Your Site Already Has the Answers. Nobody Can Find Them

What "an AI chatbot trained on your website" actually means

It means a chatbot that can only answer from your own material. Your pages, product docs, policies and help articles get indexed, and every reply is written from the passages it pulls back rather than from whatever the model happened to absorb during training. Each answer links the page it came from. When a wrong answer shows up, you can trace it to a specific paragraph and fix that paragraph, which is the difference between debugging a system and arguing with one.

Most of the market sells you the software and leaves the hard part with you. We do the opposite. The hard part (working out what your customers actually ask, finding where the answer lives, deciding what the bot is never allowed to say) is the job. The widget is the last afternoon.

The numbers nobody selling chatbots wants to print

Every page you will read today claims a chatbot resolves 80% of your tickets. Here is what the published benchmarks actually say. Median AI self-service deflection sits near 22%, across a range from roughly 8% to 45%, in a synthesis drawn from Zendesk, Intercom, Forrester, Gartner, HubSpot and Salesforce data. First-year business software deployments commonly land at 10% to 15% real deflection.

The variable that moves those numbers most is not the model, and it is not the vendor. It is whether the content underneath is current. HubSpot's State of Service data shows teams whose help centre was updated in the last 30 days reporting around 45% deflection, against 18% for teams that had not audited theirs in six months. Same technology, two and a half times the result, decided entirely by the writing.

We would rather set that expectation now than sell you a number and renegotiate it in month three.

It will not slow down the site you already have

This is the part most agencies skip, and it is the one that quietly costs you money. Published measurements put the heaviest hosted chat widgets at 500 to 750 KB of JavaScript, with lighter ones under 155 KB, and put the average main-thread cost of adding one at 300 to 600 ms. That is comfortably enough to push a passing Interaction to Next Paint score into the failing band, on every page, for every visitor, including the ones who never open the chat.

So we do not ship it that way. What loads on first paint is a button, a few kilobytes of it. The actual chat bundle is fetched when someone clicks, or during a requestIdleCallback once the page has finished the work that matters. Visitors who never chat pay nothing. The measured Lighthouse cost of a widget drops from the 9 to 16 point range to roughly zero when it is deferred this way, which is a decent trade for a splash of extra engineering on our side.

If your site is already fighting for Core Web Vitals, tell us on the call and we will measure before and after so you have the numbers rather than our word for it.

When we will tell you not to build one

Three situations come up often enough to be worth saying out loud before you book a call.

• Your content is thin or out of date. A chatbot over stale documentation produces confident, wrong answers faster than a human ever could. Fix the writing first. We will say so, and we will tell you it is a smaller job than the one you were about to pay for.
• You get under about thirty repeat questions a month. At that volume a well-organised FAQ page and a contact form beat a chatbot on both cost and customer experience. The maths does not work and we would rather tell you than take the deposit.
• The real problem is that a form is broken. Plenty of "we need AI" conversations turn out to be a checkout that fails on mobile or a pricing page nobody can parse. Cheaper to fix the thing. We have sent people away for this.

Everything else, including the awkward middle case where half your knowledge is in one person's head, is worth a thirty-minute conversation.