BCG Platinion, Energy Practice
Energy / Utilities · 2026
Agentic Billing Platform for Energy Suppliers
Contributed to an agentic billing platform for energy suppliers: agents for back-office workers, an MCP layer over the domain services, an agent control tower, and smart home energy integration.
Context
Starting point and environment
Energy suppliers run billing, invoice corrections, dunning, metering, and regulated market communication with grid operators. Much of the back-office work is repetitive casework that still depends on domain rules and human sign-off.
Challenge
Core constraint to solve
Let agents take on back-office casework without ever acting outside the rules: decisions about money and customers have to stay deterministic, auditable, and approved at the right level.
Solution
Design and implementation approach
My parts of the platform were debugging and hardening the agent layer, the agent swarm for back-office workers, the MCP server, the agentic Control Tower, and the smart home (HEMS) integration. The core principle: rules decide, the model proposes within guardrails. Agents never write directly, small actions run automatically, larger ones need approval, and every action is audited. Domain services are exposed through an MCP server that agents call with the user's own token, so role and tenant permissions apply to every tool call. The Control Tower shows costs, live runs, hand-offs, traces, and open approvals per agent. The smart home work adds a provider-agnostic device layer and a single command gateway that enforces consent, limits, idempotency, and audit, with deterministic charging and heating plans against quarter-hour prices.
Outcome
Measurable impact and delivery result
Hardened the agent layer for demos, with domain safeguards such as refusing corrections that would charge a customer more and checking bulk billing runs before they go out. The smart home integration was verified end to end against a real provider sandbox with a wallbox and an electric car, and is currently in review.
Key Takeaway
“In regulated domains the model shouldn't make the decision. Hard rules decide, agents propose, and humans approve above a threshold, which is what made autonomous agents acceptable for billing.”
Working on something similar? I take on select consulting and freelance work in agentic AI, RAG systems, and LLM reliability.