About Cadinal
Agents can act now. Cadinal designs how humans stay in control.
Cadinal is a company building the interface between AI agents and the people who answer for them. As agents start doing real work inside companies - buying a report, renewing a service, requesting data, moving money - someone still has to decide what they may do, see what they did, and step in before it matters. That shared surface, where an agent acts and a human stays accountable, is what we design.
What we build
We think a new product category is forming: not another chatbot, and not fully autonomous software, but a control surface between the two. Agents move fast and in small steps; humans set the goals, the limits, and the judgment. Cadinal builds the place where those two meet.
Every Cadinal product follows the same spine: give the agent bounded authority, give the human a legible view and a way to intervene, and keep a record both sides can trust.
Agentic AI
Agents are good at breaking work into smaller units. That changes behavior. A team may not need a seat, a contract, or an annual subscription. It may need one data packet, one transaction, one vendor action.
But smaller actions do not remove accountability. They increase the number of moments where intent, permission, evidence, and human approval have to be clear.
Why we started with payments
Payments are the highest-stakes place an agent can act: money leaves the company, finance has to reconcile it, and a wrong move is expensive to unwind. So that is where we started.
CadinalPay gives an agent bounded payment authority - what it can buy, from whom, through which route, with what evidence, and when a human must step in. It checks policy before money moves, captures receipts and tax metadata, blocks risky terms like silent subscriptions, and leaves a usable audit trail.
What comes next
CadinalPay is the first Cadinal product, not the last. We are preparing more interfaces for the other moments where an agent acts on a company's behalf and a person needs to stay in the loop.
The pattern generalizes beyond money: anywhere delegation needs scope, visibility, and a clean handoff back to a human, the same control surface applies.
What we are careful about
We do not think agents should act automatically by default. We do not think any single payment rail or model fixes the hard parts. These are tools, not the thesis.
The thesis is delegated trust: agents can act like trusted operators only when their authority is scoped, their actions are inspectable, and their evidence is attached to what they did.
Cadinal