Investors deck · v0.3 · updated 2026-09-17, press ⌘P / Ctrl+P to save as PDF.
Agents need bounded authority.
His refund agent can read the case, check the order, and propose the right next step. Pete wants it to finish the work, not just draft it.
A credential lets Pete’s agent reach the refund tool. It cannot say whether this refund is authorized now.
Pete wants the agent to finish the job.
His client is right to stop him. A permission to refund is not an answer to whether this refund is valid now.
The client cannot grant a blank cheque.
One refund may be valid while the invoice is paid. The next may be invalid because the invoice was already refunded. Many individually valid refunds may exceed the agent’s daily limit.
So the client keeps a person on the consequential step. That may be the right decision until there is a better contract.
The first live connection forces the decision.
Pete can leave the agent in draft mode. But the day his client connects it to the refund tool, somebody must decide what that agent may do without asking again.
The need is created at deployment, not at audit time. An alert after a bad action cannot undo it. The authority pattern chosen for this first connection becomes the starting point for the next one.
Then Pete meets Anne.
She does not ask the model to judge its own limits. Her team defines the states, actions, authority, and ceiling that make delegation acceptable.
Anne shows Pete the missing instrument: a named, revocable mandate for one agent, checked before each action.
Anne gives him a way to say yes.
executes if allowed
The client defines the rule. KIFF applies it before execution: runtime authorization for the proposed action. A person still decides cases that require judgment. The client’s system, not KIFF, moves the money.
The last refund is refused for a new reason.
No one needs to notice the cumulative total after the fact. The boundary checks it before authorizing the next action.
Pete tests the boundary before switching it on.
If almost every case still needs a person, Pete learns that in week one. KIFF does not promise to remove a reviewer or a queue.
The client can explain every decision.
KIFF records what was proposed, which state and rule it checked, and what it decided. Signed, tamper-evident evidence appears in the Control Room and the downloadable proof pack.
The mandate statement counts authorizations. It does not claim every authorized refund was executed by the client’s system.
Pete can change the agent without rebuilding the rule.
This is the platform thesis: each new agent can use the same operational contract instead of negotiating authority inside another prompt or application.
Pete’s project opens the door. Cloud is the business.
The MIT framework and guard make adoption possible without a contract. A founder-led engagement puts one action on rails: a free boundary review, then a fixed €7,500 first week in observe mode, followed by scoped day-rate work if the customer continues.
Cloud is the recurring layer: a production platform fee and governed-operation consumption. Domains, agents, and actions are not separate toll booths. Annual commitments follow measured usage, not a forecast invented at the first meeting.
The product is shipped. The buying decision is next.
The key test is not whether Pete likes the story. It is whether an operation owner pays to bound real delegated authority.
Be like Pete. Make delegation explicit.
Pete did not replace the client, the reviewer, or the executor. He gave the client a way to authorize consequential work within a boundary they could name, revoke, and inspect.
KIFF is raising to prove that this shipped runtime can become a repeatable production business.
Next milestone: first production customers with observed usage, measured service cost, and a decision to expand.