# AI Knowledge Base Support in 2027: 7 Checks That Keep a Support Agent's Answers Correct

> AI knowledge base support in 2027: 7 checks to ground a support agent so answers stay correct, with a 50-question test, a refusal table and a 30-day pilot.

- Canonical: https://buffersoltech.com/blog/ai-knowledge-base-support-2027
- Author: BufferSol Editorial Team
- Published: 2026-11-21 · Updated: 2026-11-21

## Key takeaways
- A support agent is grounded when it answers from your retrieved content and shows the source. Microsoft defines retrieval-augmented generation as grounding responses 'in your proprietary content', and the answer is only as correct as that content.
- Duplicates hurt more than gaps. HelpDocs says a gap yields an 'I don't know' that is recoverable, while two disagreeing articles yield a confident answer that may be the retired one.
- Companies answer for what their chatbot says. A British Columbia tribunal found Air Canada responsible for its chatbot's wrong refund advice, per CTV and the law firm Manatt; it is one small-claims decision, not binding precedent.
- Decide what the agent does when the knowledge base is silent or in conflict. Intercom says that if a limitation is not documented, its Fin agent does not know to flag it, so write limits down and write refusal rules.
- Test before customers do: 50 real questions in five buckets, scored for a correct answer, a cited source and a current policy. Re-run after every content change.

AI knowledge base support means a support agent answers customers from your own help content and shows where each answer came from. The answers stay correct only if the content is current, unambiguous and owned, and if the agent knows when to stop. This guide gives 7 checks for 2027 that keep a grounded support agent right, with the sources named and dated November 2026.

For the full build, read our [AI customer support agent guide](/blog/ai-customer-support-agent-guide). For timing, see [how long it takes to build an AI agent](/blog/how-long-to-build-an-ai-agent).

## What does it mean to ground a support agent in your knowledge base?

It means the agent retrieves your approved content at answer time and answers from it, instead of from the model's general memory. [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) defines retrieval-augmented generation as "a pattern that extends LLM capabilities by grounding responses in your proprietary content". [Salesforce](https://www.salesforce.com/blog/ai-grounding/) puts it in plain words: "In short, grounding involves connecting your AI model to trusted sources of information."

[Google Cloud](https://cloud.google.com/vertex-ai/generative-ai/docs/grounding/overview) defines grounding as "the ability to connect model output to verifiable sources of information" and lists auditability, "links to sources", among its benefits. Grounding is a promise about process, not about truth. The agent can only be as right as the passage it finds. [Tidio](https://tidio.com/blog/ai-knowledge-base) says "The AI is only as good as the articles behind it", and [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) says "An AI support agent is only as good as the documentation it reads." Four of the pages we read, all vendors, make the same point, so we treat it as settled and spend this guide on what to do about it. See our [AI agent architecture guide](/blog/ai-agent-architecture) for how retrieval, tools and approvals fit together.

## Why do grounded support agents still give wrong answers?

Because the content, the retrieval or the missing rule is wrong, and each has a different fix. [Intercom's documentation](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency) names four content causes. If help articles "don't match the exact terms or phrasing users use", the agent "may pull irrelevant or partial content". If "multiple articles cover a topic differently", it "may choose inconsistent sources at different times". Vague or prose-heavy content makes precise replies hard. Stale content "can lead to weaker AI answers".

[HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) gives the stale case in a sentence: the agent "quotes a 30-day window when you moved to 14 days in March" because "there is still an article saying 30 days". It adds that "Duplicates hurt more than gaps", since a gap gets an answer that is recoverable and two disagreeing articles get a confident answer that may be the retired one. That page is a vendor's illustration, not a measured incident. [HubSpot](https://blog.hubspot.com/service/ai-knowledge-base) says the same in one line: "An AI knowledge base is only as accurate as its content, so outdated docs yield outdated answers."

| Cause | What the customer sees | Fix | Source |
|---|---|---|---|
| Wording does not match customers' phrasing | An irrelevant or partial answer | Use the customer's words in titles and text | [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency) |
| Two articles disagree | A confident answer that may be wrong | One definitive version per topic | [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency), [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) |
| Stale policy text | Last year's rule | Owner and review date | [Tidio](https://tidio.com/blog/ai-knowledge-base) |
| Limit never written down | A promise the product cannot keep | Document the limit | [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices) |
| No rule for silence | A plausible guess | Answer only from sources; hand off | [Tidio](https://tidio.com/blog/ai-knowledge-base) |

## What did the Air Canada chatbot case teach support teams?

It taught that a company answers for what its chatbot says, and that "check the website" is not a defence. [CTV News](https://www.ctvnews.ca/vancouver/article/air-canadas-chatbot-gave-a-bc-man-the-wrong-information-now-the-airline-has-to-pay-for-the-mistake/) reported that "A B.C. tribunal found the airline was responsible for inaccurate information given to a traveller by its chatbot." [Manatt](https://www.manatt.com/insights/newsletters/advertising-law/ai-gone-wild-airline-has-to-honor-a-refund-policy) says the chatbot "suggested he book a regular ticket and request a partial refund" under the bereavement policy, and that the tribunal found the airline "did not take reasonable care to ensure its chatbot was accurate".

[Manatt](https://www.manatt.com/insights/newsletters/advertising-law/ai-gone-wild-airline-has-to-honor-a-refund-policy) adds that the tribunal member questioned "why customers should be expected to double-check the information" across the airline's website. We did not open the decision itself, and the ABA's summary returned an access error. It is a small-claims tribunal ruling, not binding precedent, so read it as a warning, not as law. This is not legal advice. The practical step for a support team is cheap: label policy text with a version and date, keep a change log and make the agent quote both. Our [AI customer service agent guide](/blog/ai-customer-service-agent) and [AI help desk guide](/blog/ai-help-desk) cover the operations side.

## What does a wrong answer cost, and how do you size it?

It costs a refund or a goodwill gesture, an agent's time to fix it and a customer's trust. BufferSol estimate, with the assumptions shown: 2,000 AI-handled conversations a month, 3 percent wrong on policy (60 conversations), 15 minutes of agent time to correct each at a loaded $30 an hour ($7.50), and a $20 goodwill credit. That is 60 x $27.50, about $1,650 a month, before any legal or churn cost. Replace each assumption with your own.

[HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) calls stale content "the single most common source of a wrong-on-today's-facts answer", and [Tidio](https://tidio.com/blog/ai-knowledge-base) says "Stale or conflicting content produces confident wrong answers", so the loss in this example is usually a content problem before it is a model problem. The point of the exercise is to compare the cost of the wrong answers with the cost of a weekly content review. BufferSol estimate: If the review takes two hours a week at $30, it costs about $260 a month, which is well below the example loss. Our [AI customer service savings calculator](/tools/ai-customer-service-savings-calculator) turns your ticket volume and handle time into a monthly figure, and the [AI agent ROI calculator](/tools/ai-agent-roi-calculator) covers the wider case.

```infographic
{"type":"stats","title":"Worked cost of wrong policy answers (BufferSol estimate, every input replaceable)","items":[{"value":"2,000","label":"AI-handled conversations a month, an assumption"},{"value":"60","label":"conversations wrong on policy at 3 percent, an assumption"},{"value":"$27.50","label":"cost of each wrong answer: 15 minutes at $30 an hour plus a $20 credit"},{"value":"$1,650","label":"monthly cost of wrong answers in this example"}],"source":"BufferSol estimate with the assumptions stated in the text"}
```

## What changes in 2027 for AI knowledge base support?

Three things change: retrieval gets smarter, grounding gets checked and the burden of proof moves to you. Retrieval first. [Microsoft's agentic retrieval page](https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-overview) says the approach "can use LLM-based query planning to break down a complex query into smaller, focused subqueries", runs them in parallel and retains references for citation. An agent no longer asks one question of your knowledge base. It asks several, so a contradiction in one corner can surface.

Second, vendors now ship checks. [Amazon Bedrock Guardrails](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html) supports "contextual grounding checks to detect and filter hallucinations in model responses" when a reference source and a user query are provided, and treats "any new information introduced in the response" as un-grounded. Third, retrieval research keeps moving: [Anthropic](https://www.anthropic.com/news/contextual-retrieval) reports its Contextual Retrieval method "can reduce the number of failed retrievals by 49%" and, with reranking, by 67%, which is Anthropic's own benchmark, not an independent one. BufferSol view: in 2027, the winners will not have the cleverest retrieval. They will have the cleanest content and the best test set. Our [AI readiness assessment post](/blog/ai-readiness-assessment) helps you check your content first.

Here are the seven checks at a glance:

| # | Check | The short answer | Evidence |
|---|---|---|---|
| 1 | One owned source per policy | One definitive version, an owner, a review date | [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency), [Tidio](https://tidio.com/blog/ai-knowledge-base) |
| 2 | Articles that retrieve cleanly | One topic per section, answer first | [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices), [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) |
| 3 | Quote and cite | The agent shows source, version and date | [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview), [Anthropic](https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) |
| 4 | A rule for silence | Say not covered; hand off | [Tidio](https://tidio.com/blog/ai-knowledge-base), [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices) |
| 5 | Access by audience | Retrieve only what the asker may see | [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) |
| 6 | A gold test | 50 questions, five buckets, a pass mark | [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/), [AWS](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html) |
| 7 | A review cycle | Owners re-verify on a schedule | [Tidio](https://tidio.com/blog/ai-knowledge-base), [Help Scout](https://helpscout.com/blog/ai-knowledge-base) |

## 1. Does every policy have one current, owned source article?

Give each policy one definitive article, an owner and a review date. [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency) advises to "Consolidate overlapping or contradictory articles", because "it's better to have one definitive version per topic". [Tidio](https://tidio.com/blog/ai-knowledge-base) says "Give every article an owner and a review date, and update articles when policies change", and to "Remove or archive old versions instead of leaving them in the source list".

[HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) calls this "usually the highest-value hour you will spend", because most knowledge bases "accumulate near-duplicates rather than being written that way". Its fix is to merge the pair, keep one URL and redirect the loser. Start with the policies that cost money when wrong: returns, refunds, shipping times, warranty, pricing and cancellation. Our [returns automation guide](/blog/returns-automation-ai) and [WISMO automation post](/blog/wismo-automation) show where those policies bite.

**Choose it first if** your knowledge base has grown for years and nobody owns it. **Skip the heavy version if** you have 20 articles and one owner; a quick read-through will do.

## 2. Is each article written so the agent can retrieve one clean passage?

Write one question per article, answer it first and keep sections to one topic. [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices) says "Each section should cover one thing", and warns that when information is ambiguous, Fin "may make inferences to provide the most helpful answer it can", "which can occasionally result in inaccurate answers". Its example of poor practice is a heading that says only "Contact us" above lines labelled just "Sales", "Support" and "Advertising", where the better version names the team in each line.

[HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) explains why order matters: "Retrieval slices your article into passages", so if the answer sits in paragraph nine, "the passage that gets pulled is the background". Its rule is "Write the answer first, then the context." [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) agrees that "RAG quality depends on how you prepare content for retrieval". [Tidio](https://tidio.com/blog/ai-knowledge-base) adds that "Short articles with one question each work better." Our [AI chatbot vs AI agent comparison](/blog/ai-agent-vs-chatbot) explains why a retrieval bot differs from an agent that acts.

A BufferSol article template that follows these rules, with each field a quoted-passage target:

| Field | What goes in it | Why |
|---|---|---|
| Title | The customer's question, in their words | Matches phrasing |
| Answer | One to three sentences, first | The passage that gets pulled |
| Policy version and effective date | "v3, effective 1 March 2027" | Lets the agent quote the date |
| Last verified | A date and a name | Proves it is current |
| Exceptions and limits | What this does not cover | Prevents invented promises |
| Owner | A named person or team | Someone to ask and to blame |
| Related articles | Links, not copies | Avoids duplicate text |

## 3. Does the agent quote the source, and is it forced to cite?

It should show the article, the version and the date with every policy answer. [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) lists "Built-in citation tracking shows provenance" among agentic retrieval features, and its agentic retrieval page says references "are extracted and retained for citation purposes". A citation is how a customer, an agent or a regulator checks the answer.

You can also force the habit in the prompt. [Anthropic's guidance on reducing hallucinations](https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) says to "Verify with citations": make the response auditable by having the model cite quotes for each claim, and "If it can't find a quote, it must retract the claim." That is a model provider's tip, and it reduces errors without removing them. For support, the rule becomes: any sentence about money, time limits or eligibility carries a quote from the article and its version. If the agent cannot quote, it does not say it. [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency) notes queries without matching content lead to weaker answers, which is exactly when the quote check should block the reply.

## 4. What does the agent do when the knowledge base has no answer?

It says the topic is not covered and offers a person, and it logs the question as a content gap. [Tidio](https://tidio.com/blog/ai-knowledge-base) warns that "Without a rule to stop, a model may fill the gap with a plausible guess", and advises: "Set the AI to answer only from your sources and to hand off when it is unsure." [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) describes the failure: "an agent with no content on the subject will often construct a plausible answer from adjacent material".

The write-it-down fix matters most. [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices) says "If a limitation isn't documented, Fin doesn't know to flag it", and [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) says a short "what this does not do" section "stops your agent making commitments your product cannot keep". Here is the refusal table BufferSol would put in the agent's rules, one row per kind of retrieved evidence.

| What retrieval returns | The agent does | The person does |
|---|---|---|
| One current article that answers the question | Answers, quotes the line, shows version and date | Nothing |
| Two articles that disagree | Does not answer; says a person will confirm | Content owner fixes the conflict |
| An article past its review date | Does not state the policy as current; hands off | Owner re-verifies |
| Nothing relevant | Says it is not covered; offers a human; logs the gap | Owner writes the article |
| A documented limit | States the limit and the workaround | Nothing |
| Content the asker may not see | Declines and explains | Nothing |
| Score below the threshold | Does not send; hands off | Team reviews the miss |

```infographic
{"type":"compare","title":"Guessing agent or grounded agent with a refusal rule","columns":["No rule for silence","Grounded with a refusal rule"],"rows":[{"label":"When content is missing","a":"Builds a plausible answer from adjacent material (HelpDocs)","b":"Says not covered and hands off (Tidio's advice)"},{"label":"When two articles disagree","a":"May pick either (Tidio)","b":"Declines and flags the owner"},{"label":"What the customer can check","a":"Nothing","b":"Source, version and date"},{"label":"Who fixes the cause","a":"Nobody sees it","b":"The content owner, from the gap list"}]}
```

For the last row, [AWS](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html) shows what a threshold looks like: with grounding and relevance thresholds at 0.7, responses scoring less "will be detected as hallucinations and blocked", and "a threshold of 0.7 for grounding and relevance is a reasonable starting point" for conversational use. It adds that you should test thresholds "with your own data". Thresholds are a product setting, not a law. Our [AI customer support agent guide](/blog/ai-customer-support-agent-guide) covers what the human should receive after the handoff.

```infographic
{"type":"steps","title":"What the agent does with what retrieval returns","steps":[{"title":"Retrieve","text":"Find the passages that match the customer's question and the customer's access."},{"title":"Check","text":"One current source, no conflict, inside its review date, above the score threshold."},{"title":"Answer or stop","text":"Answer with a quote, version and date; otherwise say not covered and hand off."},{"title":"Log","text":"Record the question, the sources and any gap for the content owner."}]}
```

## 5. Does the agent retrieve only what this customer may see?

It must, and access should be enforced where the content is retrieved, not in the prompt. [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) states the problem: "Finance data should only be accessible to finance team, even when an executive asks the chatbot", and lists "Knowledge source-level access control" among agentic retrieval's security features. [Tidio](https://tidio.com/blog/ai-knowledge-base) says to keep "personal data, internal notes and draft policies out of any source the customer-facing AI can read". [Microsoft's security overview](https://learn.microsoft.com/en-us/azure/search/search-security-overview) lists "document-level permissions" among the things you configure yourself, not something the platform does for you.

For a public help centre the rule is simple: the agent's source list contains published articles only, never internal runbooks, draft policies or ticket notes. For a logged-in customer, add account data through tools with their own permissions, not through the knowledge base. BufferSol's [Enterprise Knowledge Assistant case study](/portfolio/enterprise-knowledge-assistant) describes a retrieval assistant that "only uses documents each user is allowed to see", per its own page. Our [AI vendor security questionnaire post](/blog/ai-vendor-security-questionnaire) and [data processing agreement guide](/blog/dpa-ai-vendor) cover the questions to ask a vendor about where content is stored.

## 6. Do you test the agent on real questions before customers do?

Run a gold test: 50 real questions, five buckets, a pass mark in each, repeated after every change. [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) gives a fast version in about two hours: "Pull your ten most common support questions from ticket volume, plus the ten searches that returned nothing" and "Ask your agent all twenty." Then open the article behind every wrong, hedged or accidentally right answer. That is a good first afternoon. A launch needs more.

BufferSol's gold test uses 50 questions in five buckets, so every failure type is covered:

| Bucket | Questions | What it tests | Pass mark (BufferSol starting point) |
|---|---|---|---|
| Top real questions | 20 | The everyday load | 90% correct, cited and current |
| Not in the knowledge base | 10 | The refusal rule | 100% say not covered |
| Recently changed policies | 8 | Staleness and duplicates | 100% current |
| Conflicting articles | 6 | The conflict rule | 100% decline and flag |
| Permission traps | 6 | Access control | 100% decline |

For each question, store the expected answer, the article it should cite and the policy version. Score three things: correct, cited and current. The pass marks are our starting point, not an industry standard. A single failure in a 100 percent bucket means the content or the rules need fixing before launch. [AWS](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html) advises testing thresholds "with your own data" when settings change, and the same applies to prompts, models and content. Our [AI agent evaluation guide](/blog/ai-agent-evaluation) shows how to score and track a test set over time.

## 7. Do owners re-verify policy articles on a schedule?

They should, on a fixed cycle and on every policy change. [Tidio](https://tidio.com/blog/ai-knowledge-base) says to give every article an owner and a review date. [Help Scout](https://helpscout.com/blog/ai-knowledge-base) says "Maintenance is even more crucial with an AI tool than with a typical knowledge base", and that content should be "proofread by a subject matter expert and verified for accuracy before publishing".

Here is a review plan that fits a small team. Policy articles that touch money or eligibility: every 90 days and on any change, owned by the policy owner. Product how-tos: every release. General guidance: every 6 months. Log each verification with a date and a name, because the agent can quote the date. BufferSol view: tie the policy owner's change process to the knowledge base, so a new return window cannot go live without the article changing the same day. [HubSpot](https://blog.hubspot.com/service/ai-knowledge-base) makes "Keep it up to date" the fifth step in its guide to building an AI knowledge base, and our [AI workflow automation guide](/blog/ai-workflow-automation-guide) shows how to automate the reminders.

## What does the support agent workflow look like step by step?

It is a fixed loop with a person at each decision. This is the workflow BufferSol would build on your helpdesk and knowledge base, with the tools each step touches. The matching case study is the [AI Customer Support Agent](/portfolio/ai-customer-support-agent), which handles routine tickets and sends edge cases to a person with a summary.

1. **Receive.** The question arrives in chat or email, with the customer's identity if known.
2. **Retrieve.** The agent plans one or more searches and fetches passages the customer may see.
3. **Check.** It tests for one current source, no conflict, review date and a score threshold.
4. **Answer.** It writes the reply, quotes the line and shows version and date.
5. **Act, if allowed.** For order or account questions it calls a tool, such as an order lookup, within limits.
6. **Hand off.** If any check fails, it says so and sends the case to a person with the packet.
7. **Log.** It records question, sources, answer and outcome.
8. **Learn.** Gaps and conflicts go to the content owner's weekly list.

[Intercom](https://www.intercom.com/help/en/articles/12396892-manage-fin-ai-agent-s-escalation-guidance-and-rules) notes that escalation offers are skipped when "No human support is available", so plan the after-hours path too. See our [AI help desk guide](/blog/ai-help-desk) and [AI customer service agent guide](/blog/ai-customer-service-agent). The matching ready-made build is [Support Ticket Triage and Resolution](/done-for-you/automations/ticket-triage-resolution).

## Where does a person approve in a grounded support agent?

At four points: new or changed policy content, the refusal rules, anything above a money limit and the weekly gap list. The agent reads, answers within its rules and flags. People own the content and the rules. [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) frames the agent as an answerer over your content, so the human owner of that content is the real control.

Name the approver for each: the policy owner approves content changes, the support lead approves refusal rules and thresholds, a manager approves refunds above a limit and a content owner clears the gap list each week. Our [human-in-the-loop guide](/blog/human-in-the-loop-ai-agents) explains how to build the approvals, and the [shared support inbox automation](/done-for-you/automations/support-inbox-email) shows an email-first version.

| Decision | Owner | When | Stance taken from |
|---|---|---|---|
| New or changed policy article | Policy owner | Before publishing | [Help Scout](https://helpscout.com/blog/ai-knowledge-base), [Tidio](https://tidio.com/blog/ai-knowledge-base) |
| Refusal rules and thresholds | Support lead | Before launch and each quarter | [AWS](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html), [Intercom](https://www.intercom.com/help/en/articles/12396892-manage-fin-ai-agent-s-escalation-guidance-and-rules) |
| Money above a limit | Manager | Per case | [Manatt](https://www.manatt.com/insights/newsletters/advertising-law/ai-gone-wild-airline-has-to-honor-a-refund-policy) |
| Weekly gap list | Content owner | Weekly | [Tidio](https://tidio.com/blog/ai-knowledge-base), [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) |

## Does grounding remove wrong answers completely?

No. It reduces them, and you still need checks. [Anthropic](https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) says even advanced models "can sometimes generate text that is factually incorrect or inconsistent with the given context", and its techniques are to "minimize" hallucinations. [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices) says ambiguous content can "occasionally result in inaccurate answers" even in a purpose-built agent.

So plan for the residual. Keep the refusal table, the threshold and the gold test running. [Anthropic](https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) also says to "Explicitly give Claude permission to admit uncertainty", which "can drastically reduce false information". It is advice from a model provider and not a guarantee. The honest promise to a customer is not "always right". It is "answers come from our policy, with the source shown, and a person steps in when we are unsure". Our [AI customer support agent guide](/blog/ai-customer-support-agent-guide) sets the same expectation.

## How should the agent word a refusal?

Plainly, with the next step. A refusal that sounds like a failure invites the customer to rephrase until the agent guesses. Three templates work, with details invented for illustration. When nothing covers it: "I don't have a published answer on that, so I won't guess. I'm passing this to a teammate, who will reply today." When two sources conflict: "Our information on this is being confirmed, so I've asked a teammate to check and reply." When it is a documented limit: "We don't support bulk export for custom reports. You can export each report one by one, and here is how."

[Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices) shows the value of the third kind: a specific limit with a workaround "prevents Fin from giving a misleadingly incomplete answer", and its own example is a bulk-export limit. [Intercom's escalation page](https://www.intercom.com/help/en/articles/12396892-manage-fin-ai-agent-s-escalation-guidance-and-rules) describes configurable messages for the handoff. Keep refusals short, name the next step and give a time. Our [AI customer service software guide](/blog/ai-customer-service-software) covers the tooling.

## When should a knowledge gap trigger an escalation?

When the customer's issue needs an answer the content cannot give and the cost of a guess is high. For low-stakes questions, such as a feature how-to, "not covered, here is the support link" can be enough. For money, eligibility, safety or legal topics, hand off at once. This is where the knowledge base meets the handoff.

Pass the packet: the question, the passages retrieved, the reason the agent stopped and the customer's identity. A person who sees "retrieved two conflicting returns articles" fixes the content as well as the case. Our [AI customer support agent guide](/blog/ai-customer-support-agent-guide) covers the handoff design, and [Intercom](https://www.intercom.com/help/en/articles/12396892-manage-fin-ai-agent-s-escalation-guidance-and-rules) documents how Fin offers a handoff and when it does not.

## How do you handle conflicting articles?

Find them, merge them and redirect the old one. [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) says to "Merge the pair, keep one URL, and redirect the loser", and [Tidio](https://tidio.com/blog/ai-knowledge-base) adds the case of "an old PDF" that says something different from the pricing page: "the AI may pick either". Search your sources for the same keywords, such as "refund", "returns" and "cancel", and compare the answers by eye.

Two cheap detection methods. First, ask the agent your 10 most common questions three times each with slightly different wording and compare the cited articles; if the citations differ, there is a conflict. Second, run the "conflicting articles" bucket of the gold test. [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency) says that with overlapping articles, Fin "may choose inconsistent sources at different times", so the symptom is inconsistency, not one visible error. See our [workflow exception handling guide](/blog/workflow-exception-handling) for routing conflicts to an owner.

## How do you measure whether the knowledge base is working?

Track five numbers weekly: grounded-answer rate, citation rate, correct-refusal rate, stale-answer count and gap list length. [Tidio](https://tidio.com/blog/ai-knowledge-base) suggests a different five, including resolution rate and no-answer rate, and notes that vendors publish their own averages, "so measure yours". Ours are built around correctness, because a high resolution rate with wrong answers is worse than a low one.

| Measure | Formula | Why |
|---|---|---|
| Grounded-answer rate | Sampled answers whose claims match the cited article / sampled answers | The core accuracy figure |
| Citation rate | Policy answers with a quote, version and date / policy answers | Auditability |
| Correct-refusal rate | Questions with no source that were refused / such questions | The rule for silence |
| Stale-answer count | Answers citing an article past its review date | Content health |
| Gap list length | Open unanswered questions | Content backlog, which should shrink |

Sample 50 answers a week and score them by hand against the articles. [Tidio](https://tidio.com/blog/ai-knowledge-base) also lists "no-answer rate" as "your content gap list", which maps to our gap length. Our [AI agent evaluation guide](/blog/ai-agent-evaluation) covers scoring, and the [customer feedback and insights automation](/done-for-you/automations/customer-feedback-insights) turns ticket themes into content work.

## Who should own the knowledge base?

A named content owner, plus a policy owner for each policy, with the support lead as the final approver. Ownership is the cheapest control in this guide and the most skipped. [Tidio](https://tidio.com/blog/ai-knowledge-base) says "Give every article an owner and a review date", and our template adds the owner field to the article itself.

The policy owner is usually outside support: finance owns refunds, operations owns shipping, legal owns terms. The support content owner makes sure the article matches the policy and publishes it. BufferSol view: put the knowledge base change into the same ticket as the policy change, so one cannot ship without the other. Our [AI customer service guide](/blog/ai-customer-service) and [AI readiness assessment post](/blog/ai-readiness-assessment) cover ownership.

## How long does it take to fix a knowledge base for an AI agent?

A first pass takes days, and a dependable one takes weeks. [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) says its quick audit "takes about two hours", and finds "the same two or three faults repeating", so "a small number of fixes moves most of your answers". That matches the usual pattern: most gain comes from merging duplicates, putting answers first and writing down limits.

A realistic plan, BufferSol estimate for a small help centre of 100 articles: day 1 to 2, the 20-question audit; week 1, merge duplicates and fix the 20 articles that serve the most tickets; week 2, write the "what this does not do" sections; weeks 3 to 4, build the gold test and fix the failures. A larger help centre scales roughly with article count. [Help Scout](https://helpscout.com/blog/ai-knowledge-base) advises focusing on "high-volume issues" first, which is the same idea.

## Which tools and pages cover AI knowledge base support, and what do they say?

Most are help desk or knowledge base vendors, and each page sells its own product. We read the pages in November 2026. "Not on the page" means we did not find it, not that the product lacks it.

| Source | What it covers | What it does not give | Fit |
|---|---|---|---|
| [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices) | Content rules for its Fin agent, limits, examples | A test protocol | Intercom users |
| [HubSpot](https://blog.hubspot.com/service/ai-knowledge-base) | Definition, content types, building steps, software list | A refusal rule | HubSpot users and beginners |
| [Help Scout](https://helpscout.com/blog/ai-knowledge-base) | Eight tools and maintenance tips | Measured accuracy | Teams comparing tools |
| [Tidio](https://tidio.com/blog/ai-knowledge-base) | Challenges, measurement, owner and review date | Test method | Small online shops |
| [Salesforce](https://www.salesforce.com/blog/ai-grounding/) | What grounding is, from June 2024 | 2027 retrieval changes | Salesforce users |
| [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) | Six fixes and a 20-question audit | A bucketed test | Teams starting an audit |
| [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview) | The RAG mechanism, citations, access control | Support-specific rules | Teams building their own |

**When to choose a packaged support agent:** your helpdesk is one of the supported ones and its content rules fit your articles. **When to choose a custom agent:** you need your own refusal rules, thresholds, tests and audit log, or the agent must act in several systems. Our [AI agent pricing research](/research/ai-agent-pricing-2027) lists vendor prices, our [AI customer service software guide](/blog/ai-customer-service-software) compares options, and the [AI agent development guide](/blog/ai-agent-development-guide) covers a custom build.

## How strong is the evidence behind each claim in this post?

It ranges from a platform's own documentation to vendor marketing, and we say which is which. Weight a platform's documentation and a tribunal report above a vendor's marketing page, and weight vendor pages above an unsourced figure. We opened every page below in November 2026.

| Source | Type | What it supports | The limit we found |
|---|---|---|---|
| [Microsoft RAG](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview), [Microsoft agentic retrieval](https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-overview) | Platform docs | The mechanism, citations, access control | Describes one platform |
| [AWS Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html) | Platform docs | Grounding checks and thresholds | A product setting, not a standard |
| [Anthropic hallucinations guide](https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations), [Contextual Retrieval](https://www.anthropic.com/news/contextual-retrieval) | Model provider docs | Admit uncertainty; cite quotes; retrieval gains | The 49% and 67% figures are Anthropic's own benchmark |
| [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency) | Vendor help docs | Content failure modes | Written for its own agent |
| [CTV](https://www.ctvnews.ca/vancouver/article/air-canadas-chatbot-gave-a-bc-man-the-wrong-information-now-the-airline-has-to-pay-for-the-mistake/), [Manatt](https://www.manatt.com/insights/newsletters/advertising-law/ai-gone-wild-airline-has-to-honor-a-refund-policy) | News and law-firm summary | The tribunal finding | Secondary; small-claims decision |
| [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/), [Tidio](https://tidio.com/blog/ai-knowledge-base), [Help Scout](https://helpscout.com/blog/ai-knowledge-base), [HubSpot](https://blog.hubspot.com/service/ai-knowledge-base), [Salesforce](https://www.salesforce.com/blog/ai-grounding/) | Vendor blogs | Practices and pitfalls | Vendor interest; no measured accuracy figures |

The pattern is clear. The practices are consistent across vendors, and no source we opened publishes a measured accuracy figure for grounded versus ungrounded support answers. That is why this guide leans on testing. The same split runs through our [AI agent evaluation guide](/blog/ai-agent-evaluation) and [AI customer support agent guide](/blog/ai-customer-support-agent-guide).

## What does a before-and-after of one policy article look like?

It looks like the same facts, reordered, dated and bounded. Here is an invented returns article, an illustration and not a client page, before and after the checks. The facts are made up, and the structure follows the pages we read.

| Element | Before | After | Rule from |
|---|---|---|---|
| Title | "Returns" | "Can I return an item, and how long do I have?" | [Intercom](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency): match customer phrasing |
| First paragraph | Three paragraphs of background | "Yes. You have 14 days from delivery." | [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/): answer first |
| Version | None | "Policy v3, effective 1 March 2027, verified 2 March by A. Patel" | [Tidio](https://tidio.com/blog/ai-knowledge-base): owner and review date |
| Exceptions | Hidden in a footer | "Not covered: sale items, opened hygiene products" | [Intercom](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices): document limits |
| Older copies | An old PDF still says 30 days | Archived and redirected | [Tidio](https://tidio.com/blog/ai-knowledge-base), [HelpDocs](https://www.helpdocs.io/learn/ai-customer-support-agents/) |
| Related | Copied text from the shipping page | A link to the shipping article | [Microsoft](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview): prepare content for retrieval |

With the "after" article, the agent's reply can read: "You can return it within 14 days of delivery (Returns policy v3, effective 1 March 2027). Sale items are not covered." Every part of that sentence traces to a line a person owns. With the "before" article, the agent might quote whichever copy it retrieved. The [Air Canada case](https://www.ctvnews.ca/vancouver/article/air-canadas-chatbot-gave-a-bc-man-the-wrong-information-now-the-airline-has-to-pay-for-the-mistake/) is a reminder of why the difference matters.

If you run several agents, for example one for orders and one for policy, keep one shared policy source so they cannot disagree, as in our [single-agent vs multi-agent systems post](/blog/single-agent-vs-multi-agent-systems). For email-first teams, our [AI email automation post](/blog/ai-email-automation) shows the same rules on a shared inbox. The [AI workflow automation examples](/blog/ai-workflow-automation-examples) and [state of AI agents for SMBs research](/research/state-of-ai-agents-smb-2027) put this in context, and the [AI agent ROI calculator](/tools/ai-agent-roi-calculator) helps size the effort. The [AI agents service page](/services/ai-agents) describes how BufferSol builds and tests the whole loop.

## What do the top results on AI knowledge base support leave out?

They define the term, list benefits and rank tools; [HubSpot](https://blog.hubspot.com/service/ai-knowledge-base) names five content types and software options, [Help Scout](https://helpscout.com/blog/ai-knowledge-base) ranks eight tools and [Salesforce](https://www.salesforce.com/blog/ai-grounding/) explains grounding as of June 2024. They rarely give a test, a refusal rule or a review schedule. We read [HubSpot](https://blog.hubspot.com/service/ai-knowledge-base), [Tidio](https://tidio.com/blog/ai-knowledge-base), [Help Scout](https://helpscout.com/blog/ai-knowledge-base) and [Salesforce](https://www.salesforce.com/blog/ai-grounding/) in November 2026. A fifth result, Zendesk, blocked our plain request with a bot check, so we did not read it, and the ABA's Air Canada summary returned an access error.

The gaps this post fills are the 50-question gold test, the refusal table, the article template with a quotable version and date, the cost of wrong answers and the 2027 view on retrieval. For related reading, see our [AI customer support agent guide](/blog/ai-customer-support-agent-guide), [AI customer service guide](/blog/ai-customer-service), [AI help desk post](/blog/ai-help-desk) and [AI email automation post](/blog/ai-email-automation).

## What does the AI not do in knowledge base support?

It does not know your policy changed, it does not know what you meant to write and it cannot tell a retired article from a current one without metadata. It repeats what the content says. [Tidio](https://tidio.com/blog/ai-knowledge-base) says "The AI repeats whatever your articles say, including last year's return window."

There are limits in our evidence too. Vendor pages describe practices, and none gives a measured accuracy figure. The Air Canada ruling is one small-claims decision. The [Anthropic retrieval figures](https://www.anthropic.com/news/contextual-retrieval) are the vendor's own benchmark. And [OpenAI's retrieval guide](https://platform.openai.com/docs/guides/retrieval) shows a practical limit worth knowing: a search response returns 10 results by default and up to 50 with the `max_num_results` setting, so what the agent sees is always a slice of your content, which is why one clean answer per topic beats many partial ones. A balanced reading: grounding is a strong control, not a guarantee.

## Should you buy a packaged agent, build one or run a pilot first?

Pilot first. A pilot takes weeks and gives you your own scores on your own content. If a packaged agent passes your gold test and fits your helpdesk, buy it. If you need your own refusal rules, thresholds and audit log, or the agent must act across systems, a custom agent fits.

A custom agent wins when you want to own your system, act inside your own tools and avoid per-resolution fees. A packaged tool wins when your setup is standard and you want a vendor to carry updates. BufferSol builds the first kind, so weigh that. Our [AI automation agency pricing guide](/blog/ai-automation-agency-pricing), [AI agent development guide](/blog/ai-agent-development-guide) and [AI customer service guide](/blog/ai-customer-service) cover costs and options.

## What does a 30-day pilot deliver for knowledge base support?

In 30 days you get a working support agent grounded in your real knowledge base, tested on your own gold questions, with a person approving anything beyond its rules. The plan follows the delivery model on our [AI agents service page](/services/ai-agents).

- **You bring:** read access to your help centre and helpdesk, 200 to 300 recent tickets, the list of policies that cost money when wrong, a named owner for each and one support lead who can decide within a day.
- **We build, in about two weeks:** a prototype with retrieval over your published articles, quotes with version and date, the refusal table and a 50-question gold test scored on your content.
- **We launch, in the final stretch:** a live pilot on a first slice of tickets with a person approving above your limits, a weekly 50-answer sample, a gap list for your content owners and a report on what to fix next.

BufferSol delivery data: a working agent on your real data inside 30 days, and Done For You automations live in 3 to 6 weeks. Success is measured on your baseline: grounded-answer rate, correct-refusal rate, stale-answer count and share of tickets resolved without a person. A pilot is not a full rollout; it shows what a rollout needs. See our [Enterprise Knowledge Assistant case study](/portfolio/enterprise-knowledge-assistant) for the retrieval pattern.

## Is this right for you?

Check these signs. Three or more means a pilot is worth scoping.

- You have more than one article on returns, shipping or pricing and nobody is sure which is current.
- Your bot or agent has quoted an old policy to a customer.
- Nobody owns the help centre, or articles have no review date.
- You have no list of questions the agent must answer correctly, and no list it must refuse.
- Agents correct the same wrong AI answers every week.

The matching ready-made automation is [Support Ticket Triage and Resolution](/done-for-you/automations/ticket-triage-resolution), a Done For You build for support teams that want triage, grounded replies and escalation in 3 to 6 weeks. For policy-heavy cases, see [Refunds and Returns](/done-for-you/automations/refunds-returns), and browse the [Done For You catalogue](/done-for-you). Related reading: [AI customer service software](/blog/ai-customer-service-software) and [AI customer service agent](/blog/ai-customer-service-agent).

## Is there anything we did not re-check?

Yes. **Facts we did not re-check:** facts checked November 2026 against the pages named in this post. We did not read the Air Canada tribunal decision itself, only [CTV's](https://www.ctvnews.ca/vancouver/article/air-canadas-chatbot-gave-a-bc-man-the-wrong-information-now-the-airline-has-to-pay-for-the-mistake/) and [Manatt's](https://www.manatt.com/insights/newsletters/advertising-law/ai-gone-wild-airline-has-to-honor-a-refund-policy) reports, and we state no damages figure. Zendesk's guide blocked our plain request and the ABA page returned an access error, so neither is used. [Tidio's](https://tidio.com/blog/ai-knowledge-base) claim about resolution rates is a vendor figure we do not repeat. The cost example, review cycle, pass marks and refusal templates are BufferSol estimates or illustrations. Nothing here is legal advice.

If you want a support agent whose answers you can test, we recommend asking BufferSol for a scoped estimate of a 30-day pilot now, so your 2027 support plan starts with a knowledge base that is owned, current and scored on your own questions.

## Frequently asked questions
### What does grounding a support agent in a knowledge base mean?
Grounding means connecting an AI support agent to trusted content so its answers come from that content, not from the model's general memory. Salesforce puts it as connecting your AI model to trusted sources of information. In practice the agent retrieves the right article passage at answer time, answers from it and shows the source, so a person can check it.

### Why do AI support agents give wrong answers from a knowledge base?
Mostly because of the content. Intercom's documentation lists terms that do not match how customers phrase questions, articles that cover a topic differently, vague or overloaded writing, and stale content. HelpDocs adds that duplicates hurt more than gaps. Fixing those four causes removes most wrong answers before any model tuning is needed.

### Is a company responsible for wrong answers from its AI chatbot?
In one widely reported case, yes. A British Columbia tribunal found Air Canada responsible for inaccurate information its chatbot gave a traveller, per CTV News, and Manatt reports it found the airline did not take reasonable care to ensure the chatbot was accurate. It was a small-claims decision, not binding precedent, and this is not legal advice.

### What should an AI support agent do when the knowledge base has no answer?
It should say the topic is not covered and offer a person. Tidio advises setting the AI to answer only from your sources and to hand off when unsure, noting that without a rule to stop, a model may fill the gap with a plausible guess. Log every unanswered question as a content gap for an owner to fill.

### How do you test an AI support agent before launch?
Build a gold set of about 50 real questions, with expected answers and the article each should cite. Include questions the knowledge base cannot answer, recently changed policies, conflicting articles and permission traps. Score each answer for correctness, citation and currency, and re-run the set after every content or model change. HelpDocs suggests a faster 20-question audit.

### How often should knowledge base articles be reviewed for an AI agent?
Give every article an owner and a review date, as Tidio advises, and update it whenever the policy changes. BufferSol suggests reviewing policy articles on a fixed cycle, such as every 90 days, and on any policy change, with a log that shows when each was last verified. Stale policy text is where wrong answers come from.

### How long does it take to ground a support agent in a knowledge base?
A first working version on your real articles and tickets is a matter of weeks. BufferSol's 30-day pilot puts a working support agent on your real knowledge base: a prototype in about two weeks, scored on your own gold questions, then a live pilot on a first slice of tickets, with a person approving anything beyond its rules.

## Sources
1. [RAG and Generative AI - Foundry IQ (Azure AI Search)](https://learn.microsoft.com/en-us/azure/search/retrieval-augmented-generation-overview), Microsoft Learn
2. [Agentic retrieval in Azure AI Search](https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-overview), Microsoft Learn
3. [How to improve AI-generated answers and ensure consistency](https://fin.ai/help/en/articles/13976177-how-to-improve-ai-generated-answers-and-ensure-consistency), Intercom (Fin help centre)
4. [Optimizing content for Fin](https://www.intercom.com/help/en/articles/7860255-fin-ai-agent-content-best-practices), Intercom Help
5. [Manage Fin AI Agent's escalation guidance and rules](https://www.intercom.com/help/en/articles/12396892-manage-fin-ai-agent-s-escalation-guidance-and-rules), Intercom Help
6. [How to create an AI knowledge base that actually works](https://blog.hubspot.com/service/ai-knowledge-base), HubSpot
7. [The 8 Best AI Knowledge Base Software in 2026](https://helpscout.com/blog/ai-knowledge-base), Help Scout
8. [AI Knowledge Base: How It Works and Why It Matters](https://tidio.com/blog/ai-knowledge-base), Tidio
9. [10 Ways To Prep Your Knowledge Base for AI Grounding](https://www.salesforce.com/blog/ai-grounding/), Salesforce
10. [AI support agents: how to optimize your knowledge base](https://www.helpdocs.io/learn/ai-customer-support-agents/), HelpDocs
11. [AI Gone Wild: Airline Has to Honor a Refund Policy Its Chatbot Fabricated](https://www.manatt.com/insights/newsletters/advertising-law/ai-gone-wild-airline-has-to-honor-a-refund-policy), Manatt
12. [Air Canada's chatbot gave a B.C. man the wrong information. Now, the airline has to pay for the mistake](https://www.ctvnews.ca/vancouver/article/air-canadas-chatbot-gave-a-bc-man-the-wrong-information-now-the-airline-has-to-pay-for-the-mistake/), CTV News
13. [Reduce hallucinations](https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations), Anthropic
14. [Introducing Contextual Retrieval](https://www.anthropic.com/news/contextual-retrieval), Anthropic
15. [Retrieval](https://platform.openai.com/docs/guides/retrieval), OpenAI
16. [Use contextual grounding check to filter hallucinations in responses](https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-contextual-grounding-check.html), Amazon Web Services
17. [Grounding overview](https://cloud.google.com/vertex-ai/generative-ai/docs/grounding/overview), Google Cloud
18. [Security in Foundry IQ (Azure AI Search)](https://learn.microsoft.com/en-us/azure/search/search-security-overview), Microsoft Learn