The Quote Nobody Had Time to Win

The Company

Halden & Vor formulates and applies specialty coatings for marine and offshore equipment — the stuff that keeps a hull from corroding in the North Sea. A hundred and thirty people, deeply technical, the kind of firm that wins because its engineers genuinely understand the chemistry. Sales is treated as an interruption to the real work, which is the problem in this story.

The Pain

Tenders arrive as dense technical RFQs — substrate specs, environmental conditions, compliance standards, twenty pages of it. To quote accurately, someone has to read the whole thing, match it against the product catalogue and past projects, calculate coverage and labour, and assemble a compliant proposal. Only two senior engineers can do this well, and they're booked solid on delivery. So tenders sit. The firm quotes maybe half of what it receives, and the unanswered six-figure tenders are pure lost revenue — not lost on price, lost on capacity to respond.

The Move

This is a Custom Build. Not a connector you set up in an afternoon, not an agent you schedule — a real commissioned system, because the work crosses too many sources and the output has to be trusted.

The build: a trigger (a tender lands in the bids inbox or portal) kicks off a pipeline that pulls data sources — the product catalogue, the spec library, costing rules, and the archive of past won proposals. An LLM step reads the RFQ, extracts every requirement, matches products, drafts the technical narrative, and pre-fills the costing model. The destination is a draft proposal in the engineer's review queue, with every figure traceable back to the rule that produced it.

The senior engineer no longer reads twenty pages cold. They review a drafted, sourced proposal and apply judgement where it matters. Response capacity roughly doubles without hiring. The tenders that used to sit now get answered.

And the honest answer to "who would build this?" is: occasionally, us.

Why most 130-person coatings makers miss this

Because they file it under "AI can't do our work — the engineering is too specialised." Which is true, and beside the point. The system isn't doing the engineering; it's doing the reading, matching, and assembling that surrounds the engineering and consumes the expert's scarcest hours. The judgement stays human. The grunt work doesn't have to.

The contrarian point: the bottleneck was never quote quality — their quotes win. The bottleneck was the number of quotes two people could physically produce. A custom build attacks throughput, not quality, and throughput was the thing quietly capping the top line.

The translation

The pattern — a trigger, several data sources, an LLM step, a trusted destination — fits any firm where expensive expertise is trapped in document-heavy, repeatable assembly:

• A law firm drafting first-pass contract reviews against its own clause library and prior matters.
• An insurance broker assembling submission packs from client data, risk schedules, and underwriting guidelines.
• An architecture practice producing planning-application narratives from drawings, site data, and local-authority policy.