How to automate customer enquiries and quotations without losing control

Start with the handoffs, not the AI model

A quotation often starts in an email or message, passes through a spreadsheet, becomes a document and ends up in a CRM. The delays are usually spread across those handoffs. Automating a reply alone will not solve a missing price list or an unclear approval process.

Map one real request from arrival to follow-up. Identify which system holds the customer record, which price list is authoritative, who approves discounts and what happens when a request needs clarification. This map is the specification for the workflow.

Step 1: capture and classify the request

Capture the source message and its identifier, then extract only the fields needed for the next step. These might include the requested product, quantity, delivery location and preferred date. Keep the original message available so a reviewer can verify the extraction.

Do not treat a missing field as permission to invent it. If quantity, configuration or scope changes the price, ask for clarification or send the request to a salesperson. Separate ordinary quote requests from complaints, contract changes and unrelated messages.

Step 2: answer only from approved business information

Use maintained product information, delivery policies and service descriptions to prepare an acknowledgement or answer a routine question. If the answer is not in those sources, the workflow should escalate instead of improvising a business commitment.

An illustrative safe reply acknowledges the request and asks for a missing requirement. An unsafe reply promises a delivery date or discount without checking the relevant system. Define that distinction before go-live, then test it with historical messages after removing unnecessary personal information.

Step 3: calculate prices with rules

AI can interpret what a customer is asking for. Your approved pricing logic should calculate the actual quote. Use explicit product identifiers, quantity bands, tax treatment and rounding rules rather than asking a language model to make up a total.

Store the price-list version used for the draft. If the source data is missing or expired, stop the quotation step. Keep unusual discounts, changed payment terms and non-standard scope outside automatic approval.

Step 4: make human approval useful

A reviewer needs more than an approve button. Show the original request, extracted requirements, selected products, price source and any exception flags alongside the draft quote. Make it possible to edit or reject the draft.

Record who approved the final version and when. If the draft changes after approval, require a fresh approval before sending it. Use a stable request identifier to prevent a timeout or retry from creating duplicate quotes or sending the same message twice.

Step 5: update records and follow up carefully

After an approved quote is sent, update the relevant CRM record with the quote reference and status. Decide which system is authoritative so one automation does not overwrite a salesperson's newer changes.

Follow-ups should have a maximum number of attempts, sensible sending hours and clear stop conditions. Stop when the customer replies, declines or asks not to be contacted. Check the channel's rules and applicable marketing and data-protection requirements before sending automated outreach.

What to measure during the pilot

Start with draft-only operation for one request type. Expand permissions only after the owner has reviewed the exceptions and the measured workflow is working better than the previous process.

• Time from enquiry arrival to first useful response, not just an automated acknowledgement.
• Time from a complete request to an approved quotation.
• Share of drafts accepted without pricing or scope corrections.
• Requests requiring clarification or manual handling, and why.
• Duplicate records, duplicate sends and follow-ups that should have stopped.