A governed agentic architecture for PPA, GoO and route-to-market operations.

Context

A German power trading firm runs PPA, guarantee-of-origin and route-to-market operations. A typical request was kept intentionally simple — “Place this 50 GWh corporate PPA and source matching Guarantees of Origin” — but that instruction starts a series of regulated and financially sensitive steps: contract interpretation, certificate sourcing, GoO matching, registry redemption, commercial pricing, nominations, balancing, settlement and audit checks.

Before orchestration, the process depended mostly on email. Commercial, legal, operations, settlement and registry specialists each held their own piece of the information, and triage alone took around four hours per complex request.

The problem

The obvious answer — let a model handle requests end to end — was unacceptable. These workflows touch registries, nominations and settlement. A misrouted certificate transfer or an unapproved schedule change is not a customer-service error; it is a financial and regulatory event.

The firm needed the speed of automation with the accountability of a named person behind every action that binds the company.

What we built

Blackdown designed, built and delivered the production agentic workflow — the orchestration model, tool boundaries, approval controls and audit processes used in live operations. The orchestration layer acted as planner and router: it interpreted each request, decomposed it into auditable sub-tasks and assigned each one to the right specialist agent — contract intelligence, GoO portfolio matching, deal and pricing support, or route-to-market operations.

Execution was never the model's job. Deterministic tools — registry actions, nomination submissions, settlement calculations — ran through set interfaces, validation rules, access controls and audit logs, and any binding step ran only after a named person approved it. The agent proposed; the tool executed; the human decided.

The control model

The system was a copilot, not a fully autonomous trading engine. It could suggest and prepare actions, but nothing moved money, transferred certificates, submitted to registries or markets, or took market exposure without approval from a named person. Every step was logged with its source data, outputs and approver, so any process could be reconstructed after the fact — for the desk, for audit, for the regulator.

This was a strict rule, not a preference: the project carried significant legal and governance requirements, so there was always at least one human check of the AI's output before any binding action. The goal was not to remove human responsibility, but to make the review step clear, informed and easy to audit.

Results

Measured in production during the final two-week handover, when the system managed hundreds of live processes: time from request triage to a review-ready action plan fell from around four hours to around twenty minutes — a ~92% reduction — and 100% of binding actions were routed through named human approval, a mandatory legal and governance control.