A salesperson requests a discount in chat. Finance approves a percentage, the catalogue changes overnight, and a colleague sends a revised PDF. When the customer accepts, the order contains a different delivery term. The business has several records of agreement but no reliable answer to one question: which complete commercial version was authorised?
For sales operations teams in Belgium and the Netherlands, an AI-assisted quote desk should preserve that answer. AI can structure a request and explain deviations; deterministic services calculate prices, named people approve exceptions, and the system binds every decision to one price-and-terms version. This is a quotation-control workflow, not another follow-up sequence.
Define the commercial record before automating
Start with one quote identifier and immutable revision identifiers. Each revision references the customer legal entity, catalogue snapshot, product codes, quantities, currency, tax treatment, discount rules, cost basis, payment terms, delivery commitments and expiry. Record where each field came from and who can change it. A live price-list link is useful context, but it is not a historical price snapshot.
Microsoft's Dynamics 365 quote documentation illustrates why state matters: draft prices can update when the quote is opened or edited, whereas active and closed quotes do not follow current price-list changes. Test your actual CRM/CPQ configuration rather than assuming every saved quote is frozen.
Give sales operations ownership of the revision lifecycle, finance ownership of margin and discount policy, and legal or commercial leadership ownership of non-standard terms. A model's confidence score is not commercial authority. The customer-facing document, structured commercial snapshot and approval record must all refer to the same revision.
Build a version-bound approval workflow
- Capture the request. Attach the customer request to the opportunity and identify missing quantities, delivery location, validity period or payment terms. AI may propose structured fields from an email, but ambiguous requirements return to the salesperson. Do not let a plausible completion become a promised service.
- Pin the inputs. Select an effective catalogue revision and a dated cost basis. Store the actual inputs needed to reproduce the calculation, including rounding, currency and tax rules. Confirm whether customer-specific agreements override standard prices. A missing or expired source blocks the quote rather than silently using a cached spreadsheet.
- Calculate deterministically. A tested pricing service applies quantity breaks, permitted discounts and the approved margin formula. Keep model-generated arithmetic out of the authoritative totals. Finance decides which costs count and whether freight, implementation or rebates affect the floor; a single generic margin field may hide those differences.
- Classify deviations. Compare the proposed revision with policy. Separate price concessions from payment, liability, delivery and service-level changes. AI can suggest a classification with citations to the request and policy; uncertain or contradictory classifications go to review. Never interpret silence as acceptance of an exception.
- Route the exact revision. Show approvers the full snapshot, change summary, calculated margin and requested exceptions. Record approver identity, delegated authority, timestamp, decision and revision identifier. A chat message saying “fine” without a bound version is insufficient. Escalate absent approvers through an explicit delegation policy.
- Lock and send. Generate the customer document from the approved snapshot, retain its digest and send through an authorised account. Record recipient, channel, document identifier and send result. Restrict the automation account so it cannot grant its own approvals. A timeout creates an uncertain-delivery exception; check delivery evidence before resending.
- Reconcile acceptance. Match the customer's acceptance or purchase order to the sent revision, line items, amounts, terms and validity. A changed quantity, handwritten term or expired offer needs review. Only a matched acceptance may create the order; retain the comparison and any approved resolution.
Use optimistic concurrency or equivalent version checks at approval and send time. If another user edits the quote while approval is open, invalidate the pending decision and present a new revision. Do not attach yesterday's approval to today's PDF simply because the quote identifier stayed the same.
A reusable quote-deviation matrix
The following is a design worksheet, not a recommended discount policy. Each company must set its own limits and authorised roles. Store the approved policy revision alongside every routing decision so a later rule change does not rewrite the historical justification.
| Deviation | Required evidence | Decision owner | What invalidates approval |
|---|---|---|---|
| Price below standard list | Catalogue snapshot, discount and calculated margin | Role authorised by finance policy | Quantity, cost basis, price or discount changes |
| Non-standard payment terms | Requested terms and applicable customer agreement | Finance or credit owner | Payment schedule or customer entity changes |
| Delivery or service commitment | Capacity confirmation and proposed wording | Delivery owner | Scope, dates or service-level changes |
| Contractual exception | Exact clause and proposed replacement | Authorised legal/commercial reviewer | Any substantive clause change |
Consider a synthetic example: ten units at €100 produce €1,000 before tax. An 8% discount produces €920. With an explicitly assumed total cost of €700, gross margin on revenue is (€920 − €700) / €920, about 23.91%. These are illustrative inputs, not a client result or an acceptable margin recommendation. If quantity or cost changes, recalculate and re-evaluate the approval rule.
Keep the accepted order tied to the offer
CRM conversion is not the end of control. Microsoft's order-creation guidance describes copying products, price list and currency from an activated quote, but integration settings affect price-lock behaviour. Verify the resulting order values, not merely that an order exists.
The price-lock documentation also distinguishes locked and current pricing. In your acceptance tests, reopen the order after a catalogue change and inspect permitted edits. Compare the order snapshot with the approved quote and customer acceptance before releasing fulfilment. A legitimate difference needs a named reviewer and traceable resolution, not an automatic overwrite.
Use an idempotency key based on the accepted quote revision and customer acceptance identifier. Duplicate webhook delivery must not create a second order. When an ERP call times out, query the target for the existing external reference before retrying. Keep “accepted”, “order creation uncertain” and “order verified” as different states.
Constraints for a Belgian or Dutch rollout
Begin with one product family, one currency and one established approval policy. Configure customer-facing Dutch, French or English documents without changing the commercial snapshot. A translated clause still requires an approved equivalent; the model must not improvise legal wording. Validate tax treatment and cross-border obligations with the responsible finance or legal specialist rather than asking the assistant to decide them.
Restrict access to negotiated prices and costs, minimise customer data sent to the model, and define retention and audit access. Treat customer emails and attachments as untrusted inputs: a request to ignore discount policy is content to classify, not an instruction the workflow can execute. Preserve a manual queue for unusual bundles, missing costs and conflicting agreements.
This approach works where the company has reproducible pricing and accountable exception owners. It will not repair an undefined pricing policy or create delivery capacity. If every offer is bespoke, first standardise the decision record and review checklist; introducing autonomous quotation would only make disagreement faster.
Measure control and effort together
Establish a baseline over representative quotations before the pilot. Track request-to-approved-version time, active reviewer minutes, revisions per quote, exception age and acceptance-to-verified-order time. Separate waiting time from labour; faster elapsed time alone is not a cash saving.
Pair those measures with control outcomes: approvals without a complete revision reference, sent documents that differ from approved snapshots, duplicate orders, expired-offer acceptances and unresolved order mismatches. Review a sample with sales, finance and delivery every week. A pilot succeeds only if effort improves without weakening those controls; agree stop conditions before enabling live sending.
Explore Intyb's sales automation service for this operational boundary, our work with Belgian businesses, and the separate guide to sales follow-up in Belgium and the Netherlands. To map one controlled quotation pilot, contact the Intyb team.