Your help center is ready for an AI support agent when it has a written, current answer for each of your top contact reasons, no two articles contradict each other, and the rules an agent must follow (refund windows, eligibility, when to hand over) are written as rules rather than buried in prose. You can test this in an afternoon: take your 20 most common contact reasons from last quarter, run the 8 checks below, and score each 0, 1 or 2.

For a quick first pass, free graders read your help center from its URL. Lorikeet's Knowledge Base Evaluator and Chat Thing's Knowledge Base Grader both do this without an account. Neither replaces the ticket-led checks, because only your tickets tell you what customers actually ask.

Key takeaways

• Start from tickets, not articles. Coverage of your top contact reasons matters more than the total number of articles.

• Contradictions are the expensive failure. In Moffatt v. Air Canada, a tribunal held the airline responsible after its chatbot gave an answer that conflicted with its own bereavement fare page.

• Structure for retrieval. Zendesk and Intercom both say AI agents read your content in chunks, so each section should answer one question on its own.

• Some answers cannot live in a knowledge base. Order status, identity checks and refunds need system access and fixed rules, not articles.

• Score 13 or more out of 16 on our rubric before you put an AI agent in front of customers on those topics.

Which tools check whether a knowledge base is ready for AI?

Lorikeet's free Knowledge Base Evaluator takes a public help center URL (Zendesk, Intercom or any help site with a sitemap) or a CSV export, and returns a health score covering content coverage, article freshness, duplicates and thin content. Other free options worth running:

• Chat Thing Knowledge Base Grader gives five letter grades from a URL, headlined by an AI answerability test against likely customer questions.

• Ribbo AI readiness assessment is a 10-question self-assessment of demand, knowledge, data, channels and team, with no email needed.

• My AskAI help center audit is a free skill you run with your own AI assistant that grades each article A+ to F and flags conflicting facts.

How we chose and checked these

The checklist is built from the content guidance Zendesk and Intercom publish for their own AI agents, the failure modes in NIST's generative AI risk profile, and a public tribunal decision on chatbot accuracy. The free tools were chosen because they analyse a help center or AI readiness at no cost; we read each tool's own page on 30 October 2026 and describe only what it states. Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

1. Coverage: your top contact reasons have a written answer

The first question is not "how good is our help center" but "does it answer what people contact us about". Export last quarter's tickets, group them by contact reason, and list the top 20 by volume. For each one, find the article an AI agent would use. Mark it 2 if a complete answer exists, 1 if the answer is partial or spread across pages, and 0 if nothing exists.

Intercom's guidance for Fin makes the same point from the other direction: look at patterns in customer conversations to spot the questions customers always ask, and turn useful saved replies into articles or short snippets (Intercom). Ribbo's assessment asks a related demand question: how many of your monthly questions are repeats such as prices, hours, order status and how-to (Ribbo).

Two traps are common. Teams count articles instead of contact reasons, so a help center with 400 articles still misses the top five reasons. And teams count internal macros as coverage, when the AI agent may not be able to read them. If coverage is thin, our guide to launching an AI support agent without a knowledge base covers how to start from tickets and policies instead.

2. Freshness: every article has an owner and a review date

An AI agent will repeat an outdated article as confidently as a current one. Intercom suggests sorting content by last updated date to find the articles most likely to be wrong, then asking each owning team (product, legal, security) to review its own area (Intercom). Zendesk's guidance is blunter: remove duplicate articles and conflicting information before you rely on generative replies (Zendesk).

Score 2 if every article in your top 20 has a named owner and was reviewed after the last change to the product or policy it describes. Score 1 if it is current but nobody owns it. Score 0 if you cannot tell. Ribbo's assessment asks the same question in operational terms: when a price or policy changes, how quickly is the written version updated (Ribbo)? The honest answer for many teams is "within a month", which means the AI agent is wrong for up to a month after every change.

3. Contradictions: no two pages give different answers

Contradictions are the costliest failure because the agent can pick either answer. In Moffatt v. Air Canada (2024 BCCRT 149), the British Columbia Civil Resolution Tribunal found the airline "did not take reasonable care to ensure its chatbot was accurate" after the chatbot described a bereavement fare process that differed from the airline's own bereavement travel page. The tribunal wrote that it "makes no difference whether the information comes from a static page or a chatbot" and ordered Air Canada to pay $812.02 (decision).

NIST's generative AI risk profile names this class of problem "confabulation": confidently stated but erroneous or false content by which users may be misled (NIST AI 600-1). Conflicting source content makes it more likely. To check, search your help center for each top-20 policy (refund window, cancellation notice, fees, eligibility) and note every page that states it. Score 0 if any two differ, including old blog posts, PDFs and community answers the agent can read.

4. Duplicates and thin articles are merged or removed

Near-duplicate articles split retrieval: the agent may find the older, shorter version first. Thin articles, a two-line answer with a "contact us" link, give the agent nothing to work with. Lorikeet's Knowledge Base Evaluator and My AskAI's audit both flag duplicates and outdated or overlapping content, and Chat Thing grades article quality separately from coverage (Chat Thing; My AskAI).

Score 2 if each top-20 topic has exactly one canonical article with enough detail to answer the question without a follow-up. Score 1 if duplicates exist but the canonical one is clearly marked. Score 0 if the agent has to choose between versions. Merging is usually faster than rewriting: keep the most complete version, fold in anything unique from the others, and redirect or archive the rest.

5. Structure: one topic per section, answer first

AI agents do not read an article top to bottom the way a person does. Zendesk explains that its generative AI ingests content, splits it into smaller chunks, and uses the article title and best matching chunks to write the reply. Its advice follows from that: keep content focused on one topic, give complete answers, make content self-contained, avoid vague pronouns such as "it" or "they", and use headings and numbered lists for sequential tasks (Zendesk).

Intercom's 14-factor content readiness framework says the same in different words: use headers to divide content into single-topic sections, and restate the question in the answer so it still makes sense when retrieved out of context (Intercom). A quick test: copy any one section of a top-20 article into a blank document. If a customer could act on it with no other context, score 2. If it relies on "as above" or a screenshot with no text, score 1 or 0.

6. Policies are written as rules, not as prose

Facts can be quoted. Policies have to be applied, and applying them needs the conditions spelled out. "Refunds are available within 14 days of purchase for annual plans, except for add-ons" is a rule an agent can follow. "We're always happy to help with refunds" is not. For each policy in your top 20, write down the condition, the outcome, the exceptions and who can override it.

This is also where you discover which policies need data the article cannot hold. Linktree's early AI workflows included refund and cancellation flows tied to 72-hour policies, which depend on knowing when a given customer's purchase happened (Linktree's story). Score 2 if every top-20 policy has written conditions and exceptions, 1 if conditions exist but exceptions live in people's heads, and 0 if the policy is only implied.

7. What cannot live in a knowledge base is listed separately

Some of your top contact reasons are not knowledge questions at all. "Where is my order", "change my address", "why was I charged twice" and "verify my identity" need the agent to look up or change something in your systems. Ribbo's assessment scores this separately as data: whether customer records live in a system with an API and whether a customer can be identified reliably (Ribbo).

List these topics and decide, for each, whether the AI agent will act through your systems, collect details and hand over, or not touch it. Lorikeet, for example, runs sensitive steps such as payments and identity checks as deterministic code the agent can invoke but never alter, and replays historical tickets to show knowledge gaps before go-live (how it works). Score 2 if every action topic has a decision and an owner, 1 if some do, 0 if they are still mixed into the knowledge base.

8. Escalation rules are written down

An agent that does not know when to stop is the risk that matters most in regulated support. Write down which topics and phrases always go to a person: complaints, vulnerability, legal threats, anything clinical, and any request to speak to a human. Ribbo asks whether this list exists in writing (Ribbo).

easykind, an Australian health provider, set rules that its AI agent never mentions a product by name and escalates immediately to a human when a patient reports concerning side effects (easykind's story). Linktree's agent escalates when a user asks for a human, steps back after multiple failed attempts, and screens for frustration signals before continuing (Linktree's story). Score 2 if your list is written and testable, 1 if it is informal, 0 if it does not exist.

How do you score a help center for AI readiness?

Add up the 8 checks for a score out of 16. This is our rubric, not an industry standard, and it is deliberately strict on contradictions and escalation.

Two rules override the total. Any 0 on check 3 (contradictions) or check 8 (escalation) means fix that first, whatever the score. And score per topic as well as overall: a help center can be ready for billing questions and not ready for account security. Ribbo's self-assessment uses similar bands (70 and above means ready to pilot, below 45 means groundwork first), which is a useful cross-check from someone else's model (Ribbo).

Free tools that check your knowledge base

Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim. Every tool below was checked on its own page on 30 October 2026.

Chat Thing's grader is the most complete single check because it tests whether an AI can answer likely customer questions from your docs. The Lorikeet evaluator is the fastest way to see freshness and duplicate problems across a whole help center, and its CSV mode runs in your browser, so the file never leaves your machine (Lorikeet). Its limit is that it scores content health, not answers, so pair it with checks 1 and 6 above. My AskAI's audit goes deepest per article but takes setup. Ribbo's is the right one when the question is the whole operation, not just the help center.

What still needs a human?

A readiness score tells you the content is in shape. It does not decide which topics an AI agent should own. People still need to make these calls:

• Policy sign-off. Someone accountable for each policy must confirm the written rule matches what the business will honour, because the Air Canada decision shows the business owns what its chatbot says.

• Regulated and clinical topics. Compliance teams decide what the agent may say; in healthcare the agent escalates to clinicians rather than giving clinical advice.

• Judgement calls and exceptions. Goodwill credits, hardship and complaints usually stay with people, with the agent gathering context first.

• Ongoing ownership. Ribbo asks who will own the agent day to day; "whoever has time" is the answer that predicts drift.

Questions to ask before you switch an AI agent on

• Which of our top 20 contact reasons score 2 on every check, and are we launching only on those?

• Who reviews a policy article the same day the policy changes?

• Which topics need system access, and has engineering agreed to the APIs the agent will call?

• Can we replay last month's tickets against the agent and read the failures before customers see it?

• What is our written list of always-escalate topics, and how will we test it?

For more on keeping content healthy after launch, see our practitioner's guide to knowledge bases, how teams prevent AI hallucinations in regulated support, and how a head of AI evaluates AI agents for support. If your score is 13 or more on your top topics, you can start a 30-day Lorikeet free trial and replay your own tickets before anything goes live.