Check the Front Door You Didn't Build

Picture the AI line in a customer service budget: a chatbot for the website that answers questions about orders, pricing and policies, and a forecast of how many calls it’ll take off the team.

There’s something cheap worth knowing before that line is settled: what the AI assistants your customers might ask instead already say about you. It’s a front door to your company that you didn’t build and never see. It won’t make the decision for you, but it gives you your own evidence to make it with, instead of a survey’s average.

What the surveys settle, and what they don’t

When Gartner surveyed 199 customer service leaders in April and May 2026, their AI spending was up 38% while their overall budgets had grown just 2%, and looking two years out, Gartner reports, respondents predicted GenAI chatbots will become the most valuable service channel.

A separate Gartner survey, of 3,566 B2B and B2C customers in February and March 2026, looked at what customers actually do. When resolving a customer service issue, they were approximately three times more likely to use a third-party AI tool, the ChatGPT kind, than a chatbot the company provided. Use of those outside tools during service interactions has nearly doubled in the past year, while use of company chatbots has stayed statistically unchanged since 2022.

That release gives no share of customers and doesn’t say what they asked, so the finding isn’t that most customers skip your bot. It’s that customers resolving a service issue are more likely to turn to an assistant that isn’t yours. The leaders are betting on their own chatbots, while the customers Gartner asked were likelier to reach for an assistant they already had. Neither result is about your company.

Gartner’s own conclusion is blunt. That release told service leaders, in its subheading, they “should rethink standalone chatbot investments,” and Eric Keller, a senior director analyst at Gartner, said many company chatbots are still designed mainly to answer questions while customers increasingly expect AI to help them take action, “such as booking an appointment, submitting documents or updating their account.”

What no survey says is which of your customers’ questions still need a bot of yours. Ask the assistants yourself and the answers separate: some come back good enough, some come back wrong, some come back too vague to be any use, and some come back as general advice, because the real answer needs a customer’s own account or a change in your systems.

That last group is what a budget line is really about, and it isn’t a contest with ChatGPT. Those customers come to you anyway, by phone, email or your app, because nobody outside can see their account. The mix differs in every company, which is why it’s worth an afternoon before the money is committed.

Expect it to spill past customer service, since what it costs and how you compare with a competitor are pre-sale questions, and some of what comes back will land on marketing’s pages.

What the outside assistant actually does

When someone asks ChatGPT about your return policy, it doesn’t check with you. It answers either from what its model already learned, which can be out of date, or from a search of the web. OpenAI’s documentation says ChatGPT may search on its own when a question needs current information, typically rewrites the question into search queries of its own, and may include links to the pages it used. OpenAI also says plainly that “search results and citations can be incomplete, outdated, or incorrect.” And a wrong answer can arrive in the same confident, finished tone as a right one. Nobody at your company sees that conversation. What you can do is look.

A check you can run in an afternoon

1. Get two answers before you start. Ask whoever runs your website whether your robots.txt file, or your hosting, blocks either OpenAI’s search crawler, OAI-SearchBot, or Google’s Google-Extended. OpenAI says a site’s content is only included in ChatGPT’s search summaries if it isn’t blocking that crawler. Google describes Google-Extended as how publishers manage whether crawled content is used to train Gemini and to ground its answers. A yes explains a run of thin or missing answers from that assistant, and only that one. If either is blocked for a reason nobody can name, unblocking it and coming back in a month is cheaper than reading answers that never saw your pages.

2. Collect the real questions. Ask the people who answer customers every day for the questions they hear most, and pick ten in the customers’ own words, since that’s how they’ll be typed. Ten is enough to show a pattern without turning the check into a project. Keep the mix roughly as your team hears it, since the piles at the end will only be as honest as this list, and it’s a list of your most common questions rather than a proportional sample of your queue. Cover both kinds: general questions anyone could ask (what it costs, what the return policy is, how you compare with a named competitor) and questions about a customer’s own account (why was I charged twice, can you move my delivery).

3. Ask all ten the way a customer would. Type each one in plain words into ChatGPT and into Google’s Gemini, and paste every answer into one dated document. Use a browser where you’re not signed in to ChatGPT, since its web search works signed out, so your own saved memories don’t shape what comes back. Do the same in Gemini if it lets you. ChatGPT can still take a person’s location into account, and Gemini’s answers may differ from person to person too, so treat what you see as a sample rather than a transcript.

4. Sort the questions into four piles. Two assistants means two answers each, and they’ll often differ, so sort the question rather than the answers, in this order. If neither could answer it without seeing the customer’s account, it belongs in the fourth pile. Otherwise, if either answer was wrong, the question is Wrong, since one wrong answer is enough to reach a customer. If nothing was wrong but nothing was usable either, it’s Vague or missing. What’s left is Right.

• Right. At least one assistant got it, and neither got it wrong. Nothing to fix today, and it stays on the list for the next check.
• Wrong. Look for the right answer on your own website. If it isn’t stated clearly on a public page, that’s page work. If it is, open any links under the assistant’s answer to see where the wrong version came from: an old page of yours, which can be updated or taken down, or someone else’s, which can’t, leaving your page to carry the current answer and the date it was updated. And if there are no links at all, on a question your page answers clearly, from an assistant you aren’t blocking, the answer probably came from what the model learned earlier. Note it and look again next time.
• Vague or missing. If you’re blocking one of the assistants, start with step 1 rather than here, since a thin answer may be about access rather than your pages. Otherwise, check whether the answer is on your site at all. If it only lives in a price sheet your team emails, a proposal template or someone’s head, no outside assistant can quote it, so decide whether it should be public. If it’s already on a clear public page and you aren’t blocking that assistant, there’s nothing to fix today: note it and look again next time. A rough test: if you’d say it to a prospect on a first call, it can go on a page. If it’s negotiated customer by customer, it can’t.
• Needs the customer’s account. The assistant offers general advice rather than an answer, because the real one needs this customer’s details or a change in your systems. A double charge usually lands here. “Can I change my plan” may not, since help pages can explain how.

What the piles tell you

The piles point in two directions. Wrong and missing answers point at your website, apart from the ones where nothing on your side is at fault, and the fourth pile points at your own systems. The website work and the systems work are separate, so neither has to wait for the other.

The page work is the wrong answers your pages could have prevented, plus the missing ones you decide should be public. Your pages are the one source behind those answers that you control. Nobody can promise when, or whether, an assistant picks up a change, which is worth knowing before it’s sold to you as a quick win.

The fourth pile is where your own AI would earn its money. For now, an assistant working from your public pages can’t see a customer’s account or change anything in your systems, so general advice is the most it can offer there. That leaves two reasonable readings of the pile.

One is the narrow bot: scope it to those questions and leave the rest to your pages. The other is Gartner’s, which goes further. Keller said leaders should “redesign digital support around conversational, action-oriented experiences, rather than treating GenAI as a standalone chatbot,” across digital and voice channels. That’s a bigger programme than a bot on your website. The check won’t choose between them, but it tells you which of your customers’ questions sit in that pile, which is where both readings start.

Either way, Keller’s advice in a later Gartner Q&A is worth borrowing: start with targeted, tested deployments on the issue types it handles reliably, whatever you end up funding, and expand only once it performs, and give customers a visible path to a person. In practice, hand the fourth pile to whoever would build or buy it, ask which of those questions it can reliably resolve, start there, and look at how many of those conversations ended without a person stepping in before anyone widens the scope. And any of it is something someone has to keep right after it launches.

Where this leaves the budget line

Ten questions on one day won’t give you a percentage, and they don’t need to. One wrong answer is enough to know a wrong answer is possible, a pile that’s mostly wrong is a pattern, and anything in between is a reason to run the ten again next month, and again whenever a price, a policy or a product changes.

What moves the number is the comparison between the piles, not a threshold worth inventing. Read the comparison as a direction rather than a measurement, since the ten were picked rather than counted, and ask your helpdesk for ticket counts by topic if you want it firmer. If the questions needing a customer’s account are the bigger group, the case for building something looks stronger than the case for writing, and the check has named where it would start. If the page-work piles are bigger, the writing comes first and the build can wait for the next check. If almost everything comes back right, neither job is urgent, and you have a date to look again.

None of that decides the budget line for you, and the surveys behind it don’t either. What the check gives you is the one thing the surveys can’t: what the assistants your customers use are saying about your company this week, and which of those answers are yours to change.