A quote format that rejects prices without a source

The central risk in quoting automation is an unsupported price. One invented number in an offer can erase margin or damage the client relationship. A model asked to price a line with missing data may still produce a plausible answer.

A common control is an instruction to quote prices only from the price list. That instruction can fail on an unusual case, especially in a long context.

Enforce the rule in the result format

In our RFQ quoting system, every result must pass a format that rejects prices without an approved source.

A quote line can exist in exactly two forms. PricedLine carries a price and a mandatory source field: a cited precedent from quoting history or an approved pricing rule. The type's constructor simply fails without a legal source. EscalatedLine represents a line that could not be priced, and it has no price field. It carries a description of the missing basis instead, plus a question ready to send to the estimator or the client.

For generated text to become a quote line, it must pass one of the two validation paths. The system has no valid result type for an unsupported price.

Where the agent comes in

The model and workflow handle the work that needs iteration. A customer's RFQ may be incomplete and written in their own words. The system breaks it into lines and parameters, then searches for a precedent, checks the rules, and refines the specification until it exhausts the available sources. The purpose of every tool call is recorded before execution, so the team can trace the run later.

The workflow has two valid outcomes: a sourced price or an escalation with a concrete question. It cannot return an approximate price without a source.

When the pattern creates value

The best fit is a quoting process with recurring materials, operations, and parameters, where estimators repeatedly consult the same precedents. Repetition in the product matters more than similarity in wording. Two parts described with similar language may have different tolerances, machining steps, or inspection requirements. Matching must use cost-driving parameters rather than text resemblance alone.

Review the quote history before implementation. A usable precedent should retain the catalog version, date, currency, quantity, material, operations, and the estimator's decision. An old price without those conditions can be misleading. Every precedent needs a validity rule or a fresh approval path.

High-mix, one-off work can still benefit from structured RFQ intake and better escalation even when few lines receive an automatic price. The useful outcome may be a clean specification and a precise question for the estimator.

A staged implementation

The first stage can stop after line extraction and missing-parameter detection. The team checks units, quantities, materials, tolerances, and operations. A mistake here invalidates any later price lookup.

The second stage introduces precedents and rules for one product family. Each source needs an owner, an effective date, and applicability conditions. Estimators should assess both the proposed prices and the quality of escalations. “More information required” is weak; a useful handoff asks for the missing tolerance, material grade, or delivery target.

The third stage adds approval of the complete sheet and integration with CRM or quoting software. Sending the offer remains a deliberate action. Correct line prices do not prove that discount, lead time, freight, and commercial terms are ready.

Interpreting the measurement

Zero unsupported prices demonstrates the data contract. It does not demonstrate competitiveness, current margin protection, or compatibility with the client's machine park. Those require separate measures: variance from estimator decisions, corrections after review, useful escalations, and time to a complete offer.

Model-reading cost is a small part of the economics. Maintaining rules, improving quote history, and assigning experts to exceptions are likely to matter more. A pilot should show whether preparation time saved exceeds the effort of reviewing suggestions.

RFQ automation checklist

1. Which parameters drive cost for each product family?
2. Does every precedent retain date, quantity, currency, and applicability conditions?
3. Who may approve a pricing rule, and when does it expire?
4. Is a line without evidence structurally unable to carry a price?
5. Does every escalation ask the estimator or customer one specific question?
6. How will the team measure corrections, source freshness, and quote lead time?
7. Who approves the complete sheet, discounts, delivery terms, and sending?

The measurement and its limits

The test covered 60 generated quote lines. Forty received a price backed by a precedent or an executed rule; the other 20 reached the estimator without a price. No priced line lacked a source, and the result format rejected both attempts to construct one. A separate small-model RFQ read using gpt-5-mini consumed 714 units of model text, took 5.57 seconds, and cost about PLN 0.0034.

This evaluation checks the type contract and the graph. It does not decide whether a historic price still protects today's margin; that remains with the estimator and salesperson. What the system can enforce is that every quoted number has a checkable origin.

A reusable pattern

The pattern transfers to other domains. Wherever one unsupported field would be expensive, ask whether the result contract can require its basis. An answer without a citation, an accounting entry without a precedent, and a record without a source can all fail validation before they enter the process.

Full measurements and system screens are on the case page. If quoting at your company depends on two people, a free process scan is a good first step.