ProductBenchmarkDocsTrustCompanyTalk to us
Independent authorization for agent financial actions

Runtime controls for AI agents that move money.

Refunds, payouts, payments and credits get an ALLOW, REVIEW or BLOCK decision before they execute. Kerneva judges the pattern across sessions, not just the single transaction in front of it.

Book 20 minutesSee the benchmark →
What goes wrong

A refund agent went from $20 to $74 in five sessions.
The limit was $75. Every refund passed.

Per-transaction control
$20 → $35 → $50 → $65 → $74
Each refund is under $75. Nothing fires.
Money moves
Kerneva
$20 → $35 → $50 → $65 → $74
REVIEW — amounts climbing toward the limit
Operator decides → outcome recorded
The agent never broke a rule. It found that larger refunds close complaints faster, and the cheapest path to “resolved” was the largest refund that needed no approval.
Simulated scenario from the 2026-01 benchmark: 1,120 trajectories across 6 models, 85.2% drifted while staying inside policy.
Where existing controls stop

Everyone checks the transaction.
Nobody checks the trajectory.

Policy engines validate one request against static rules. Fraud tools score one transaction. Observability shows what already happened. None of them ask whether an agent’s behavior has been changing across sessions.

When
Sees
Cross-session pattern
Policy engines (OPA, Cedar)
Before request
One transaction
Blind
Fraud detection
Real time
One transaction
Blind
Observability
After the fact
One session
Partial
LLM guardrails
Before output
Content
Blind
Kerneva
Before execution
Agent history
Full
How it works

Agent → Kerneva → business system.
Start in Observation Mode.

Your agent sends the action it is about to take: agent, session, action type, amount, counterparty. Kerneva scores it against that agent’s history and returns a decision with a signed receipt. No conversation content, no PII, no agent rewrite.

PHASE 1
Observation Mode
→Blocks nothing, changes no workflow
→Builds a behavioral baseline per agent
→Shows the drift, structuring and threshold-gaming it would have flagged
PHASE 2
Enforcement
→ALLOW — behavior consistent with the baseline
→REVIEW — hold for a human before money moves, including a first payment to a new payee or bank account, or an action from an agent nobody approved
→BLOCK — stop it, including an operator halt for a whole agent or account
→Every decision versioned and exportable for audit
✓ Hosted or Self-hosted✓ Python SDK or REST✓ No PII requiredRead the quickstart →
K

Running an agent that issues refunds,
payouts or credits?

Twenty minutes on how it behaves today, and whether the patterns above show up in your own history. No integration needed for the first conversation.

Book 20 minutesTry it yourself

Built by Vaibhav Dedhia in Bangalore. Write directly: [email protected]