Which manufacturing orders can an AI quotation assistant price automatically, and which need review?
Treat “AI prepares a quotation” and “the business commits a price to a customer” as separate permissions. A standard order may progress towards controlled automatic release only when required inputs are complete, the configuration is approved, the calculation is reproducible and commercial exposure remains within delegated limits. Missing drawings, first-time processes, exceptional lead times, high value, weak margin or conflicting terms require human review. Confidence claimed by a model is not a release control.
The business context
This framework is for manufacturers whose prices depend on specifications, materials, process routes, quantity, delivery and customer terms. Sales may spend days consulting engineering and costing teams. A model that produces a total quickly can still omit an operation, use an obsolete revision or promise an impossible lead time.
Tata Consultancy Services describes a manufacturing GenAI sales use case that gives representatives configuration-compatibility insight and creates tailored quotations for selected configurations. The page describes assistance and augmentation, not unrestricted autonomous commitment.
The NIST AI Risk Management Framework treats risk as a combination of likelihood and impact and makes risk tolerance contextual to the organisation and use case. It calls for stronger management of higher-impact risk and continued assessment even in lower-risk settings. The control model below applies those principles to quotation; it is an inference for product design, not a threshold prescribed by NIST.
Decide what the assistant is authorised to do
Separate four layers:
- Intake: extract product, quantity, material, delivery and missing information from an enquiry.
- Recommendation: identify the product family, allowed options, similar completed work and engineering questions.
- Draft calculation: invoke approved material, labour, waste, subcontracting, freight, tax, discount and margin rules.
- External release: issue a customer-facing offer with validity, scope, exclusions and commercial terms.
Most first releases should automate the first three and keep the fourth with an authorised employee. After a narrowly defined class of standard orders proves stable on new work, the business may grant limited automatic release for that class.
A practical decision matrix
| Condition | Action | Why |
|---|---|---|
| Standard product, complete fields, approved combination, current price rules | Draft automatically; consider limited release after evidence | Inputs and arithmetic are reproducible |
| Standard product with expedite, special packing, split delivery or unusual credit | Prepare draft, require commercial review | Terms change cost or commitment |
| Missing drawing revision, material, tolerance or quantity | Do not price; ask for the named fields | A precise price has no reliable basis |
| New process, non-standard combination, substitution or trial build | Engineering and costing review | Historical similarity does not prove manufacturability |
| Value, discount, margin or liability exceeds delegation | Escalate approval | Consequence of error is higher |
| Model, rule service or source data is unavailable or unexplainable | Stop automation and route manually | The result cannot be audited |
Management sets the limits for its products and contracts. AI should not invent discounts. Where the final price follows deterministic rules, use AI for extraction and recommendation, a calculation service for versioned arithmetic, and accountable people for exceptions and commitments.
Build a connected evidence set
Do not collect quotations alone. A useful historical record connects the enquiry, drawing and product revision, selected route, cost components, applied rules, approval, issued offer, customer changes and later evidence of repricing, rework or delivery failure. A past quote without actual process and cost data may preserve an old mistake rather than a profitable pattern.
Version product-option compatibility, units, materials, machine and process capability, standard time, waste, subcontracting, packing, freight, tax, currency, discount authority, validity and customer agreements. Assign a business owner and update rule to each source. Retain the versions used for every quotation so a later update cannot rewrite history.
For PDFs, drawings and emails, retain source location and extraction signals. High confidence is not business validation. Critical dimensions, material, quantity and revision may require explicit confirmation. Conflicting units or document versions should stop the workflow.
Design review as a product capability
A review screen should show more than a total and an Approve button. Present the enquiry, extracted fields, missing or conflicting items, rule versions, cost and margin composition, the reason for any similar-order reference, and which work came from AI versus deterministic rules. Classify overrides such as customer term, engineering correction, purchase-price change or management exception.
Corrections can enter an evaluation set, but they must not silently change production rules. Product, sales, engineering, costing and finance should own named fields and decisions. The employee who makes the external commitment remains traceable.
Verify a first release
Choose one product family and completed historical orders. Hold out later examples for blind evaluation rather than tuning and accepting on the same set. Measure field extraction, missing-item detection, invalid configuration detection, calculation agreement, recall of work that should be reviewed and review time.
Classify errors by consequence: omitted operation, wrong drawing revision, unit conversion, missing subcontract or freight, unauthorised discount, infeasible delivery and absent terms. A single overall accuracy number can hide the serious classes. A critical error may be a reason to pause automatic release even when the aggregate looks strong.
Acceptance should demonstrate that:
- identical inputs and rule versions produce a reproducible result;
- missing critical information prevents a sendable price and creates a precise question;
- invalid combinations, stale prices, weak margin, high value and exceptional terms route correctly;
- reviewers can inspect sources, calculation, versions and changes;
- validity, scope, tax, delivery and exclusions reach the customer document;
- retries, concurrent changes and delivery failure do not create conflicting valid offers;
- reporting distinguishes drafts, overrides, rejections and subsequent corrections by risk class.
Roll out by authority, not by headline accuracy
Start with intake, missing-information prompts and an internal draft. Add calculation for a controlled product family. Consider limited release only after new-order evidence is stable. Before each expansion, review error classes, time saved, response time and realised margin. A source or pricing-rule failure must disable automatic release without disabling the manual process.
Human review is not evidence that the product failed. For consequential exceptions, it is part of the intended operating model. The automation earns its value by concentrating specialist time on orders that genuinely require judgement.
Before commissioning the production workflow, use the AI demonstration acceptance guide to set samples, blind evaluation, pass criteria and stop conditions for the selected product family.
Sources
- Tata Consultancy Services: TCS GenAI for Manufacturing — the Sales and marketing section directly describes configuration insight and tailored quotations that augment sales representatives. Accessed September 14, 2026.
- NIST: AI Risk Management Framework 1.0 — contextual risk tolerance, impact-based prioritisation, evaluation and monitoring; it does not set a manufacturing quotation threshold. Accessed September 14, 2026.
- Wavesteam: 3D mould quotation case — one implemented combination of model interpretation, specialist parameters and approval workflow; its results are not transferred to another manufacturer. Accessed September 14, 2026.
This is a product-scope and acceptance framework. Authorised business owners remain responsible for price, tax, contract, quality and delivery commitments.