Making AI-generated renewable energy offers commercially trustworthy.

Context

A European power-trading and renewable-origination business prepares proposals for renewable assets, PPAs and direct-marketing opportunities. An internal AI tool — using large language models, retrieval and internal data — could already produce a fluent, well-structured proposal.

Fluency is not correctness: an AI proposal can read perfectly and still be commercially wrong, carrying mistakes in generation profiles, forward curves, balancing and imbalance exposure, regulatory treatment or PPA valuation.

The problem

In origination, a plausible-but-wrong number is not a typo — it is a mispriced position. A mispriced offer can leave value on the table, lose deals that could have been won, or book risk nobody intended at the point of submission. The question was never whether the AI could write an offer. It was whether the system could help produce one an experienced originator would stake their name on.

Our approach

Blackdown worked in an AI consulting role, helping the firm move its internal AI tool towards a commercial decision-support workflow its desk could trust. We focused on the layer beneath the interface — pricing accuracy, evaluation, governance and human review — so that every figure the system generates or explains traces back to approved data, is checked against existing tools, and is reconciled with the pricing logic the firm's analysts already use.

In that workflow, drafting stays separate from pricing: the system can help draft, but it is never the sole source of a number. Proposals that reconcile within tolerance go forward; the ones that need a closer look are routed to the right analyst, with originator review as the final backstop — never fluency as a proxy for correctness.

A real evaluation baseline

We validated the system against a pilot set of 24 historical origination opportunities — previously priced deals, reconciled against the pricing logic analysts already use. We checked whether it used the right inputs, applied assumptions consistently, produced explainable pricing deltas and landed where an experienced originator would expect. Of the proposals that needed further scrutiny, the system correctly flagged 91% for analyst review before they reached an originator — so the desk knew what the system caught, and what it didn't.

Results

Against a pilot set of 24 historical deals, 78% of AI-generated proposals reconciled to analyst pricing within ±3% with no intervention; the system flagged 91% of the proposals needing scrutiny before they reached an originator, and first-draft proposal time fell by 38% — from 16 to roughly 10 hours per proposal.

Alongside the evaluation baseline, we defined the control points needed for production: traceable assumptions, approved data sources, regulatory grounding, human review of commercial logic, and a data-governance review of the third-party research tooling. This gave the business a practical basis for deciding when an AI-generated proposal is ready for originator review and when it needs further analyst input.