Summarize with AI:ChatGPTPerplexityGoogle AIClaude

Quick Answer

A customer support website needs a searchable knowledge base, visible contact options, live chat, ticket submission with status, and an AI assistant that answers from your content and hands off to a person.

The handoff decides whether the rest works. An assistant such as hue replies in live chat and to new tickets from the knowledge you approved. When it is not sure, it says so. Then it passes the customer to a person with the full conversation attached, so nobody starts over.

Did this answer your question?
Three things people believe. Myth or fact?
Call each one, then see how other readers called it.
1 A help center for a county or a tax authority needs the same visible route to a person as a help center for an online store.
2 An AI assistant removes the need to keep help center articles current.
3 Even with an AI assistant live, the site still needs staffed chat, with a person one click away.

The first test for any support site is plain: can a visitor with a common question find the answer, or a person, from the page where they land? That gap is the subject of this article. It is not a gap in effort. It is a gap in content, in routing, and in how visibly a person stands behind the page.

I set out what the site needs, part by part: a searchable knowledge base, contact options a visitor can see without hunting, live chat, ticket submission with a status the customer can check, and AI self-service that knows its limits. Each part is judged the same way. Does it end the visitor's problem, or move it cleanly to someone who can?

The AI part deserves the most care, because trust is won or lost there quickly. hue, the AI assistant discussed later, answers customers in live chat, replies to new tickets (including email that arrives as tickets) and responds to phone calls, which remain a beta feature, all from the knowledge you approved. If nobody is free to staff the chat behind it, HelpSquad, LiveHelpNow's sister company, staffs live chat with trained agents from $185 a month, live in about 48 hours, with the chat software included.

Data handling is the second test buyers apply, and rightly so. Your data is used to answer your questions and is not used to train AI unless you choose to allow it. By default, each client's data answers only that client's own questions.

Service leaders, for their part, are not planning to empty the queue of people. So the design question is never AI or people. It is how a visitor moves from one to the other without starting again, and that depends first on what the help center actually says.

A customer support website succeeds when a visitor can find the answer alone or reach a person in one step. Everything else is secondary. I would judge any help center against those two outcomes before looking at layout, widgets or vendor logos.

The difficulty is that much support content was written for search engines or for agents, not for the visitor with a stuck order or a locked account. An AI assistant does not repair that. The fair objection is that an assistant will make things up, and the safeguard is specific: it should answer only from connected systems and approved documents, show its source and say when it does not know.

So the practical question for any buyer is which pages the assistant reads, and who checks them. At LiveHelpNow, the website crawl that feeds the AI assistant starts from the sitemap, reads English pages by default, and applies priority rules by page section along with exclusion patterns. It runs on a schedule or on demand. The crawled pages, and the questions and answers drawn from them, can be browsed in the admin, so a support lead can check what the assistant knows before a customer ever asks.

Price is the other early question. My recommendation is to compare the cost of an AI assistant with the cost of one more hire, to buy from a provider that publishes its prices on its own page, and to read the fine print on what counts as a resolution. That definition matters, but it matters less than the pages the AI reads, which is where I begin.

Two support agents wearing headsets look at a computer monitor together in a bright office.
Support agents with headsets work side by side, ready to step in when a customer needs a real person.

Want an AI assistant that answers from your content and still hands off to a person?

hue, our AI assistant, answers chats and new tickets, and phone calls in beta, from knowledge you approved; when unsure, it hands the customer to a person with the full conversation.

A claim check runs after each answer and can replace any reply its sources do not support. That matters. Gartner's 2026 survey found 87% of customers say companies using GenAI in service must offer a route to a human agent, so a person is always one click away. Plans start at $39 per seat a month with AI agent assist included, and customer-facing AI starts at $99 a month.

What should a customer support website include?

A customer support website needs six parts: a searchable knowledge base, clear contact options, live chat, ticket submission with status, AI self-service, and a visible route to a person.

Before choosing any tool, I would put four plain questions to the current site, in this order:

  1. Can a visitor search for an answer from the first page they land on?
  2. Can the visitor see, without scrolling, how to reach a person?
  3. Can a customer submit a request once and later check its status without calling?
  4. Do the chat widget and the agent logins protect personal data by default?

Cost is the reason these questions matter now. APQC's benchmark across 1,330 companies puts the median cost of an in-app chat contact at $4.00, and that is a channel where the customer is normally already signed in. Every answer the site fails to surface becomes a contact that someone must staff and pay for. Small leaks add up.

ComponentWhat it does for the visitorWhat goes wrong without it
Searchable knowledge baseAnswers common questions at any hourSimple questions arrive as chats and tickets
Clear contact optionsShows chat, email and phone routes in one placeVisitors hunt for a number, then give up
Live chatReaches a person while the visitor is still on the pageQuestions wait in an inbox for hours
Ticket submission and statusCaptures details once and shows progressCustomers call to ask where their request stands
AI self-serviceAnswers from approved content after hoursNo help at night, or a bot that guesses
Visible route to a personLets the visitor escalate in one stepFrustration builds behind a closed loop

The common assumption is that a support website is a contact page with a form attached. I see it differently. It is a resolution system, and the controls underneath the layout matter as much as the layout itself.

Security: In our own platform, agents sign in with two-factor authentication by a code sent through email or SMS, and a trusted device is remembered for 30 days. Personal data is encrypted at rest with AES-256, and the chat widget applies PII masking. Those are the defaults I would expect any vendor to show on request.

Durability: The technology behind LiveHelpNow is covered by US patents 9,178,950 and 9,584,375, which matters less for marketing than for continuity. A support site is a long-lived asset. The platform under it should be built to last.

A review of 4 sources behind this section points the same way: cost, findability and data protection are decided by structure, not by adding more pages. If you are starting with the content layer, our knowledge base software for support teams shows how articles, search and chat share one source of answers.

The harder question is why so many help centers already have all six parts and still leave most visitors without a resolution.

Why do so few customers solve their problem on a support website?

Nearly three in four customers try self-service at some point. When Gartner counted how many of them left with the problem actually solved, the number was far smaller.

Gartner surveyed 5,728 customers in December 2023 and published the results in August 2024. Of those customers, 73% used self-service at some point, yet only 14% of issues were fully resolved there. Even issues customers called "very simple" were fully resolved in self-service only 36% of the time. "It's concerning to see that so few fully resolve there," said Eric Keller, a senior director of research in Gartner's customer service and support practice.

Customers who use self-service at some point 73%
"Very simple" issues fully resolved in self-service 36%
All issues fully resolved in self-service 14%
Most people try it, and few finish there. The top bar is a share of customers. The other two are shares of issues. Gartner survey of 5,728 customers, December 2023, published August 2024.

So where does self-service break? Mostly at finding things. The most common reason for failure, in 43% of cases, was that customers "couldn't find content relevant to their issue." Separately, 45% of customers who started in self-service said the company "didn't understand what they were trying to do." Keller described journeys that feel "too rigid to deal with the complexities of their service issues." More articles alone fix neither complaint: the customer still has to land on the right one, and the website still has to grasp what the customer means.

Many support teams now hand that job to an AI assistant. George Dilthey runs a 16-person support team at Clay, a data enrichment company, where 90% of customers who come through support talk to AI first. In a January 2026 interview, he called the AI the smaller part of the problem.

"Every AI support tool is using the same underlying models. Nobody's got a huge advantage when it comes to the actual AI. Most of it depends on how good your content is."

George Dilthey, head of support at Clay, 2026

We build an AI assistant that answers from a company's own knowledge base, so we have a stake here, though Clay uses another vendor's tool. "Historically, knowledge bases have always been last priority," Dilthey said. One would "go out of date" until someone overhauled it every few months or years. Now he calls the knowledge base "the measure of intelligence for your custom AI." Gartner reached a similar recommendation: let support reps "create knowledge as part of the issue resolution workflow, rather than as a separate process." An assistant built on stale articles gives stale answers, and gives them confidently.

Dilthey's team also decides which conversations the AI never handles. Account deletion requests go to Clay's privacy team, and refund requests skip the AI. "That's just a bad experience," he said. He passed on one AI-native vendor partly because it felt like "a bit of a black hole," with "very little control over processes." Customers want that exit too. In a Gartner survey of 3,566 customers in February and March 2026, 87% said it is essential for companies using GenAI in customer service to offer an option to reach a human agent. The handoff must also carry the conversation. A former customer service rep, writing about phone support in 2012, described being "volleyed from one department to another" and asked to repeat details already given: one person's account, but a clear picture of a handoff without context.

How a support website can close the gap

  1. One prominent place to start: search, or an AI assistant that says it is an AI. Gartner recommends a "single digital concierge" as the main entry point.
  2. Answers come from a knowledge base the support team keeps current while it resolves tickets.
  3. Sensitive requests, such as refunds or account deletion at Clay, go straight to a person.
  4. Anything the pages cannot settle moves to a person in one step, along with the conversation so far.

Drawn from Gartner's 2024 recommendations and Clay's routing as described by its head of support.

Put the Gartner numbers next to Dilthey's experience and the same weak spot shows up: the distance between a customer's question and the right answer, or the right person. Search that misreads intent lengthens it. Stale articles lengthen it, and an AI assistant repeats each one. A hidden route to a person lengthens it for exactly the cases no page can settle. Self-service already gets plenty of use. What it lacks is a path that finishes.

  • Track whether visitors who use your help pages still open a ticket or chat about the same issue; page views and searches alone hide the gap Gartner measured.
  • Search your own help center using the words customers write in their tickets, and list the common issues where nothing relevant comes up.
  • Make writing or fixing an article part of closing a ticket, as Gartner recommends, instead of leaving it to a quarterly cleanup.
  • Before switching on an AI assistant, ask which content it will answer from, how out-of-date articles get flagged, and what it says when it does not know.
  • Write down which requests should skip the assistant entirely, the way Clay routes refunds and account deletion, and check that the handoff to a person keeps the conversation.

How we checked this

We used two Gartner press releases summarizing customer surveys, a January 2026 interview with Clay's head of support, and a 2012 blog post by a former customer service rep. None of the figures are ours. Gartner's resolution survey was fielded in December 2023, so it cannot show how much newer AI assistants have changed resolution since. The press release summarizes a longer client report we did not read. Clay is one company, and Dilthey's views reflect his own team's experience. The 2012 post describes one person's phone support experience. We sell customer service software with an AI assistant, so we have a commercial interest in how help centers and handoffs are built. Still unknown: whether well-maintained knowledge bases with AI assistants raise the full-resolution rate across companies. None of our sources measured that.

  1. Gartner, survey on self-service resolution, August 19, 2024.
  2. Gartner, survey on access to a human agent, August 4, 2026.
  3. Yash Tekriwal, interview with Clay's head of support, Know Your Growth, January 12, 2026.
  4. Budget Savvy Diva, a former rep on customer service fails, April 22, 2012.

Why do new support tools so often fail to fix the help center?

New support tools fail when their content is hard to use and data rules come last; a healthcare team, for instance, needs a HIPAA business associate agreement before chat launches.

It is tempting to assume the next tool will close the gap. The evidence suggests the gap sits elsewhere. In its Tech Tide report on contact centers for Q4 2024, Forrester reviewed 20 customer service technologies and found that knowledge management, agent augmentation and automated quality monitoring had been on the market for years without generating their expected value, because they were too hard to use. Generative AI, through summarization and personalization, gave those categories new life. Forrester also observed that newer tools such as conversational AI often looked promising in pilots and then disappointed at scale, because organizations paid too little attention to adoption, change management and governance.

Software alone rarely fixes a help center. I see three tensions behind that finding.

Usability: A knowledge base that agents find awkward to update goes stale. Stale articles produce wrong answers, whether a person or an AI assistant reads them aloud. The content layer has to be easy to maintain, or it decays quietly.

Governance: A pilot runs on enthusiasm. Production runs on ownership, review cycles and clear rules for who approves an answer. Without those, the same tool that impressed the steering committee drifts within a quarter.

Data retention: AI self-service needs content, yet support conversations carry personal data that regulated teams must not keep longer than necessary. In our platform, each account sets an automatic data purge that runs anywhere from immediately after an interaction to 30 days. Healthcare customers are offered a HIPAA business associate agreement, so the platform is HIPAA-ready with a BAA rather than a liability waiting for an audit.

Fragmentation makes all three worse. When the public help center, the agent macros and the management reports each draw on a different source, they disagree, and customers notice. Our approach is to let one knowledge base serve three audiences: customers asking about their own accounts, agents who receive suggested replies and next steps, and managers who ask about the business in plain English. One source means one correction fixes every channel. You can see how that works in our AI customer support software, where the same approved content drives the assistant and the agent view.

The practical lesson is uncomfortable. A support website is an operating discipline, not a purchase. A team that budgets for the license but not for content ownership is likely to end up shopping again.

And when the tools do fail, customers feel it as the oldest complaint in service: being passed from one place to another and asked to repeat everything they already said.

How do you keep the route to a person visible and the data safe?

Show a one-step route to a person on every page, pass the conversation along, and secure agent logins with SAML single sign-on through Microsoft Entra ID, Google Workspace or Okta.

The complaints have not changed much in a decade. A 2012 post on the Budget Savvy Diva blog, written by a former customer service representative, listed five service fails, led by hold and transfers and by phone trees. The author described being passed between departments and asked to repeat details already given. At one credit union, pressing 0 did not reach an operator; callers needed a special code. The biggest problem, in the author's words, was "the inability to fix a problem."

A website can rebuild that credit union phone tree without anyone noticing. A chatbot with no exit is the same trap. So is a contact page that hides the phone number three clicks deep.

I would resolve it with three changes, in this order.

Visibility: Place the route to a person on every help article, the chat window and the ticket form. Text messaging, email and phone belong on one contact page, each with honest expectations about response time. Nobody should need a code.

Continuity: When an AI assistant hands off, the person who picks up should receive the conversation summary. The customer should never retell the story. That single design choice removes the repeat-yourself complaint the 2012 post described.

Trust: Visitors share account details only when they believe the site will protect them. Buyers in government and healthcare ask the hardest questions here. The security options agencies ask us for are consistent: SAML SSO, two-factor authentication, PII encryption, IP allow-listing, scheduled data purge and a HIPAA BAA. On county and city government websites, the visitors are residents and taxpayers, and procurement teams check every one of those items before a chat widget goes live.

Training is the other trust question. We use customer data to train AI in only 1 case, when the customer opts in. Otherwise, your data is used to answer your questions and is not used to train AI unless you choose to allow it. Our customer website privacy policy sets out how visitor data on a client site is handled.

In summary, visibility gets the visitor to a person, continuity spares them from repeating themselves, and trust makes them willing to share what the agent needs. Each one is a design decision, not a staffing decision.

Trust brings the visitor as far as the chat window. What they find before they get there depends on how the knowledge base is built.

What should support leaders do with their help center next?

Build the help center and the human route together, because the evidence points toward AI that answers routine questions while trained people stay in place for everything the pages cannot settle.

This forecast is not mine alone. In a March 2025 poll of 163 service leaders, Gartner found that 95% planned to retain human agents to define the role of AI. It also predicted that by 2027, 50% of organizations that had expected to cut their service workforce significantly would abandon those plans. Staffing is not going away. So I would ask one question of every help page: where does a visitor go when this page does not answer?

A brief aside. I have little patience for any setup that hides the contact button to make self-service figures look better. A visitor who cannot find a person is not a solved case, only a quieter complaint.

So my request is specific. List the questions your agents answer most often this month, check whether the help center answers each one in plain words, and place the route to a person beside every answer it does not. Then let the AI assistant read those pages, beginning with the questions your visitors ask most.

Frequently Asked Questions

What else do people ask about customer support websites?

Most remaining questions concern measurement, cost, AI handoffs and getting started, and each answer below points to something a support team can check on its own site.

What is abandonment rate, and does it apply to a support website?

Global Response's call center glossary defines abandonment rate as "the percentage of callers who hang up before a Brand Specialist answers, or before they make a selection in an interactive voice response (IVR) unit." It is the inverse of answer rate. The glossary also counts calls dropped after a set wait separately, because immediate hang-ups "do not reflect an adequate opportunity for performance." I would track the chat version of both: visitors who leave a queue before anyone replies.

Which requests should an AI assistant not handle?

Requests that carry money or account risk, such as refunds and account deletions, are better sent straight to a person. Some fast-growing software companies already route exactly these away from AI. Decide that list before launch.

Will customers accept an AI assistant on a support site?

Acceptance is higher when the assistant identifies itself as an AI assistant and a person is one click away. With LiveHelpNow, every answer carries a handoff flag, and the handoff includes the conversation summary.

How much does it cost to run?

LiveHelpNow offers a free trial for each product, so a team can test the part it needs before paying, and a demo can be requested at livehelpnow.net/demo. AI pricing in the market is often charged per outcome or per conversation, so read what counts as a resolution.

How do I get started?

Start with the questions your agents answer most often. Then decide where the route to a person sits on each page, and book a demo at livehelpnow.net/demo to see the assistant read your own content.

Written by

Michael Kansky

Founder

Michael Kansky is a serial entrepreneur, software founder, and AI-driven business operator with more than two decades of experience building companies at the intersection of customer engagement, automation, software, digital services, and data-driven growth.

Connect on LinkedIn