
Reports demonstratedRecord values, probabilities and assumptions so fixed-input calculations can be checked. Keep each affected party’s account separate.
Recorded analysisEngineer flags, holds and refusals for authorized actions in supported workflows. The agent retains its task; the customer defines the boundary.
In developmentThe Logic Receipt
See what was known, what was estimated, who may be affected and what remains unresolved.
Analytical reports are demonstrated. Production integrations and customer-authorized controls are being engineered.
What’s next
HBM moves to release. ADL moves through engineering and workflow evaluation.
HBM / Client Alpha
Beginning October 2026Existing clients begin using the Alpha version as HBM moves toward full release.
Ask about HBM AlphaHBM / Full release
April 2027Our six-month release plan brings HBM to trained professionals: frame the choice, model experience, compare changes and measure behavior.
Discuss HBM and practitioner trainingADL / Healthcare pilot program
Bentham AI will select six healthcare pilot partners. With each organization, we will define the decision workflow, operating boundaries and measures of success.
Replay and shadow decisions; examine evidence and consequences.
Test available context and supervised agent interaction.
Test customer-authorized flags, holds or refusals in supported workflows.
Assess decision quality, workload, implementation and adoption.
Engineering and workflow testing precede activation. Each phase has a separate customer release decision.
Discuss a spring 2027 pilot
The founder
David Marx brings systems engineering, human factors and law to consequential decisions. His work includes mishap-investigation methods for Boeing and NASA, prospective risk modeling and the development of Just Culture.
At Bentham AI, that foundation is being applied to understanding human choices and independently examining AI decisions.
Start a conversation
Show us the human behavior you want to change or the AI decision you want independently examined.