> ## Content Index > Fetch the complete content index at: https://www.customersuccesscollective.com/llms.txt > Use this file to discover other available public pages before exploring further. # Why your knowledge base decides whether AI works in customer support - URL: https://www.customersuccesscollective.com/ai-knowledge-base-customer-support/ - Published: 2026-09-14T13:00:12.000Z - Updated: 2026-09-14T13:00:12.000Z - Description: AI can only retrieve, reason with and communicate the information it's given. That sounds obvious written down. It's somewhat less obvious when a vendor demo promises a chatbot that resolves 80% of tickets, and considerably less obvious once the budget's already been signed off. - Author: Grace Gupta - Tags: Customer support, AI and automation, Articles During a recent [Customer Support Summit](https://events.customersuccesscollective.com/), it was revealed that a VP had been pressed by their [C-suite](https://www.customersuccesscollective.com/chief-customer-officer/) about expediting their AI rollout: *“Can't you just make this thing work without a knowledge base?”* It's a fair question to ask if you've never built an [AI support deployment](https://www.customersuccesscollective.com/will-ai-replace-customer-support-video/). But it's also the wrong question, and the answer explains most of the difference between support teams whose AI actually works and teams whose AI keeps making things up. New research from our [**AI in Customer Support Leaders Report**](https://www.customersuccesscollective.com/ai-for-customer-support-leaders/), which surveyed support leaders alongside speakers from the Customer Support Summit series, backs that up with numbers: - Only **32%** of support leaders describe their knowledge base as fully AI-ready: structured, maintained and reliably retrieved. - **42%** say theirs is only partially there, meaning AI can find information but not consistently enough to trust. - **17%** say their knowledge isn't AI-enabled at all. That's a majority of teams asking AI to answer customer questions from a knowledge base that isn't built for the job. [![AI for Customer Support Leaders 2026 Report](https://storage.ghost.io/c/6f/66/6f66cab9-7939-4355-91c0-409dddffd4d9/content/images/2026/09/CSC_AI_for_Customer_Support_Leaders_Report_Assets_Article--1-.jpg)](https://www.customersuccesscollective.com/ai-for-customer-support-leaders/) ## Why knowledge quality matters, not the AI model AI can only retrieve, reason with and communicate the information it's given. That sounds obvious written down. It's somewhat *less* obvious when a vendor demo promises a chatbot that resolves **80%** of tickets, and *considerably* less obvious once the budget's already been signed off. **Megan O'Donoghue**, VP Global Support at Bazaarvoice, saw exactly how much the knowledge layer mattered once her team fixed it: > "When we pointed our bot at our KCS knowledge base, and we had been implementing KCS for over a year before that, the quality of our responses increased by about 80%." Bazaarvoice didn't switch AI platforms or upgrade to a newer model. The AI stayed the same. The information it was pulling from got better, and the results changed by **80%**. That pattern holds across the research: teams with fully AI-ready knowledge cluster in the highest deflection bands, while teams without AI-enabled knowledge sit stuck at 0 to 10% deflection. At this [stage of AI adoption](https://www.customersuccesscollective.com/ai-adoption-in-customer-success-teams/), the knowledge base is the ceiling everything else operates under, well ahead of any single other factor. ****Your knowledge base could be costing you 80% in AI performance** That's what fixing it did for Bazaarvoice. Get the full [****AI for Customer Support Leaders 2026 Report**](https://www.customersuccesscollective.com/ai-for-customer-support-leaders/) for the deflection rate data by AI-readiness tier, plus six leader interviews in full. [Download your copy ](https://www.customersuccesscollective.com/ai-for-customer-support-leaders/) ## What the AI knowledge base problem actually looks like It rarely looks like one big obvious gap. **Katherine Gutierrez**, Sales Operations Manager at Oticon, described the more common version: > *"We had duplicate* [*content*](https://www.customersuccesscollective.com/customer-success-value-led-content/)*, outdated content, content that no one had looked at in years, spread across departments with no single source of truth."* That's usually years of ordinary documentation habits, nobody's fault in particular, adding up into something AI can't reliably use. What changed wasn't a new tool. *"When we finally centralized it and built a structure agents could actually use, onboarding time dropped by half and escalations went down significantly,"* Gutierrez said. Fixing the structure paid off for [human agents](https://www.customersuccesscollective.com/human-first-approach-at-work/) before AI ever entered the picture, which is worth sitting with. Good knowledge infrastructure helps everyone using it, not just the bot. The research backs up where teams are actually putting effort. **57%** have restructured articles for [AI consumption](https://www.customersuccesscollective.com/ways-to-use-ai-in-customer-success/), adding clear titles and tagging. **55%** have implemented retrieval-augmented generation (RAG). **48%** have consolidated into a single source of truth. **38%** have built out a taxonomy for products and issues. Those are the four investments showing up most, and they're also, not coincidentally, the four things Bazaarvoice and Oticon both did before their numbers improved. **One in five teams** in the research **have made none of these investments.** That group is almost certainly sitting in the lowest deflection band, and it's unlikely to move without doing this work first, no matter which AI platform or model it switches to next. [![Free resources worth paying for: Templates, reports and expert content – no card required – with a Customer Success Collective Insider membership.](https://storage.ghost.io/c/6f/66/6f66cab9-7939-4355-91c0-409dddffd4d9/content/images/2026/06/Free-resources-worth-paying-for-Templates--reports-and-expert-content-----no-card-required.--2-.jpg)](https://www.customersuccesscollective.com/insider-membership-plan/) ## Problems with AI knowledge bases (and the fixes) **Olga Rais**, Senior Director of Technical Support Americas at Autodesk, named the failure mode that worries leaders most: > *"Hallucinations happen. The systems deliver the most probable next answer, and they don't know when they're wrong. That is the scary part."* 85% of support leaders in the research name hallucinations as their top AI trust concern, more than 20 points ahead of anything else on the list. There are actually two separate failures hiding under that one word. 1. AI answering confidently and incorrectly, which in a [B2B account relationship](https://www.customersuccesscollective.com/build-and-maintain-cross-functional-relationships-between-customer-success-marketing-product-and-sales/) can do damage that's hard to undo. 2. The second is AI refusing to answer when a good answer genuinely exists in the knowledge base, defaulting to an escalation that wastes everyone's time. Both trace back to the same root cause: [thin knowledge structure](https://www.customersuccesscollective.com/bad-customer-data-hygiene/) with no guardrails on top of it. The fixes aren't identical, but they start in the same place. **Burak Kebapci**, Sr. Director of AI Customer Support at Cardlytics, put the accountability question bluntly: > *"What failed was not the model. It was the boundaries. If you don't have proper guardrails, AI will confidently guess and send the wrong answer."* A capable model without guardrails just produces capable-sounding wrong answers, and knowing what the AI should refuse to touch is a leadership decision, not something to hand off entirely to an implementation vendor. [![AI-first customer support playbook template](https://storage.ghost.io/c/6f/66/6f66cab9-7939-4355-91c0-409dddffd4d9/content/images/2026/07/AI-first-support-playbook.jpg)](https://www.customersuccesscollective.com/ai-first-support-playbook-template-framework/) ## Successful AI knowledge base implementation **George Sullivan**, Senior Director of Support and Customer Education at Clio, described what the payoff looks like once the knowledge work is actually done: > *"We're getting 80% independent non-human resolution with the AI support agent. And because of the audit trail, we're able to go in and see where we have knowledge gaps. We've been able to work cross-functionally from support into customer success and customer education to figure out where our gaps are, because now we have insights we didn't have before."* The AI's own audit trail is now pointing Clio's team to where its knowledge is still thin, beyond cleaner retrieval alone, and that's feeding back into what gets fixed next. That's the pattern across every [high-deflection](https://www.customersuccesscollective.com/high-customer-support-ticket-volume/) team in the research: knowledge work first, deployment second, then a continuous loop where the AI's performance points to the next gap worth closing. Teams that reversed that order, deploying first and discovering their knowledge gaps through [customer complaints](https://www.customersuccesscollective.com/turn-customer-complaints-into-product-gold-dust/), ended up doing the same work anyway. They just did it later, at greater cost, with a failed pilot behind them that made [the next round of investmen](https://www.customersuccesscollective.com/how-do-you-prove-the-value-of-customer-success/)t harder to justify internally. [![State of Customer Success 2026: Get your copy](https://storage.ghost.io/c/6f/66/6f66cab9-7939-4355-91c0-409dddffd4d9/content/images/2026/07/CSC_State_of_CS_2026_Report_Assets_Article--1-.jpg)](https://www.customersuccesscollective.com/state-of-customer-success-report) ## Sort out your foundational base first, not after The teams furthest along in their [AI deployments](https://www.customersuccesscollective.com/ai-adoption-in-customer-success-teams/) didn’t necessarily move the fastest. Some might have. But they succeeded because they got their sequencing right. Every high-deflection team in the research we undertook did the knowledge work first, *then* deployed AI. The teams that skipped it got disappointing results. They then had to go back and build the foundation they should have started with, on a longer timeline and with less internal confidence than they had before the failed pilot. That's the expensive way to learn the lesson that one[ C-suite](https://www.customersuccesscollective.com/speaking-the-language-of-the-c-suite/) member asked about. You can, technically, make AI answer questions without a [proper knowledge base](https://slack.com/intl/en-gb/blog/productivity/what-is-an-ai-knowledge-base-tools-features-and-best-practices) behind it. It'll just make things up while it does it, and someone will notice before you'd like them to. If there's one place to put budget and headcount before the next AI rollout, it's here: - Restructure the articles - Consolidate the sources of truth - Build the taxonomy None of it’s glamorous, and none of it will show up in a vendor's product demo. It's also the single clearest predictor, across every team in this research, of whether the AI you deploy on top of it actually works. --- ## Get the full picture before your next AI rollout This article covers a fraction of what's in the [AI for Customer Support Leaders 2026 Report](https://www.customersuccesscollective.com/ai-for-customer-support-leaders/): deflection rates broken down by knowledge-readiness band, the complete interviews with Bazaarvoice, Oticon, Autodesk, Cardlytics and Clio, and the specific investments separating high-deflection teams from the rest. [ Get your copy ](https://www.customersuccesscollective.com/ai-for-customer-support-leaders/)