now open Business partner program: evaluate HCA free on one action or application. Apply
control_origin: HUMAN_CONTROL_LIKELY

Know who controlled the action.

Human Control Assertion (HCA) is a new evidence technology for the agent era. It tells any system whether a person or an AI agent controlled the interaction behind a consequential action, and signs that evidence so others can verify it.

why now

Agents now act inside human sessions.

An employee may be signed in while an agent operates the application. Identity records show the legitimate user, and the approval log shows a click. Neither says whether a person or an agent produced it.

79%

of US executives report AI agent adoption in their company.

PwC AI Agent Survey, 2025
33%

of enterprise software will include agentic AI by 2028, up from under 1% in 2024.

Gartner forecast, 2025
15%

of day-to-day work decisions will be made autonomously through agentic AI by 2028.

Gartner forecast, 2025
25%

of enterprise breaches will be traced back to AI agent abuse by 2028.

Gartner forecast, 2024

Adoption signals and forecasts, not measurements of the problem HCA addresses.

business outcomes

Give agents more work. Keep proof of who decided.

HCA plugs into the approval and security tools you already use. What changes is what you can prove, and what an agent can no longer do in someone's name.

Automate more, safely

Let agents prepare payments, deployments or orders while the decisions that matter stay human, and provably so.

scaleagent workloads

Audit-ready approvals

Answer "who approved this, a person or an agent?" for auditors, customers and regulators, action by action.

proofper action

No agent signs for a human

An agent can't pass itself off as the approver, and unexpected agents in an employee's session reach your security team.

less riskof unauthorized actions

Fast to adopt

An open format and SDK that fit your existing identity, approval and security stack. Nothing to rip out.

openno lock-in
try it

Approve this payment. Then let an agent try.

A toy version of the classifier runs in your browser. Move your pointer, click Approve and see the assessment. Nothing you do here is sent anywhere. For the real product, open the live demo ↗

finance app / payments / review
Supplier payment pay_7841prepared by agent
beneficiary
Northwind GmbH
amount
€12,500.00
account
DE89 ···· 3000
prepared by
finance agent
policy
human approval
expires
10 min

Your pointer trail is drawn here. It never leaves this page.

hca · control origin

Toy model for illustration. It says nothing about HCA's measured performance.

how it works

Four steps from the click to the side effect.

HCA is one evidence layer. Identity, authority and execution stay with the systems that already own them. Hover a step to play it.

01

Collect

A versioned collector reads interaction timing and input provenance. Features are extracted locally, with no typed content.

02

Link

The event linker ties the observed confirmation to the exact frozen action and a fresh server challenge.

03

Classify and sign

HCA assesses control origin, states evidence quality and limits, and signs the result. It abstains when evidence is inadequate.

04

Assemble and execute

Your systems add identity, authority and the decision. The executor checks the full certificate right before the side effect.

See the architecture and the assertion format →

use case one

See where the check sits in an approval.

An agent prepares a €12,500 supplier payment and policy reserves the approval for a person. Turn HCA on or off, choose who clicks Approve, and run the workflow.

IT operations, procurement, healthcare and agent control awareness →

classifier research

Help us train an honest classifier.

HCA's accuracy is an empirical question, so we measure it before we claim it. The Human Control console collects consented, labeled interaction sessions, from people and from agents, to train and evaluate the classifier.

  • Consented and labeled. Every session records who or what actually produced the interaction.
  • Timing, not content. No passwords, message bodies or screen recordings.
  • Kept apart from production. Research capture has its own notice, permissions and data flow.
  • Results published with limits. Datasets, coverage, thresholds and failed cases, not a headline accuracy.
business partner program

Evaluate HCA on one action, free.

We're looking for organizations where agents touch consequential actions, and for platforms that own a confirmation step or a workstation agent. Together we measure HCA against your current controls, in shadow mode.

priceFree during evaluation
modeShadow, nothing blocked
scopeOne action or application
you getResults vs current controls