AI chat, live chat, or a help centre: which answers what

These three are usually compared on features, which is the least useful axis. They differ in what a marginal answer costs, what it costs to create the answer in the first place, and which questions they are physically capable of handling. Those three facts decide the mix.

The help-centre illusion

A documentation page is the cheapest channel per visit and the most expensive per answer, because almost nobody finds it. Gartner (2024) surveyed 5,728 customers and found only 14% of issues fully resolved in self-service — and that in 43% of the failures, the customer could not find content relevant to their problem at all. Even for issues customers themselves described as very simple, the figure only reached 36%.

So the honest model of a help centre is not "this covers 60% of our questions". It is "this contains the answer to 60% of our questions, and resolves a seventh of them". The rest leak back into an inbox.

The middle slider is the point of the figure. Writing more documentation is good and you should do it — but as a deflection strategy on its own it is close to a rounding error, because the failure is discovery rather than coverage. That is the gap an assistant fills: it is mostly reading the pages you already wrote, to the people who could not find them.

A help centre answers the question your customer knew how to search for. That is a much smaller set than the questions they have.

What each one can physically do

One row deserves more than a table cell. Only a person notices they are wrong. A help centre is silently out of date and an assistant answers a question about a policy you changed last week with total confidence. This is the strongest argument for keeping a human in the loop and for making the handoff easy to reach, rather than something the assistant offers only after three failed attempts.

Speed is the real trade

The cost argument is usually overstated and the speed argument understated. Gorgias's current benchmarks across a thousand-odd ecommerce brands put the all-industry median first response at 6.3 hours (2026), with brands automating 30% or more replying roughly ten times faster. It is their own platform data and their own commercial interest, so treat the magnitude sceptically — but the direction is not in doubt, and an answer at 2am is worth more than a cheaper answer at 2pm tomorrow.

The counterweight is real too. In a randomised trial over 680,676 chats, conversations the AI finished alone were 64.6% shorter and rated 0.858 points lower out of five (2026). Faster, and more irritating. Which is a trade you can make deliberately — the measurement guide has a figure for exactly where that line sits.

What to do first

Start by reading your last hundred tickets and sorting them into those three piles. It takes an hour and it is more informative than any vendor's calculator, including ours. The pile in the middle is what an assistant is for; if it is small, your problem is documentation, and if it is large, you have been paying people to do retrieval by hand.

Then measure rather than assume. We take the same view of support channels that Zinx Signal takes of landing pages — its post on measuring revenue rather than clicks makes the same argument in a different domain: the metric that is easy to collect is rarely the one that pays for anything. What you train it on decides how big that middle pile can get.