How to choose an AI customer service platform: 7 questions that get past the demo

Every customer service tool now says it has AI. That makes the category harder to shop, not easier. The demos look alike: a chat bubble, a fast answer, a happy customer. The differences show up three weeks after launch, when the agent meets your real questions, your edge cases and your busiest Saturday.

This guide is a checklist for getting past the demo. It applies whichever tool you choose.

Start with your conversations, not the feature list

Export or copy your last fifty to one hundred customer conversations. Tag each one with a single label:

• Product: sizes, specs, compatibility, stock, comparisons
• Order: status, address changes, cancellations
• Returns: eligibility, exchanges, refunds
• Advice: anything that needs judgement or a person
• Booking: consultations, fittings, demos

The mix tells you what you're buying. A store where seventy percent of questions are product and order questions needs strong grounding and order lookups. A clinic where most conversations are bookings needs scheduling first. A jeweller needs a clean path to a human and probably a video call.

The seven questions that separate tools

1. Is the AI grounded in your sources?

Ask whether you can restrict answers to approved sources only, and whether you can see which source each answer came from. If the vendor can't show you a citation for an answer in the demo, assume the agent will improvise in production.

Look for a fallback reply, the thing the agent says when it doesn't know, and check that you control its wording.

2. Can it take actions, safely?

Answering "where is my order?" needs an order lookup, not a paragraph about shipping times. Check what the agent can actually do: look up orders, search products with live stock, start an exchange, capture a lead, call your own API.

Then check the safety rails. Anything that changes an order or shares personal data should have a verification step (confirming the customer's email or phone) and a confirmation step before acting.

3. Which channels share one agent and one inbox?

If you sell on WhatsApp or Instagram, ask whether those conversations use the same agent, the same knowledge and the same inbox as website chat. Separate bots per channel mean every policy change is made three times.

4. What does handoff look like?

Ask to see a handoff from the agent's side and the teammate's side. The person who picks up should see:

• the full transcript,
• a short summary of what the customer wants,
• the order or product in question,
• why the agent handed off.

Also check the reverse: can the agent take routine follow-ups back once the person is done?

5. Do you need calls?

Some purchases need a voice or a face. If that's you, check whether calls start from the chat, whether they need a download, whether they can be recorded with consent, and whether there's an AI option, such as a voice agent or video avatar, for out-of-hours callers.

6. How will you know it's working?

Look for conversation-level analytics, not just totals. You want to open any session and see what was asked, what the agent retrieved and where it went. The metrics worth watching are covered in measuring AI agent performance.

7. What will it cost at your volume?

Pricing models differ more than prices do. Some tools charge per seat, some per plan, some per AI resolution, and some by video hours. Model your expected monthly conversations, handoffs and seats on each. Our AI chatbot pricing guide walks through the maths.

Red flags in a demo

• Every answer is perfect, but the vendor won't let you type your own question.
• There's no way to see sources or traces.
• "Handoff" means sending an email to a support address.
• The AI's persona can't be changed, or it claims to be human.
• Pricing depends on a usage metric the vendor can't estimate for you.

Run a one-week test

Shortlist two or three tools and run each on the same high-traffic pages for a week:

1. Connect your FAQ, shipping and returns policies, and product data.
2. Write a persona and a fallback reply.
3. Set handoff rules: ask for a person, refunds over a value, angry sentiment.
4. Read every conversation at the end of the week.

Score three things: questions answered correctly without a person, handoffs that arrived with enough context, and conversations that ended in an order, booking or captured lead. The tool that wins on your own conversations is the right choice.

If LiveCart is on your list, the install guides take a few minutes and our comparison pages say plainly where other tools are stronger.