Chatbot projects usually get judged on the model. Which provider, which version, how well it understands phrasing. That is almost never why they underperform.
The bots that fail commercially generally understand the visitor fine. They fail because they were built to reduce support load and are being measured on revenue, and those two goals pull in opposite directions.
Deflection and conversion are opposing goals
A deflection bot succeeds when the visitor goes away satisfied without contacting anyone. A conversion bot succeeds when the visitor takes a commercial action.
Most bots are specified by whoever owns support cost and evaluated by whoever owns revenue. The design optimises for containment — answer the question, avoid the handoff, close the conversation — and containment is precisely what you do not want when the person asking is a buyer.
A visitor asking "do you handle sites on Shopify" is not raising a support ticket. They are qualifying you. Answering "yes, we support Shopify" and closing the conversation is a successful deflection and a lost enquiry.
The fix is not a better answer. It is recognising commercial intent and changing what the bot does next.
The four faults
1. It cannot complete anything
The most common and the most expensive.
A bot that answers questions but cannot book, quote, check an order or start an application has added a step. The visitor asks, receives an answer, and still has to go and do the thing. Every genuine buyer ends up in the same queue they were trying to skip, having spent two minutes first.
This is not a model problem. It is an access problem: the bot has not been connected to the calendar, the order system, the CRM. Those integrations are unglamorous and they are the entire difference between a bot that converts and one that chats.
The test: can a visitor finish something without leaving the conversation? If not, the bot is a search interface with extra latency.
2. The handoff loses everything
The second most damaging, and the most fixable.
A visitor spends four messages explaining their situation. The bot cannot help. They are passed to a human — who opens with "hi, how can I help?"
Everything just collected is discarded, and the visitor now has to decide whether to type it all again or leave. Many leave, and they leave more annoyed than if the bot had never existed, because the effort was wasted.
A handoff that carries the full transcript, the identified intent and any collected details converts better than a materially smarter bot that hands over nothing. This is a plumbing problem with a large commercial return, and it is usually skipped because it is not the interesting part of the build.
3. It is placed where nobody has a question
Bots get deployed site-wide by default, which means they appear where nobody needs them and behave identically where they are urgently needed.
- Homepage: visitors are orienting. A bot popping up interrupts
- Pricing or service page: high intent, specific questions. This is where it earns
- Checkout or enquiry form: friction point. Real value, if it can actually resolve
- Blog or resource pages: reading, not buying. Interruption costs the read
Placement is a bigger lever than prompt engineering, and it costs nothing to change.
4. It is measured on the wrong thing
Conversations handled, deflection rate, satisfaction with the bot.
Every one of those can be improved by making it harder to reach a human. The dashboard improves while revenue falls, and because the metric is going the right way, nobody investigates.
Measure completed outcomes: bookings made, quotes issued, orders located, qualified handoffs accepted by sales. Those cannot be gamed by containment, and they are the numbers the business actually cares about.
Connecting a bot to the systems that let it finish something is most of the work and almost none of the demo, which is what
assistants built to complete actions rather than deflect them.The strongest counter-argument
There is a real case against everything above: for many businesses, deflection is the correct goal.
A company drowning in repetitive "where is my order" enquiries genuinely benefits from a bot that answers them and nothing else. Support cost falls, response times improve for the enquiries that need a person, and no revenue is lost because those conversations were never going to produce any.
That is a legitimate and often excellent use, and I do not want to argue against it.
The failure is not deflection. It is deflection deployed without segmentation, then judged on conversion. A bot that treats "where is my order" and "do you work with companies like mine" identically will handle the first well and lose the second — and the second is worth far more.
If the bot cannot tell the difference between a service enquiry and a sales enquiry, it will optimise for whichever is more common. That is almost always support.
What to build instead
For most sites, in this order:
- Answer the common questions on the page itself. A clear pricing page and a real FAQ remove more friction than any bot, cost nothing to run, and rank in search — which a chat widget never will
- Make the primary action obvious and fast. Many bots exist to help people find a booking link that should have been unmissable
- Then, if there is genuine volume, build an assistant that can complete the two or three most common actions and hand off cleanly on everything else
Step three is worth doing only after one and two, because a bot in front of an unclear proposition makes the confusion interactive rather than resolving it.
If you already have one
Four checks, in order of expected return:
| Check | Question |
|---|---|
| Completion | Can a visitor finish anything without leaving the conversation? |
| Handoff | Does the human receive the transcript and the intent, or start from nothing? |
| Segmentation | Does it behave differently for a sales question than a support one? |
| Measurement | Is it judged on outcomes, or on conversations handled? |
Most bots fail the first two. Fixing those requires integration work rather than a model change, which is why it tends not to happen — and why the conclusion is so often "AI did not work for us" when AI was never the constraint.
Keep reading
This argument connects to a few other things worth reading before you make a decision.