Why We Don't Let Our AI Quote Prices

There is a tempting feature every business wants the moment it adds an AI chatbot: let it quote prices. A customer asks "how much for a booking website?" and the bot fires back a number, day or night, no staff involved. It sounds like the whole point of automation. It is also the fastest way to turn a helpful assistant into a liability, and it is a line we deliberately do not cross in the chatbot we built to qualify leads without inventing prices. This is the narrow reason why, and how we solved it.

A language model is built to sound right, not to be right about your prices

This is the part most people miss. A language model does not look up your price and read it back. It predicts fluent, plausible text one word at a time. Asked for a number, it will produce one that sounds reasonable in context, because sounding reasonable is exactly what it is trained to do. Whether that number matches your actual pricing is, from the model's point of view, beside the point.

So it will answer. It will almost never say "I don't know." It will give you a crisp, confident figure delivered in the same calm tone it uses for everything, and there is no visible difference between the times it happens to be right and the times it is guessing. That confidence is the trap. A hesitant wrong answer you would catch. A polished wrong answer walks straight out the door to your customer.

A wrong price is not a small mistake

If a chatbot gives a slightly wrong opening hour, you correct it and everyone moves on. A wrong price is different, because a price is a near-commitment. The moment a customer sees a figure attached to your business, they anchor to it. Every wrong quote lands as one of three problems, and none of them is minor.

• The undercut. The bot quotes low. Now you either honor a number that loses you money or you tell a customer the real price is higher, which reads as a bait and switch even though nobody intended one.
• The overquote. The bot quotes high. The customer decides you are expensive and messages a competitor instead. You never even hear about the sale you lost, so you cannot fix it.
• The contradiction. The bot quotes a figure that clashes with your website, your proposal, or what your sales person said yesterday. Now your own tools disagree in front of the customer, and every number you give afterward is suspect.

That last one is the quiet killer. Trust in a business is built on consistency, and the fastest way to lose it is to be caught contradicting yourself. A chatbot that improvises prices is a machine for manufacturing those contradictions, at scale, while you sleep.

The fix: separate the talking from the quoting

The mistake is asking one system to do two very different jobs. Talking to a customer, understanding what they want, and asking good follow-up questions is exactly what a language model is good at. Committing to a number is not a language task at all. It is a business decision that has to follow your real rules every time, not a plausible-sounding guess.

So we split them. Our chatbot is allowed to be the brilliant conversationalist. It reads the request, works out whether the customer needs a simple site or an e-commerce build, notices what detail is still missing, and asks for it in plain language. What it is never allowed to do is invent the figure. When the conversation reaches the point of a number, it hands off.

• The model listens and gathers. It clarifies scope, pulls out the details that actually change the price, and makes sure the customer feels heard rather than interrogated.
• A separate, deterministic step owns the number. The actual pricing logic runs on its own, follows the exact same rules for every customer, and cannot be talked into a discount by a clever prompt. This logic stays server-side, and we do not expose the figures or the formula behind it.
• A person owns anything real logic should not decide. Where a project genuinely needs judgment, the bot routes the customer to a proper guided estimate or a human, rather than papering over the gap with a made-up number.

From the customer's side, none of this feels like a handoff. They get a smooth conversation that ends in a real, defensible next step, either an accurate estimate or a route to someone who can give one. What they never get is a number the business would have to walk back.

What this buys you

This is not caution for its own sake. Drawing the line here is what makes the automation safe to leave running unattended, which is the whole reason to have it.

• No rogue discounts. The bot cannot be flattered, argued, or prompt-engineered into a price you never authorized.
• No invented figures. Every number a customer sees traces back to real logic or a real person, so it is one you can stand behind.
• No awkward walk-backs. You are never in the position of correcting your own chatbot in front of a paying customer.
• A pricing conversation you actually control. Your rules live in one place and change in one place, instead of drifting inside a model's guesses.

It also means the chatbot does the thing it is good at: capturing intent, qualifying the enquiry, and handing your team a warm, well-understood lead. It does the talking. Your pricing does the quoting. Neither one pretends to be the other.

The short version

A language model asked for a price will answer, confidently, whether or not it has any business doing so. Since you cannot tell the right guesses from the wrong ones without checking every answer, the only safe design is to keep the model out of the number entirely. Let it converse. Route the figure to logic or a human. That single boundary is the difference between an AI that helps you close and one that costs you deals. For the full picture of how the whole assistant is put together around that boundary, read how we built a chatbot that qualifies leads without inventing prices.