Should your business use AI for customer support? An honest assessment

AI is worth deploying for customer support when a large share of your questions are the same questions, the answers are documented, and being wrong occasionally is survivable. It is a poor idea when your queries are genuinely varied, your answers live in people’s heads, or a wrong answer costs real money.

Most businesses can work out which they are in an afternoon, and almost none do before buying.

How do you tell which one you are?

Take your last two hundred support conversations. Sort them into groups by what was actually being asked, not by how they arrived.

You are looking for the shape of the distribution. If a small number of question types account for most of the volume, AI has something to work with. If two hundred conversations produce a hundred and fifty distinct problems, it does not — you have a complex support function, and a bot will handle the easy tail while your people keep doing the hard part.

Then ask a second question about the top groups: is the answer written down anywhere? If the answer to “why is my order delayed” requires checking three systems and using judgement, that is not a documentation gap — it is a genuinely hard task, and automating it is a much larger project than a chatbot.

Where does it work well?

Works well	Why
Product specifications and availability	Factual, documented, verifiable
Order and dispatch status	Lookup from a system, not judgement
Policy questions — returns, warranty, delivery	Written down, stable, repetitive
Routing to the right team	Classification, which models do well
Out-of-hours acknowledgement	Something beats nothing at 11pm
First-language responses	Language handling is a genuine strength

Where does it go badly?

Anything involving money or commitment. Quoting a price, confirming a discount, agreeing a delivery date. If a customer can reasonably hold you to what the bot said, a person should say it.

Complaints. An angry customer wants to be heard by someone who can do something. A bot asking clarifying questions escalates the situation rather than resolving it.

Technical judgement. “Which adhesive for this substrate” looks like a lookup and is not — it depends on conditions the customer has not mentioned and does not know are relevant. A confident wrong answer here produces a failed installation.

Anything where being wrong is expensive. The question is not how often it is wrong. It is what happens when it is.

What does it actually save?

Less than the marketing suggests, and in a different form than expected.

Deflection rates quoted by vendors are not comparable to your business — different question mixes, different definitions of resolution, and figures produced by people selling something. Ignore them and measure your own.

What tends to be real:

• Coverage outside working hours. Often the biggest genuine gain, because the alternative is nothing.
• Peaks absorbed. Launches and campaigns stop overwhelming the team.
• The dull questions removed, which improves retention among support staff more than it reduces headcount.
• Faster first response, which affects satisfaction independently of resolution.

What is usually not real is a headcount reduction. The volume that gets deflected is the easy volume, and what remains is harder and slower per conversation. Teams get better work, not smaller.

There is also a new cost from 1 October 2026 if you run this on WhatsApp: messages inside the customer service window become chargeable. A verbose bot is now a direct expense — see the billing change.

What has to be true before you start?

1. Your common answers are written down, or someone is going to write them. This is the project, more than the software is.
2. There is a human to hand over to, during hours customers are actually awake.
3. Someone owns it — reviewing unanswered questions monthly and keeping the knowledge current.
4. You have decided what it must not answer, and it refuses those firmly.
5. You can measure it — resolution without handover, messages per conversation, and complaints about the bot itself.

If the first point is not true, start there. A business that documents its top twenty answers properly gets a real benefit even if the AI project never happens — and the AI project is far more likely to work afterwards.

Common questions

Will customers object to talking to a bot?

They object to being trapped with one. A bot that answers quickly and hands over on request is generally accepted. One that loops, deflects and hides the route to a person generates more anger than the original problem.

Should we tell customers it is AI?

Yes. They work it out within two messages, and having pretended costs you more than disclosing would have. It also sets expectations, which makes them more tolerant of limitations.

Can we start small?

You should. One question category, answered reliably, is worth more than broad coverage that is often wrong — and it lets you measure something real before widening scope.

What about personal data in support conversations?

It is being processed, so it falls under your data protection obligations like anything else. Know which provider handles it, where, whether it is retained and whether it could be used for training. Get that in the contract rather than the sales conversation.

More on this: when a language model is the wrong tool and AI on WhatsApp. See also AI chatbots.