Irreversible actions wait for a person
Can a bad prompt trigger a financial transaction you can't take back?
Not if the system is built correctly. Kyver classifies every action an AI system can take by cost and reversibility. Reading and drafting run freely. Anything that moves money, deletes records or contacts customers is held for approval by a named person, who sees the request, the reasoning and the data in one place.
Issue refund $2,340.00
Order SO-88213, damaged in transit. Customer sent photos of 3 of 12 units. The agent matched the order and the carrier claim and drafted the reply.
Policy checks
- 14:31:07PROPOSEPROPOSE refund for order SO-88213routed, by service-agentroutedby service-agent
- 14:31:07CHECKCHECK refunds.v4, limit $500held, over limitheldover limit
- 14:31:07CHECKCHECK proposer can't approvepassed, agent excludedpassedagent excluded
- 14:31:08ROUTEROUTE to service managerrouted, named approverroutednamed approver
How Kyver handles AI governance
Inventory every action
We list each action every AI system can take and rate it by financial impact and how hard it is to reverse.
Approval gates you define
You set the thresholds: dollar amounts, record types, recipients. Above them, the action is held for sign-off by the right person.
Separate the duties
The agent that proposes a payment can't approve it. The rules you apply to staff apply to software.
Keep policy in version control
Rules live in a repository with full change history, so an auditor can see exactly what the policy was on any date.
Training is a tendency, not a control
Most AI pilots start with the model and add controls later, if ever. The result is an assistant that can send the wire, issue the refund or email the whole customer list because a request sounded reasonable.
Model providers improve their safety training with every release, but training only makes good behavior more likely. Your controls have to hold on the day the model is wrong.
Built in alongside governance
- SecurityCatch AI attempts to reach data or take actions outside their permissions, and block them before they run.
- CostRun most AI work on hosted models that aren't metered by the token, so the monthly cost is predictable.
- TraceabilityRecord every AI session step by step, so failures can be investigated and fixed instead of guessed at.
- MaintainabilityAn outside team with no internal politics, that stands behind what it builds and keeps improving it.
Questions about AI governance
What is human-in-the-loop AI?
A design where an AI system prepares an action and a person approves it before it takes effect. It is used for actions that are costly, regulated or hard to reverse.
Won't approvals slow everything down?
Only for the actions that warrant them. Most AI work (research, drafting, summarizing, matching) is reversible and runs without a gate. Approval requests arrive with everything the approver needs, so most take seconds.
How does AI governance relate to compliance frameworks?
Governance controls produce evidence for the frameworks you already answer to, such as SOC 2 change management or financial controls testing. During the assessment we map each control to the requirements you name.
Start with the assessment
Fixed price, agreed up front. Implement the plan with us within 6 months and we refund the fee.