Nobody owns your help centre, and your AI shows it

If your AI support agent is underperforming, look at your help centre before you look at the model.

An AI agent can answer only from the content it is given. When an article is out of date, a human agent usually recognises the problem and works around it. An AI agent cannot. It relies on the article and gives an incorrect answer with the same confidence as a correct one.

Organisations often spend weeks selecting an AI vendor but give little attention to who maintains the content that vendor’s product depends on.

What the evidence shows

The consequences can be legal as well as reputational. In 2024, Air Canada’s chatbot told a customer that a bereavement discount could be claimed after travel. The airline’s own policy page stated the opposite. British Columbia’s Civil Resolution Tribunal held Air Canada liable and rejected its argument that the chatbot was responsible for its own statements.

Gartner surveyed 5,728 customers and found that only 14% of service issues were fully resolved in self-service. Even for issues customers described as very simple, the figure was 36%. The most common reason self-service failed, in 43% of cases, was that customers could not find content relevant to their problem. This is a help centre problem before it is an AI problem.

The risk also extends beyond the company’s own channels. Gartner’s 2026 research found that customers are about three times more likely to ask a tool such as ChatGPT, Gemini or Copilot about a service issue than to use a company’s own chatbot. If those tools can read your public help centre, an outdated article misleads customers there as well.

Vendor research suggests that content freshness makes a material difference, although I could not confirm the figures at source. HappySupport, a help-centre software provider, reports that help centres updated in the last 30 days reach 45% deflection, compared with 18% for those not audited for six months. It credits HubSpot data but does not link to it, and I could not find the figure in HubSpot’s published reports. Both figures are also measured as deflection, which overstates resolution.

Even if the figures hold, they compare different teams, so they do not prove cause on their own. Teams that keep their help centre current are likely to do other things well too. It is nonetheless reasonable to expect content quality to be a significant factor, as it is the one source the AI consults on every query.

Why nobody owns it

In many organisations the help centre has no clear owner. It was written at launch and is updated mainly in response to customer complaints.

The causes are familiar. Product teams ship changes without informing support until tickets arrive. Agents know the correct answer, but writing an article does not count towards their targets, so it is not done. Content work is deferred whenever the queue is busy, and in most support teams it rarely is.

Before AI, an outdated article generated a few additional tickets. With AI, it produces an incorrect answer for every customer who asks that question, at any hour, until it is corrected.

What ownership looks like

Assign a single named owner. A committee or a shared channel will not keep the content current. In a large team, that person can lead a small rota, but one name should sit at the top.

Protect dedicated time. If content work happens only when the queue is quiet, it will not happen. Allocate a fixed portion of the week in the rota and treat it like any other shift.

Link content updates to product releases. Every release should include a check on which articles it affects. Request a seat in release planning, or at minimum a copy of the release notes before they go live.

Make gaps easy to flag. Agents identify incorrect or missing articles every day. Provide a tag or a form that takes seconds to complete, and ensure someone acts on what is submitted. Customer feedback on articles, such as a rating with an optional comment, is equally valuable and should be prioritised alongside ticket trends and agent flags.

Prioritise reviews by usage. The articles your AI cites most often matter far more than those rarely read. Review the most-used articles first and most frequently, and set a review date for every article, at least annually.

Retire outdated content. An outdated article that remains live is more harmful than no article, because the AI will find and use it.

In the organisations where I have led support, the help centre was either missing, limited to product documentation or used only for onboarding. None of them helped customers troubleshoot an issue. Assigning clear ownership, with a team leader or senior engineer reviewing every article, produced measurable results. At one high-volume B2B software and hardware company, cases fell by 20%. Repeat questions and escalations from partners to senior engineers both decreased, and annual savings, measured against the baseline, ranged from approximately £50,000 to more than £100,000, depending on team size and resource costs. New agents also became productive more quickly, by an estimated 10%.

The most important lesson is to plan for what happens after launch. Agree a lifecycle for articles and a standard approach to maintenance at the start of the project, not after it. If you are buying a tool, make the ability to manage that lifecycle a technical requirement.

AI can also reduce the maintenance effort. Most AI knowledge tools can flag articles that need updating and identify frequent questions in your cases that have no article yet. A general assistant such as Claude or Copilot can also review your contact data and indicate which articles would have the greatest impact.

Where to start

If AI is already in use, pull the list of articles it cited last month. Sort the list by how often each article was used, and add the date each was last updated. Articles near the top with old dates are the first priority. Assign an owner for these articles this week.

If you are still selecting a vendor, include content ownership, and the time it requires, in the business case alongside the licence cost.

Put it into practice: Audit your top 20 help articles in an afternoon