Route incoming customer inquiries with AI

The short answer

Route inquiries in two passes. First, classify each new message into a small set of categories. Second, send only the categories that need a person into the inbox a person actually watches. Everything else gets an automatic acknowledgement and a scheduled follow-up.

That is the whole idea. The AI is doing triage, not conversation.

What problem this solves

Most small businesses do not lose inquiries because they never reply. They lose them because a hot lead and a newsletter question sit in the same unread pile. The owner triages by mood and time of day instead of by urgency.

A routing workflow makes the priority order explicit and repeatable.

Who this is for

• You receive more than roughly ten inquiries a week.
• Inquiries arrive through more than one channel (email, website form, social messages).
• One or two people handle all of them.
• You already have a written idea of what a good lead looks like.

Who this is not for

• You receive a handful of inquiries a month. A label and a daily check are enough.
• You have no shared inbox yet. Fix that first; AI triage on a personal inbox creates more confusion than it removes.
• Your inquiries are mostly regulated personal data you are not comfortable sending to a third-party model. Use local rules-based filtering instead.

What you need before starting

1. A list of five to seven inquiry types, in your own words. For example: new job request, quote request, existing customer issue, billing question, vendor pitch.
2. A definition of "urgent". Revenue at risk, a deadline, or an existing customer down are common choices.
3. One inbox or help desk where every channel lands. Routing across four apps is the problem, not the solution.

The workflow

1. Consolidate the intake

Forward website forms and social messages into the shared inbox. Do this before adding any AI. One queue is the foundation; the model only sees what reaches the queue.

2. Write the categories as instructions

Turn your five to seven inquiry types into short definitions with examples. Be explicit about the default. A useful prompt shape is:

Classify this message as one of: quote_request, existing_customer, billing, vendor_pitch, other. Return only the label. If the message is ambiguous, return other. A message is urgent if it mentions a deadline within seven days or an existing customer who cannot use the product.

Vague instructions produce confident nonsense. Concrete examples do most of the work.

3. Add a confidence rule

Route anything the model is unsure about to a human, not to the automatic path. A "maybe" that reaches a person costs a minute. A "maybe" that gets an automatic reply costs a customer.

4. Build the routing rules

In your help desk, send each label to a folder, assignee, or SLA. Keep the automatic replies short and honest: confirm receipt, state when a person will reply, and give a way to escalate.

5. Review the first two weeks

Read every classified message for the first ten working days. Count mistakes. Adjust the category definitions, not the model. Most failures are ambiguous instructions.

6. Keep a human on the send button

Do not let the system send a substantive reply on its own in the first month. Draft, review, send. Automate the acknowledgement and the routing, not the relationship.

Tools

You do not need a specialised product. A help desk with rules and an AI triage feature is enough; many shared inboxes now include one. A general assistant works for the first pass if you are willing to paste messages in manually, but that does not scale past a few dozen inquiries a week.

What it costs

Entry plans for shared inboxes and help desks commonly sit in the low tens of dollars per seat per month, and several have usable free tiers. Verify current pricing before you commit; it changes. The larger cost is the setup afternoon and the two weeks of review.

Limitations and privacy

• Classification is not understanding. Sarcasm, mixed-language messages, and forwarded threads will confuse it.
• Do not send more personal data than the task needs. If you must process sensitive information, check the vendor's data-processing terms and your own obligations first.
• Keep an audit trail. If a lead was misrouted, you need to be able to see why.
• This workflow does not replace a CRM. It decides where a message goes; it does not remember the customer.

A cheaper alternative

Rules-based filters are free and predictable. Search for "quote" or "invoice", route by sender domain, and use a form field for inquiry type. If your categories are stable and your volume is modest, rules may outperform a model for a fraction of the effort.

Do you actually need AI for this?

Often, no. If your inquiry types are stable and your volume is under roughly twenty a week, a short form with a dropdown and two inbox rules will do the job. AI becomes useful when messages are unstructured, when the mix shifts often, or when the volume makes manual reading the bottleneck.

The honest test: if you can write the routing rules on one page and they stay true for a month, use the rules.

How to measure it

Track three numbers for four weeks before and after:

• Time from first contact to first human reply.
• Percentage of inquiries that received any reply within one business day.
• Percentage of hot leads that were routed to the right person on the first attempt.

If those do not move, the workflow is not working, regardless of how good the classification looks in a demo.

Notes on sourcing

This guide describes a general workflow and is labelled researched, not tested on your systems. Cost ranges are approximate and must be verified against current vendor pricing. No performance or revenue claims are made here.