AI quotation agents: make every proposed line verifiable

Start with a repeatable offer whose rules are already understood. A standard maintenance proposal is a more manageable pilot than a project whose scope changes after a site visit. Gather representative requests and ask an experienced estimator to describe the decisions they actually make, including what they need to clarify before calculating a price.

Define the permitted output: an internal draft, a proposal awaiting approval or a document ready to send. Keeping transmission under human control narrows the first review. CRM and ERP automation can attach the draft to its opportunity and retrieve supporting information while preserving existing approval permissions.

Separate interpretation from commercial calculation. The agent extracts quantities, finds possible references and asks for clarification. A pricing service then applies authorized rules. The function calling documentation explains how a model requests tools that the application executes. The application remains responsible for what those tools actually do.

A traceability record should let another person reconstruct a line. Retain the request excerpt, selected reference, unit, quantity and catalogue version. Include the rule used to turn those inputs into an amount, followed by applicable taxes and exclusions under the company's approved configuration. Store these details with the draft version.

An agent's retrieved document may help explain a product. The rate used for calculation needs an authorized, dated and identifiable source. When the catalogue changes, an open draft needs a defined review path. Specify whether a new version invalidates the quote, requires recalculation or respects an existing price reservation under an explicit commercial rule.

A missing price creates a question for the sales team. So does an ambiguous unit: two cases can mean different package sizes. Add explicit pause reasons for an outdated catalogue, a discount outside delegated authority, an unreadable attachment and incomplete scope. The case should explain what is missing and identify who can resolve it.

The model must not treat a customer's email as permission to grant a discount. Incoming requests are data to interpret; verified identity and business rules determine authority. AWS guidance on excessive agency recommends permission boundaries for agent workflows. Enforce those boundaries inside the services the agent calls.

The draft should distinguish confirmed lines from assumptions. Avoid showing a reassuring final total while an important item remains unresolved. Prepare a clarification request and a clearly labelled partial version instead. A custom CRM and ERP can manage revisions and approval thresholds when those requirements exceed the current application's capabilities.

Imagine a maintenance supplier receives: “Add two visits and use the same package as last year.” The earlier contract may contain a replaced rate. The two visits might be additional visits or the new total. The agent retrieves the old contract and asks a question about frequency. It waits for the answer before calculating.

Once the scope is clear, the pricing service uses the current catalogue and retains the earlier contract as context. If the salesperson wants to honour the previous price, they follow the approval rule. This scenario illustrates a desired behaviour. It reports no actual Daillac customer, time saving or executed product test.

Build a set of synthetic or anonymized requests with reference answers approved by sales. Include a complete request, missing price, ambiguous unit, prohibited discount and revised catalogue. For each case, write the expected decision: produce a draft, request clarification or refer the case to a named person. Record the reason as well as the destination.

Review extracted fields, amounts and escalation decisions separately. An accurate total can still quote the wrong product. Clear prose can omit an important exclusion. Measure review effort and corrections as part of the exercise. Keep rejected requests in the denominator so the pilot does not report only its easiest cases.

A custom web application can display the original request, proposed lines and evidence together. The approver should be able to change a selection and see amounts recalculated by the authorized pricing service. A catalogue or scope change after approval must follow the agreed rule for reapproval before the document is sent.

Set conditions before the pilot: no unauthorized discounts, invented prices or transmission without required approval. Define an explicit commercial scope and a route back to manual processing. A material error should pause affected cases while an assigned owner investigates. Make the pause visible to the people waiting for proposals.

The decision record includes catalogue versions, evaluated cases, corrections and sales approval. The company can then extend one product family at a time. Review frequency should follow price changes and observed exceptions, with an additional check whenever the pricing service changes. Keep a representative set of difficult requests for each subsequent evaluation.

Estimate the cost of keeping the evidence usable. Someone must maintain catalogue ownership, remove obsolete references and resolve conflicts between sales documents. These tasks belong in the operating budget. Assigning them before the pilot prevents the estimator from becoming the informal administrator of an unattended pricing database.