From dashboards to decisions: the closed-loop KPI agent playbook

A dashboard tells you the KPI is red. That is the easy part. The hard part is everything between "red" and "back in range" - the cross-module variance attribution, the supplier-by-supplier collections sequence, the markdown vs supplier-return scenario, the segregation-of-duties checks, and the audit trail that survives the next quarter close.

DivetIQ is built so that loop closes inside the platform. Specialist agents - one per KPI - run the playbook below against the same headless APIs your finance, procurement and AR teams call. The numbers in the walkthrough come from the live working-capital workflow that DivetIQ ships as a default playbook.

Detect

The Cash Position agent evaluates a 13-week liquidity forecast against the customer's policy floor on a continuous basis, drawing from the AI-BI/DW semantic layer rather than nightly extracts. When week-6 cash projects €1.8M below the floor, the orchestrator wakes the specialists wired to that KPI's recovery playbook.

No batch job. No "we'll see it on Monday." The threshold is the trigger.

Attribute

DSO and Collections agents trace the variance to its real cause: 73% comes from three overdue receivables totaling €902K, aged 28–42 days. The agents draft a personalised collections sequence per customer through the same OpenAPI endpoints the AR clerk uses - the customer object, the dunning template, the payment-portal link.

The output is not a chart. It is three queued actions, each with its own confidence score and the row-level lineage that produced it.

Plan and optimize

Two specialists work in parallel.

The Working Capital agent queues three suppliers for a 10–15 day payment-term deferral. Before proposing, it checks the contract LCM record, the supplier SLA, and the current ESG and risk scores. Only suppliers with no early-pay discount and headroom in their terms qualify. €700K freed.

The Inventory agent scores six SKUs with over 180-day turn against markdown vs supplier-return scenarios. The recommended mix yields the highest cash return while keeping safety stock for fast-movers intact. €2.1M working capital recoverable.

Neither agent executes anything yet.

Approve

The three plans - collections, AP deferral, inventory liquidation - bundle into one CFO approval ticket. SoD checks pre-run. Full lineage attached: which event triggered which specialist, which API each one called, what RBAC/ABAC scope each call needed. A one-click modify path is available before approve.

€4.6M combined plan. One approval. The autonomy threshold is configurable per agent and per role - some companies run collections fully autonomously up to a euro limit and require approval only above it. The point is the threshold is yours to set, and every action below the line is still logged.

Execute and learn

On approve, the agents fire through the same APIs your humans use - the collections sequences post, the AP runs reschedule, the markdowns appear in WMS. Every call is in the audit log with the agent that made it, the reasoning trace, the inputs, and the result.

Eleven days later the KPI returns to range. The orchestrator records which moves actually moved the needle - which supplier deferral was accepted, which markdown sold, which collections email got the wire transfer - and updates the playbook for the next breach. The knowledge graph keeps the lineage so the next quarter's audit can reproduce every decision.

This is the difference between dashboards that report problems and a platform that resolves them. Every agent action stays inside the same governance perimeter as human action - RBAC/ABAC, segregation-of-duties, EU AI Act–ready transparency notices, per-agent kill switch. The agent is not an exception to your controls. It is a user of them.