RFQ Automation in Manufacturing: What to Automate Before the Estimator Signs Off

Can AI read drawings and draft quotes?

Yes, within limits. We recommend testing extraction first on typed title-block fields and flat BOM tables, on a 2D drawing or a native CAD export, and routing hand-marked redlines, multi-level assemblies without a machine-readable BOM, and GD&T (geometric dimensioning and tolerancing) call-outs to an estimator by default.

Plan for RFQ packages that bundle a drawing PDF, a spreadsheet BOM, an email with quantities and a due date, and sometimes a reference to a prior part number.

• Hand-marked redlines on a drawing — route these to an estimator to read the markup before anything downstream treats the dimension as confirmed.
• Multi-level assemblies with no machine-readable BOM — route these to an estimator to reconstruct the parts list from the drawing.
• GD&T call-outs — route these to an estimator, since interpreting them correctly takes engineering judgment.

Automation earns its place by assembling the readable parts of a package into one structured record — part number, revision, quantities, material and a routing candidate (the operations and work centers the part needs) — and flagging anything it can't read with confidence. Cost estimators already read blueprints and technical documents to prepare estimates ^[1]; automation's job is to do that reading faster on the requests where the drawing is clean, not to replace it on the requests where it isn't.

The RFQ automate-vs-estimator table

AI quoting in manufacturing moves through the same five stages whether a person or a system does the work: intake, extraction, costing, drafting the quote, and approval. The table below assigns each stage automate, assist, or human-only, so a team can see where to point software and where to leave the decision alone.

In Dynamics 365 Business Central, for example, you can negotiate through a sales quote and convert it to a sales order or sales invoice when you agree on the sale ^[2]. We recommend keeping every automated draft at the quote stage, so nothing becomes an order until that agreement.

Which quote decisions must stay human?

We recommend that pricing exceptions, margin and discount calls, and the decision to quote a request at all stay with the estimator and sales leadership. BLS lists among cost estimators' duties calculating, analyzing and adjusting estimates, and working with sales teams to prepare estimates and bids for clients ^[1]. Software can hand an estimator a cost-based draft; it shouldn't hand a customer a price.

• Setting the final price and any deviation from standard rates or margin targets.
• Deciding whether a request is quoted at all, or declined, deprioritized, or routed to a different team.
• Approving a drafted quote before it leaves the building, the same way Business Central lets you add an approval workflow that can require a manager to approve large sales to certain customers ^[2].
• Excluding requests your compliance team or counsel flags for review (see the callout below).

What data do we need before automating quoting?

Four inputs decide whether a quote lane is ready to automate before any software touches it. Each one is a data-quality problem first and a tooling problem second. We recommend fixing the data first; it's useful to estimators whether or not the automation goes ahead.

How do you measure quote-turnaround impact?

We recommend measuring the change against your own history: a published figure won't reflect your request mix, so treat every metric below as one you track yourself.

• Cycle time from RFQ receipt to quote sent, split by request lane, since a lane with clean drawings and a lane with messy ones shouldn't be averaged together.
• Win rate by lane, before and after automating that lane, over enough weeks that one unusually busy or quiet period doesn't decide the comparison.
• Estimator touch time per quote: minutes spent correcting an automated draft versus minutes spent building one from scratch.
• Rate at which an automated draft is corrected before approval — a high correction rate on one lane is a reason to check the data readiness checklist above before treating it as a tooling problem.

Where to start: a first scoped workflow

Build any connection between the quoting workflow and the ERP's own records through a documented interface, not a direct copy of the database that will drift out of date. Dynamics 365 Business Central, for example, supports this kind of external connection through its own point-to-point Connect apps, which Microsoft says are typically created using standard REST APIs to interchange data ^[3].

1. 01
   Pick one RFQ lane

   Start with a narrow, high-volume request type — repeat parts from existing customers, for example — rather than every incoming RFQ at once.

2. 02
   Build the structured record first

   Get intake and extraction working reliably on that lane before costing touches it, and route anything low-confidence to a person instead of guessing.

3. 03
   Wire costing to the real rate table

   Connect the costing step to the same routing and rate table your estimators already use, through the documented interface described above, rather than a copy that will drift.

4. 04
   Draft, don't send

   Have the system produce a quote draft inside the ERP's own quote record rather than emailing or posting anything to a customer directly.

5. 05
   Run it alongside your estimators

   For a defined period, have estimators review and correct every automated draft, so disagreements between the system and the estimator show up before go-live, not after.

6. 06
   Measure before scaling the lane

   Confirm the cycle-time and correction-rate baseline on the first lane before adding a second request type to the automated workflow.

When this doesn't apply

• Highly engineered, one-off assemblies where every job needs new routing and material research from scratch — there's little repeat structure for automation to learn from.
• Very low RFQ volume, where building and maintaining the extraction and costing pipeline costs more estimator time than it saves.
• Controlled drawings or defense-related RFQs that your compliance process excludes from automated handling, per the callout above.
• A BOM or rate table that isn't trustworthy yet — fix the data readiness problem first; automating on top of bad data just produces wrong drafts faster.

Frequently asked questions

Related reading

• For how a drafted item gets reviewed and approved by a named person before anything moves forward, see document review with human approval.
• The AI agent implementation service describes how a scoped workflow like the one above is built, tested, and handed over.
• On why every drafted item should reach a named person before it's acted on, see human-in-the-loop AI governance.
• If RFQ automation is one of several AI candidates competing for budget, the 360° AI Blueprint ranks it against the rest.

Sources

1. U.S. Bureau of Labor Statistics, "Cost Estimators: Occupational Outlook Handbook" (accessed 2026-09-30)
2. Microsoft, "Overview of tasks to manage sales - Business Central" (accessed 2026-09-30)
3. Microsoft, "API (v2.0) for Dynamics 365 Business Central" (accessed 2026-09-30)