How Do You Build an AI Workflow for Business Enquiries?

Build it around three jobs rather than one. It drafts, it flags, and then it stops. The workflow reads the enquiry, works out what type it is and drafts a reply from your own source material. It flags whatever the customer left out. Then a person sends, because prices, commitments and bookings stay with the person whose name goes on them.

The part that goes wrong is quieter than any of that. It's the step where the workflow decides an enquiry needs no reply at all.

TL;DR

• Speed is what people buy. Consistency is what keeps paying. One agreed answer per recurring question beats a faster version of three different answers.
• Write the answers before you write the prompt. The hard part of an enquiry workflow is agreeing what your business's answer actually is, and that work holds its value whether or not a model ever touches it.
• The step that files an enquiry as needing no reply leaves nothing behind to check. Give it a visible list and a written reason, never a delete rule, and read that list every week.
• Flag, don't fill. Any figure, date or scope decision that isn't already in your source material gets marked as a gap instead of estimated.
• Default the uncertain enquiry to a draft. A reply nobody needed costs somebody half a minute. A dropped enquiry costs the job, and you never find out it happened.
• Start on one enquiry type you handle every week, and test it against enquiries you have already answered so you have something to compare the draft with.

What does an AI enquiry workflow actually do?

It reads an enquiry, decides what kind it is, drafts a reply out of material you supplied, and marks what's missing. Everything past that point belongs to a person.

That's a narrower definition than the category usually gets. AI workflow automation means putting AI models inside automated workflows to handle the steps that need judgment, rather than replacing the whole process with a model. 1 The AI reads and interprets; the automation moves things around. 2

Zapier's analysis of 375 mid-market and enterprise companies found that AI inside a workflow is almost always doing one of four things: drafting something for a human to read, filling in records from messy input, making the call on where work goes next, or turning a request into a task someone owns. 3 All four turn up in an enquiry inbox, and three of the four hand their output to a person.

That same analysis found AI accounted for about 18% of workflow steps in those companies, with the rest running on ordinary rules, filters and data moving between apps, and that keeping AI to the reasoning steps cost 71% less to run than routing everything through a model. 4 Design accordingly. Most of an enquiry workflow is plumbing: form to record, record to task, task to notification. The model belongs at the reading and drafting steps and nowhere else.

One thing this is not: the chat window on your website that talks to customers directly. That's a different build with a different risk profile. What follows is the workflow behind the inbox.

Why can't rules handle an enquiry?

Because the enquiry is written by someone who doesn't know your process, and rules need fields.

Say this lands in your inbox on a Sunday night:

Hi, we've got a leak in the back room and water coming through the ceiling, can someone come out this week and how much roughly?

Three requests, none of them in a field. No address, no idea what the roof is made of, no photo, and nothing telling you whether the sender owns the building or rents it. Traditional automation follows fixed rules and does exactly what you tell it, even when the input is obviously junk. 5 It's good at getting that message into a record and telling somebody it arrived. It cannot read it.

Every enquiry has that shape. It arrives in ordinary language, it's missing about half of what you need to answer it, and the sender expects one answer rather than a form to fill in. So the first real step in the workflow is classification: what type of enquiry is this, what is it asking for, what's absent, and does a person need it in their hands right now.

A shared mailbox makes the whole thing harder, because an address like enquiries@ or admin@ belongs to nobody in particular. The sorting and summarising side of that problem has its own write-up in our guide to AI email management.

Consistency is the gain that lasts, not speed

Faster replies are the reason most businesses build one of these. The reason to keep one running is different, and it shows up in the second month.

Ask three people in the same business what the warranty covers, whether the first visit is chargeable, or whether you take work in Gympie, and you can get three answers. Each one sounds reasonable on its own. Together they mean the business holds three positions and doesn't know it. Customers find out on the second email, usually the one where they quote your colleague back at you.

An enquiry workflow only functions if the answers come from one place, which means building it forces a conversation most businesses have never finished: what is our answer to the twenty questions we get asked every week? Writing those down is unglamorous and it's most of the value. Do that and you've built something that pays off with or without AI in the loop. The drafting step is just what makes people use it.

We used AI to automate SEO, enquiries and admin for Glen House Homoeopathy, and enquiry handling is where I'd start in most businesses for exactly this reason. The longer version of that job is in how I automated a health practice. None of it belongs to health practices in particular. Three things hold wherever an enquiry is written by a person instead of filled into a form: it arrives in plain language, it rarely contains everything needed to answer it, and the reply has to say the same thing whoever picks it up.

In my own business, Searchmaxxed, we built AI workflows for creating blog posts, doing SEO and preparing custom proposals. A proposal has the same shape as a quote reply. The client's situation sits on one side, our own approved facts sit on the other, and the document has to agree with both of them.

If the inconsistency you're looking at is spread across a team rather than sitting in your own drafting, that's the case for private AI training at your workplace, where the people who answer the enquiries agree the answers in the room and build the workflow on top of them.

The step that files an enquiry as "no reply needed" is the one to audit

Every other step in an enquiry workflow leaves evidence behind. A draft can be corrected. A flag can be chased. A CRM record can be opened and read back. The step that decides a message needs no response leaves nothing at all, and nothing is also exactly what a lost enquiry looks like.

Deciding not to reply is an action. Marking a message as spam, closing it as a duplicate, filing it as already answered in the thread, or letting an archive rule take it are all decisions with a consequence, and the consequence is silence. Silence has no reviewer. So the accuracy of that step is the one number in your workflow that nobody ever reads out loud.

The general form of that rule, which holds for any step whose output is "no action required", is set out in our piece on AI approval workflows. Here is what it means for an inbox.

• The workflow never deletes and never archives. The no-reply pile goes into a folder, a view or a sheet that exists on purpose and that somebody can open.
• Every item in it carries a written reason: supplier pitch, duplicate of Tuesday's enquiry, already answered below. A reason is something you can disagree with. A silent filter isn't.
• Somebody skims that pile. Weekly is enough for most businesses, and you're reading for one thing only: enquiries that should have had a reply.
• Count those. That count is the only measure of the step's accuracy you will ever get, and it's the number to watch through the first month.
• When the workflow can't tell, it drafts. A reply nobody needed costs half a minute of somebody's attention. A dropped enquiry costs the job and never announces itself.

Published accounts present this step as a straightforward win. One describes a workflow that triages inbound email, filters out spam, identifies qualified opportunities and routes each lead to the right rep. 6 The filtering is genuine and the time saved is real. It's also the step with the least visible failure mode in the whole workflow, which is a reason to instrument it rather than a reason to trust it.

What should the workflow draft, and what should it flag?

It drafts anything that already exists in your source material. It flags everything that doesn't.

Flag, don't fill, is the instruction that does the most work here. If the price isn't on the rate card, the draft says a price needs setting rather than producing an estimate. If the customer didn't give an address, the draft asks for one rather than inferring a suburb from the phone number. A model asked to produce a complete reply will produce a complete reply, gaps included, so the Rules input has to make the gap itself the desired output.

The last row of that table is the one to write first. Published guidance on what shouldn't be fully automated puts customer complaints alongside financial approvals and sensitive client data, in the class of work where AI collects information and prepares a draft while a person reviews and approves the final action. 7 In an enquiry inbox that isn't a compliance formality. A complaint answered with a cheerful template is worse than a complaint answered slowly.

Where should an enquiry workflow stop?

At anything that commits the business: a price, a date, a scope decision or a liability.

Zapier's own rollout guidance frames it as matching the guardrail to what the output does next. A drafted message needs a reader, a write into your CRM needs validation, and an AI step deciding what happens next needs a clear line for when a person steps in. 8 That's the right test, because it grades the step by consequence rather than by how clever it is.

Vendors selling agent platforms draw a similar line, routing routine questions to the assistant and sensitive decisions to human approval. 9 The argument is never about whether the line exists. It's about where "routine" gets set, and that's a judgement about your business rather than a setting in a tool.

My position for a business with a name on the van: the sending decision stays with a person for every enquiry carrying a number, a date or a promise. Not because drafts are unreliable, but because the costs are lopsided. A held draft costs you minutes. A quoted price you can't honour costs you the margin, because you'll honour it anyway.

Day to day that looks like this. Before anyone opens the inbox, the replies are written, the gaps are marked, the records exist and the tasks are assigned. The person reads, corrects and sends. Zapier describes a retailer where a manager reviews every AI-drafted support reply before it goes out and about 70% are sent without an edit. 10 Read a number like that as a description of their setup rather than a target for yours. The part to copy is the shape: review then send, not send then hope.

Reading a draft against your own source material without re-reading every document is a skill in itself, and we've set out the specific checks in how to check AI output.

How do you write the instructions once so the next enquiry reuses them?

Write five inputs: Context, Task, Source, Rules and Output.

• Context describes the business, the customer and the situation.
• Task names the job, which here is a draft reply and a list of gaps, not a sent email.
• Source supplies the notes, documents and approved facts it can use, and only those.
• Rules set the limits, including the instruction to flag missing information instead of filling it.
• Output describes the shape of the finished draft, including where the flags go and how they're marked.

Write them for one enquiry type, not for enquiries in general. The worked examples sit in our guide to writing a five-input AI prompt.

Source is where the consistency argument comes back to bite. If your twenty recurring answers live in one document, Source is a file path and you're nearly done. If they live in four people's heads, that's the job to do before the workflow, and no prompt gets you out of it. Zapier puts the same point more bluntly: AI tolerates a certain amount of mess, but data that's inconsistent, scattered across tools or written in a dialect only your team understands needs fixing first. 11

How do you test it on real enquiries before it touches a customer?

Run it on enquiries you have already answered, because you know what a good reply looks like. You sent one.

Pull twenty from the last month or two, spread across the types you genuinely receive, and include the awkward ones: the one-line enquiry, the one asking three questions at once, the supplier pitch dressed as a customer, the complaint. Run each one through and compare the output with what you actually sent.

Count three things:

1. How many drafts you'd have sent with light edits.
2. How many gaps it named correctly, against how many it quietly filled in.
3. How many it put in the no-reply pile, and whether you agree with every single one.

The third is the one that gets skipped, for the reason above. Zapier's pilot advice covers the rest of it: test with real inputs, get feedback from the people doing the work, adjust the prompts and guardrails, and write down what is and isn't working. 12 Australian automation firms sequence it the same way, building one process, running it for two to four weeks, measuring the result and only then moving to the next. 13

Then keep testing, because the workflow won't tell you when it goes stale. Your prices change, your service areas change, a new question starts arriving every week and the instructions know nothing about any of it. We've written up the ongoing side of that in how to test and maintain AI workflows.

Which enquiry should you build first?

The one you answer every week, where you already know the correct answer, and where the reply comes out of documents you control.

Good candidates: the quote reply that follows a site visit, an availability question, a scope question, a standard reply to a new customer. Leave alone, for now: anything whose answer changes with the customer, anything that touches a complaint, and anything where you'd have to invent the policy before you could write the Source input. If the full job runs across several stages and systems, build the drafting step first and connect things up once the draft is trustworthy. Our walkthrough of building your first AI workflow takes that one step from a blank page to something you can run again.

An enquiry is a good thing to bring to a workshop, because you'll have real material to test the draft against and you'll know immediately whether the answer is right.

You can join the workshop registration for the next date, and you keep the instructions and the tested example afterwards. Private onsite training runs at your team's workplace instead, scoped for roughly 4 to 20 people, with a fixed quote after we've talked through your team's tasks, experience and location.

Frequently asked questions