Automated enquiry processing: from receipt to proposal

Where do enquiries most often stall?

Between receipt and someone taking ownership. The enquiry arrives, but it is unclear who will handle it, what is missing and when the next step is due. An automatic acknowledgement does not solve this alone.

On Reddit, a business owner described slow routing from a shared inbox to salespeople. This is one person's experience, not market research. Its practical question nevertheless applies widely: can you identify an owner and current status for every enquiry received?

Break the journey into receipt, clarification, pricing, quote and customer response. Define minimum information and ownership for each stage. Only then decide which steps AI should handle and which need conventional automation.

What should AI extract from an enquiry?

Extract details that affect the next decision. For installation work, that may mean location, work type, dimensions, timing and available documents. For goods, it may mean product codes, quantities, variants and delivery method. A universal list of twenty fields usually burdens the customer unnecessarily.

Every detail should retain its origin. “Next Friday” requires the message date and sometimes a follow-up question. A product in a photograph may look similar without being identical. The agent therefore separates confirmed values, estimates and missing information.

For an incomplete brief, ask only for information needed for the next step. If a customer is still exploring options, requiring a billing address makes little sense. A well-designed conversation becomes more specific gradually and retains what the person has already said.

What do international examples show?

In its VELUX case study, HubSpot describes three separate changes: creating illustrative visualisations went from 2–3 weeks to minutes, savings of 15 minutes per report and waiting time between consultations fell from 15 to 5 minutes. These are not measured results of automated pricing, nor a promise of equivalent savings for another company. VELUX case study.

“Breeze Assistant was easy to use and improved the customer journey.” — Allyson Zook, Programs Marketing Manager at VELUX, excerpt from the cited study.

An interesting newer example is Hexa, from Y Combinator's Spring 2026 batch. It describes connecting distributors' enquiries to products and ERP. This is a startup's presentation, not independent proof of performance. The useful design principle is the connection between steps: receiving text is the beginning, not the end of the process.

How should AI connect to pricing and CRM?

AI should interpret the request and select the right supporting information. A specific system must supply current prices and commercial restrictions. The total is then calculated reproducibly according to rules.

A quote has its own identifier and version. If a customer changes the quantity, the document already sent must not be silently overwritten. The CRM must show which version they approved and which details the order is based on.

Which exceptions should be designed in from the start?

Duplicate enquiries, missing contacts, nonexistent products, inaccessible attachments, non-standard discounts and deadlines beyond capacity. These are not minor edge cases. They often account for the most expensive part of sales administration.

In the n8n Community an audit author highlights, among other issues, missing duplicate checks before writing to HubSpot. The list reflects the author's experience, not a guarantee of a complete security audit. For this scenario it suggests a concrete test: a repeated form submission must not create a new sales opportunity without recognising the connection.

Handle simultaneous work by salespeople explicitly. If a salesperson agrees a different scope by phone, the agent must not continue from the old message. The current sales status is decisive, not simply the last text in an email thread.

How can the system continue without constant manual work?

Autonomy can be assigned by case type. A standard enquiry with complete information can proceed automatically to a quote within a defined price limit. Non-standard cases return for clarification or go to a designated salesperson. People then handle the reason for the exception, rather than every routine step.

Permissions should be technically constrained: available products, maximum discount, authorised recipient and rules for changing dates. A prompt saying “be careful” is not enough. The application must reject actions outside the approved scope even when the model proposes them.

Record the confirmed outcome when sending a message. If the provider times out, the agent should check the status first. Retrying an uncertain send may leave the same customer with two different quote drafts.

What should you measure beyond first-response speed?

Measure time to a usable quote, completeness of briefs, lost enquiries and corrections. A quick “we'll be in touch” can improve first-response metrics without accelerating the actual decision.

Illustrative example: a company receives 120 enquiries a month and spends 12 minutes collecting information for each: 24 hours. Saving eight minutes per case produces 16 hours of gross savings. Administration, corrections and review of unclear cases must be deducted. Measure the effect on closed deals separately; it cannot be inferred automatically from completed forms.

What should a first working pilot look like?

Choose one service or product group with stable rules. Prepare historical enquiries and finished quotes for a real comparison. Include rejected cases, customers with incomplete details and requests that changed during negotiations.

The pilot should demonstrate the entire flow, even within a limited scope: receipt, clarification, checks, a prepared quote and traceable status. Once that foundation is reliable, you can add lead automation and further steps handled by an AI sales agent.