Skip to content

Put AI agents to work. Keep control as they scale.

KLA applies your controls as agents work, sends held actions to their owner, and samples completed work for independent quality review.

A 30 minute working session to scope one workflow.

  • A production decision in four weeks
  • Evidence an auditor verifies offline
  • Your infrastructure, your models

Programs and memberships

NVIDIA Inception ProgramAssociation Monégasque des Compliance OfficersMonacoTech Startup Program

Standards participation

Member: AFNOR CN IA

What you have after four weeks.

  1. Week 1Scope
  2. Week 2Configure
  3. Week 3Governed traffic
  4. Week 4Production decision

A production decision

The SAFR Production Sprint takes one agent through scope, configuration, and governed traffic. Week four ends with a verified evidence bundle and a documented production decision.

See sprint scope and pricing

Review with named owners

Held actions reach their owner in Decision Desk with the rule that applied. Completed work is sampled for independent quality review, and each verdict records the reviewer and the reason.

Proof for your auditor

Evidence Room exports the period as a Sealed Evidence Bundle. Your auditor checks it offline with a standalone verifier, without access to KLA.

Figures shown are example data.

Agents run the work.
Your people make the decisions.

KLA evaluates actions on connected routes against your policy before the tool runs. Actions that need approval go to the person responsible. Completed work is sampled for independent quality review.

CommandToday · 14:32 · Example data
Actions checked
12,480
Allowed
11,902
Warned
131
Sent for approval
412
Blocked
35
AgentWorkActionsTo reviewQuality
Treasury payments agentSupplier payments3,21018898% correct
Onboarding agentBusiness onboarding2,9409685% correct
Claims triage agentMotor claims4,10510296% correct
Client service agentAccount requests2,2252699% correct

Decision Desk6 waiting

Compliance officer L2

Pay invoice INV-2048

€420,000

R3 · Payment above €100,000 · One approval

  • Quality reviewerReview onboarding report KYB-1182Sampled · Release 4 · New version
  • Compliance officer L1Pay Ostrava FreightR1 · Payee added in the last 30 days
  • L2, then MLROFund the Arden share purchaseR5 · Payment above €1,000,000 · Two approvals

Routine actions run on your policy.Exceptions wait for the person who owns them.Sampled work gets independent quality review.

Decision Desk

Each held action reaches the person who owns it.

The reviewer sees the request, the rule that applied, and the context on one screen. Approval releases the action and rejection stops it. Blocked actions never reach the tool.

Decision Desk
Invoice INV-2048

€420,000Supplier payment

Requested by
Treasury payments agent
Operation
Submit payment
Supporting document
Supplier invoice

Approval threshold€250,000

Treasury reviewer

Required approval role

Review payment

The amount triggers the payment-approval rule. The payment stays on hold while treasury reviews the request.

Treasury review required

Policy outcome: require approval

ExecutionPayment submission is heldWaiting for the reviewer’s decision.
Approve or reject the payment in the example.Explore human approval

Rules that check more than a limit.

Rules compare records, judge what a document or name means, and set what an agent may do on its own. Every rule ends in one of four outcomes.

  • Compare sources

    Whenan agent releases a payment

    Ifthe beneficiary on the instruction differs from the approved account record

    Require approvalPayments operations

  • Judge meaning

    Whenan agent opens a customer account

    Ifthe legal name reads like a placeholder, such as “Test Company Ltd”

    BlockAccount not opened

  • Watch volume

    Whenan agent exports customer records

    Ifthe export holds more than 1,000 records

    WarnExport runs, flag recorded

  • Keep routine work moving

    Whenan agent issues a refund

    Ifthe amount is under €500 and the order shows as delivered

    AllowRefund issued

Continuous assurance

A case can pass every check and still be wrong.

KLA samples completed work for quality review by someone who did not approve it, starting with every decision of a new agent version. When incorrect verdicts push quality below your target, a Finding opens, an owner tracks the fix, and later samples show whether it worked.

Assurance Center
Finding · Business onboarding

79%correct decisions, target 85%

Agent version
Release 4
Sampled this week
14 reports
Marked incorrect
3 reports
Owner
Head of onboarding

Policy checks on these reports14 of 14 passed

Quality reviewer

Independent of the approver

Three reports passed every check and were wrong.

  1. Day 1

    Release 4 promoted to production

    Sampling starts at every decision for the new version.

    Onboarding agent
  2. Day 8

    3 of 14 sampled reports marked incorrect

    Reason recorded: the report omits an address mismatch.

    Quality reviewer
  3. Day 8

    Finding opened

    Correct decisions at 79%, target 85%

    Assurance Center
  4. Day 10

    Remediation Plan approved

    Prompt update, shipped in Release 5

    Head of onboarding
  5. Day 24

    Fix verified

    27 of 28 sampled reports correct over 14 days

    Finding resolved
EvidenceVerdicts, the Finding, the Remediation Plan, and the verification are recorded.Example data
A change to the review rate needs approval from a second person.See how sampling adapts

Govern an existing agent in three steps.

  1. const result = await kla.evaluateToolCall({
      runId: 'inv-2048',
      gateId: 'supplier-payment',
      toolName: 'submit_payment',
      toolArgs: { amount: 420000, currency: 'EUR' },
    });
    
    result.decision; // 'require_approval'

    Step 1

    Connect the agent

    Add the KLA Node.js or Python SDK. The agent asks KLA for a decision before each consequential tool call.

    Review integration routes
  2. Whenan agent submits a payment

    Ifthe amount exceeds €250,000

    Thenrequire approval from Treasury reviewer

    Step 2

    Write the rule

    Your team sets limits, permitted destinations, and approvers in Policy Builder, in plain sentences.

    See policy controls
    • AllowThe tool runs.
    • WarnThe tool runs and the warning is recorded.
    • Require approvalThe action waits for its reviewer.
    • BlockThe tool is not called.

    Step 3

    Act on the decision

    Every action gets one of four outcomes. Each outcome, approval, and tool result is recorded as evidence.

    See human oversight

Independent verification

Proof an auditor can check without access to KLA.

Each action keeps its request, the rule that applied, the reviewer’s decision, the tool result, and any quality verdict. Lineage Explorer shows them in order.

Evidence Room exports them as a Sealed Evidence Bundle. A standalone verifier checks the bundle with no network access.

See what the verifier checks
Lineage Explorer
Invoice INV-2048

Payment execution

Submission accepted

  1. Payment requested

    Submit supplier payment · €420,000

    Treasury payments agent
  2. Approval required

    Amount exceeds the €250,000 threshold

    Payment limits · version 3
  3. Treasury approved

    Invoice and approval scope confirmed

    Treasury reviewer
  4. Tool accepted submission

    Submission accepted; settlement not shown

    Payment tool

Common questions

Ask about your own workflow in the first session.

Book a demo
What does KLA do when an agent tries to act?
KLA evaluates the action against your policy before the tool runs. The result is allow, warn, require approval, or block. Actions that need approval wait in the Decision Desk for the person responsible.
What happens while an action waits for approval?
The tool does not run. The reviewer sees the request, the rule that applied, and the supporting context on one screen. Approval releases the action and rejection stops it. Both outcomes are recorded.
Which agents can KLA govern?
Agents you run through KLA Runtime, and agents in your existing stack. Teams add the KLA Node.js or Python SDK, and KLA checks each consequential tool call before it runs. OpenTelemetry adds a record of what the agent did.
What if an approved action still produces poor work?
KLA samples completed work for quality review by someone who did not approve it, and the reviewer records a verdict and a reason. When incorrect verdicts push quality below your target, a Finding opens, an owner tracks the fix in a Remediation Plan, and later samples show whether it worked.
Who writes the policies?
Your team defines the allowed operations, destinations, and approval requirements in Policy Builder.
What evidence does KLA keep?
Each action, the rule that applied, the reviewer’s decision, the tool result, and any quality verdict. The record shows what happened and who decided. Quality review shows whether the work was right. Evidence Room exports a Sealed Evidence Bundle when an auditor or supervisor asks.
Can an auditor check a bundle without access to KLA?
Yes. The export ships with a standalone verifier that checks the bundle with no network access. To confirm the bundle came from KLA, the auditor pins its keys against a key fingerprint KLA publishes.
How long does it take to reach production?
The SAFR Production Sprint runs four weeks on one workflow. It ends with a verified evidence bundle and a documented production decision.
What happens after the four weeks?
A positive production decision defines your subscription. Your controls, quality review, and evidence record keep running as releases and models change, and the next workflow runs on the same platform.
Where does KLA run, and with which models?
KLA runs as a managed service or on your own Kubernetes cluster. Your agents use the model provider you choose: OpenAI, Azure OpenAI, Anthropic, or a model you host behind an OpenAI-compatible endpoint.
What does the first session cover?
Thirty minutes on one workflow: the agent, the systems it touches, the approval owner, and the evidence you need.

Bring one workflow.
Scope the four weeks.

We review the agent, the systems it touches, the approval owner, and the evidence your risk team needs. You leave with the integration route and the scope of the sprint.

Book a demo

A 30 minute working session to scope one workflow.

KLA: Put AI agents to work and keep control as they scale