Vertical AI: A Practical Guide

Suppose you're building an AI returns assistant for online stores. A retailer would use your product to help its customers return things they've bought.

Let's follow one imagined customer to see what your product would need to do.

They've bought a pair of shoes. The shoes arrive, but they're too small. The customer opens the store's app and tells the assistant, “These don't fit. Can I return them?”

Your assistant politely explains the return policy. The customer still has to find their order, check whether the shoes qualify, and work out how to send them back.

Now imagine a different version of your product. After the customer signs in, it shows their recent orders. They select the shoes. It checks whether the shoes can be returned under the rules that applied when they bought them, asks what went wrong, and helps them submit a return request. They can see what happens next.

Both versions might use the same AI model. The second can also use the order system, apply the store’s rules and save the request.

It's a useful place to start thinking about vertical AI.

Start with the shoes

Vertical AI means AI built for a particular industry or kind of work. Here, we're looking at it from the perspective of a team building a returns assistant for retailers, not a team building an online store. The retailer is your customer; the shopper uses your assistant to get help. This is an imagined product, not a report of a real customer's experience.

A general assistant can explain what returns usually involve. A product built for this job needs to work with the actual order and the store's rules.

Which pair is being returned? When was it delivered? Was it sold as non-returnable? Has a return already been started?

The customer shouldn't need to know which of those questions matters or explain the store's own policy to its assistant. They should be able to say the shoes don't fit and get help with the right next step.

The retailer decides its return policy. Your team needs to make sure the assistant can use that policy and connect to the order system. If the policy is unclear, you need to resolve it with the retailer. The model can't settle it for you.

A good reply still leaves work to do

Suppose your assistant lets the customer ask for a return, but isn't connected to the system that records it.

The assistant says, “Your return has been approved.” But nothing has changed in the store's system. No request was saved. The customer waits for instructions that never arrive.

The reply sounded helpful. The product failed.

A better version would save the request and show its status. If it couldn't save it, it would say so. If a member of staff needed to review it, the customer would know they were waiting for a decision.

This is ordinary product work. AI doesn't make it less necessary.

Shopify's return rules, for example, let merchants set return windows, fees and final-sale exclusions. Its documented return process treats creating a return and issuing a refund as separate steps. A request to send something back isn't the same as money reaching the customer's account.

Those are existing software capabilities, not evidence that an AI assistant improves returns. Before adding AI, your team should understand what the store can already do and where customers still need help.

Where AI could help

For a straightforward return, selecting an order and pressing a button may be easier than having a conversation. There's no reason to make someone type a paragraph when a couple of taps will do.

But people don't always arrive with a straightforward request.

Go back to our customer. This time they say, “The shoes are too small, but I used a discount code. If I send them back and buy the next size, will I lose the discount?”

Now the product has something useful to untangle. The customer wants shoes that fit and wants to know what a replacement will cost. They're asking about a return, but an exchange may be what they need.

Your assistant could recognise that, look up the relevant options, and explain them using the actual order. If the store's rules don't answer the discount question, it should pass it to someone who can decide, along with the order and the customer's question.

The AI helps interpret what the person means. The store's software supplies the order details and applies its rules. The retailer's staff handle exceptions your product isn't allowed to resolve.

We don't need the model to invent a policy. We need it to help the customer navigate the one that exists.

What you need to agree with the retailer

Before your team builds the discount feature, the retailer needs to tell you how discounts carry over to exchanges.

Your team needs to connect the assistant to the right order and give it a way to pass unanswered questions to the retailer's staff. Agree who will receive those questions and what the customer will see while waiting.

These choices can look less exciting than the conversation in a demo. But without them, the customer gets another explanation and another person to contact.

They also cost time to maintain. If each retailer needs weeks of custom work before your assistant is useful, include that effort when deciding whether you can run the product at a sensible cost.

Suppose you’re building an AI returns assistant for online stores. Start by working with one retailer. Find out where its existing return process leaves customers needing help. If customers can already return an item easily, leave that path alone. Pick a problem they struggle with, then test whether your assistant helps them finish without creating more work for the store’s staff.

With the retailer, follow a few real requests from the first message to the final result. Did the customer finish? Did they have to explain the problem again? Did staff spend less time on the request, or just different time?

Compare that with the store's existing return page and any assistant it already uses. If a simple page handles the job better, the retailer may not need your assistant for that task. If its existing chat works well once it has access to the order, that integration may be enough.

Try the return yourself

The person with the shoes doesn't care whether you call your product vertical AI. They want to know whether they can send the pair back, what it will cost, and what to do next.

As the team building the assistant, walk through those steps with the retailer before choosing how it will speak.

Use a test order and try to return one item through the retailer's current process. Follow the request until the customer gets a decision and the store has a record it can act on. Notice every point where someone has to look elsewhere, repeat themselves or guess what happens next.

Choose a gap your product can actually help with. Agree with the retailer what needs to change, then build and test that part of the return.