Decision Desk

The Decision Desk (/approvals-inbox) is where people make the calls that machines should not make alone. When the KLA policy engine returns a require_approval decision, the agent's action does not proceed. Execution pauses and KLA opens an Escalation, a manual-intervention object carrying everything a reviewer needs to act. The work item the reviewer ultimately resolves is a Decision Request. The Decision Desk is the queue, the review surface, and the audit record for every one of these moments.

This matters because require_approval is one of KLA's four policy outcomes, in precedence order: allow, warn, require_approval, block. The first two let work continue; block stops it outright. Only require_approval hands the decision to a human, and the Decision Desk is where that human stands.

ℹ️ Note

The Decision Desk does not create the rules; those live in **Policy Builder**. It is the runtime destination for actions a policy deliberately routed to people instead of resolving automatically.

How an Escalation reaches the desk

flowchart LR
  A["Agent action"] --> B{"Policy decision"}
  B -->|allow / warn| C["Continue"]
  B -->|block| D["Stop"]
  B -->|require_approval| E["Open Escalation"]
  E --> F["Decision Desk queue"]
  F --> G{"Reviewer acts"}
  G -->|approve| H["Resume execution"]
  G -->|deny| I["Terminate run"]
  G -->|re-route| J["Senior reviewer"]
  J --> G

The agent's run is held, not failed, while the Escalation waits. The instrumented agent (via the Govern in Place SDK) or the managed proxy (Run through KLA) blocks on the pending decision, so no side effect occurs until a person resolves it.

Working the queue

The inbox is built for triage at volume. Reviewers filter and sort across four dimensions:

Dimension	What it does
Priority	Surfaces critical Escalations ahead of routine `warn`-adjacent reviews.
Team	Routes items to the right queue (claims triage, refund approval, document review) so finance never sees legal's work.
Agent	Isolates every pending action from a single agent, useful when one agent is behaving unexpectedly.
State	Separates open Decision Requests from resolved history for audit and trend review.

Routing rules are declared in policy, so an Escalation lands in front of the team that owns that risk by default rather than in a shared pile.

Reviewing a Decision Request

Opening an item gives reviewers the full context behind the paused action, with no switching tools and no guessing:

Prompt and instructions: the input that drove the action, including genai.system.instructions, so the reviewer sees what the agent was told to do.
Tool payload: the exact genai.tool.name and genai.tool.parameters the agent intended to invoke (the refund amount, the record to write, the document to send).
Triggering rule: the specific policy that returned require_approval, with its reason codes, so the decision is grounded in a named control, not a hunch.
Lineage link: a jump straight into Lineage Explorer to replay the full Lineage Record leading up to this step.

From here the reviewer takes one of three actions:

Approve: releases the hold; the agent resumes exactly where it paused.
Deny: terminates the run; the action never executes.
Re-route / escalate: hands the Decision Request to a more senior reviewer or another team without losing context.

SLA and the audit angle

Every Decision Request carries timing. Reviewers see how long an item has waited, and overdue Escalations surface against their service-level target so nothing rots in the queue. That timing is part of the record, not just an operational nicety.

🛡️ Important

Every decision is evidence. Each approve, deny, and re-route is captured as an OpenTelemetry span, redacted by the KLA Collector, and written to the ImmuDB cryptographic ledger. The reviewer, the action, the triggering rule, and the timestamp all flow into **Sealed Evidence Bundles** and **Control Packs**.

The result is a defensible answer to the auditor's hardest question, who approved this, on what basis, and when, without anyone reconstructing it after the fact. Human judgment becomes a first-class, tamper-evident part of the governance record.

💡 Tip

Pair the Decision Desk with **Assurance Center**. A rising rate of `require_approval` Escalations on one agent is often an early **Assurance Alert**, a signal to revisit the policy or the agent before reviewers drown.