Beginner's guide to AI knowledge bases

Week two for your new starter. A client asks whether they can share screenshots of the dashboard in a case study, and she knows there's an NDA somewhere in Notion. She searches "NDA". Nineteen results, four of them templates, two of them from a client who churned in 2023. "I'm sure someone dealt with this last year. Was it on Slack?" She asks in #general instead, and someone answers from memory, confidently, forty minutes later.

That's the gap a no-code AI knowledge base fills. Strip the marketing off it and the thing is simple: you point it at documents you already have, someone asks a question in plain English, and they get an answer in seconds with a link to the paragraph it came from. The answers are constrained to your material rather than the open internet (the technical word for that is grounding).

The "no-code" filler just means that you don't need to be technical whatsoever. You do the setup yourself in an afternoon. Connect the places your documents already live — Google Drive, Dropbox, Notion, direct file uploads, a few website URLs — put your logo and colours on it, decide who's allowed in, publish it to a URL. Nobody writes anything. Nobody raises a ticket with IT.

You've probably got good documentation. That's rather the problem: there's a lot of it, it's spread across four tools, and searching it rewards people who already know where the answer lives.

Why this isn't a help centre with a chat box on it

A wiki and a help centre both do the same job — they store things, and they match the words you typed against the words on the page. Which works right up until the question isn't phrased the way the document is.

Say you install boilers. A customer emails: does the warranty still stand if the boiler was fitted by a subcontractor? Your engineer searches "warranty", gets the 40-page installation terms PDF, and starts scrolling. The word "subcontractor" appears twice, once in a section about invoicing. The answer is in clause 9.3, phrased as "works carried out by an approved third-party installer". Keyword search never had a chance, because the customer and the document don't use the same nouns.

An AI knowledge base handles that differently. It looks for passages that mean the same thing rather than passages that spell the same thing, pulls the two or three that are actually relevant, and writes an answer from them with the clause reference attached (retrieval first, then answer — the industry calls this retrieval-augmented generation, and you can safely never say that out loud).

The citation is the whole point. Without it you have a confident paragraph from an unnamed source, which is worth less than nothing to anyone whose job involves being right.

The adoption numbers say something odd

AI use across UK businesses with 10 or more employees hit 35% in June 2026, nearly triple the rate in September 2023, when it sat at around 12%. Large language models were the most-used technology of the lot, at 18% [1].

Then the same data set undercuts itself. Only 10% of those businesses describe their AI use as 'extensive', and the average adopter is using 1.6 AI technologies, up from about 1.4 in September 2023 [1]. Nearly everyone has tried something. Almost nobody has built anything.

The US picture explains why. Business AI use ran between 17% and 20% from December 2025 to May 2026, but split by headcount it stops looking like an average at all [2].

“By firm size, 37% of firms with 250+ employees and 32% of firms with 100-249 employees reported current AI use, against a national average of 19.8%.”

US Census Bureau

Business Trends and Outlook Survey, 2026

Building a grounded, branded, access-controlled knowledge base used to mean a discovery phase, a data engineer, a security review and a number with a comma in it — so the firms with 250 people did it and the firms with 30 didn't. That constraint has quietly gone. If you can create a Notion workspace and invite your colleagues to it, you can stand one of these up before lunch and delete it if you don't like it.

Which raises the awkward question for anyone currently scoping a project: what exactly are you buying six weeks of consultancy to find out?

"Won't it just make things up?"

Microsoft analysed 105,000 Copilot chats in February 2026 and found 15% of them were people trying to find information. In the same research, 86% of surveyed workers said they treat AI output as a starting point rather than a final answer [3].

That instinct is correct and it's also a tax. Every answer your team has to go and verify from scratch costs most of the time the answer was supposed to save. Cited AI answers are what collapses that: the engineer reads the sentence, clicks through to clause 9.3, sees it in context, and replies to the customer. Ten seconds of checking rather than ten minutes of hunting.

So when you trial something, don't test it on questions you don't know the answer to. Test it on five you do, and one where you know two of your documents contradict each other. The second test is the useful one. A tool that cites both versions and shows you the conflict is behaving properly. A tool that picks one and sounds certain is telling you something about how much scrutiny it will survive.

“Analysis of 105,000 Copilot chats in February 2026 found 15% support finding information, alongside 49% supporting cognitive work, 19% working with people and 17% producing work. 86% of surveyed workers treat AI output as starting points, not final answers.”

Microsoft

2026 Work Trend Index Annual Report, 5 May 2026

The parts we'd warn you about

Most of these tools import a snapshot rather than maintaining a live connection. If you rewrite the returns policy on Monday morning, the assistant is still answering from Friday's version until someone re-imports it. Worth knowing before you tell 60 people it's authoritative.

Second, contradictory documents. If there are two expenses policies in Drive — the 2024 one and the one HR rewrote in March — you will get answers drawn from both, correctly cited, and the citation is the only reason you'll notice. An AI knowledge base surfaces your documentation problems rather than fixing them. That's genuinely useful, and it is not the same as being solved.

Third, scope. One assistant pointed at every file you own gives worse answers than three assistants pointed at HR, sales collateral and support docs respectively. Narrow beats comprehensive, every time.

And if the answer to your top question lives in one person's head and nowhere else, no software helps. Deloitte and eGain found that 92% of organisations they surveyed still fail to consistently capture knowledge from soon-to-be retirees [4]. Write it down first. Then index it.

Is an AI knowledge base safe? Ask about the plumbing, not the model

This is where most evaluations go wrong. The worry people voice is "could the AI leak our data", and the evidence points somewhere much more boring.

IBM found that more than 20% of organisations reported a breach targeting AI models or applications, and the two most common causes were compromised APIs, applications or plug-ins, and cloud misconfigurations affecting AI workloads — 27% each [5]. Neither of those is a clever model doing something unexpected. Both are ordinary infrastructure failures masked by AI.

Which makes AI knowledge base security a procurement conversation you already know how to have. Is the provider ISO 27001 certified, and can they show you the certificate? Is your content used to train models — the answer you want is a flat no, in writing. Is data encrypted in transit and at rest? And who can actually get into the assistant: a public link, anyone with an @yourcompany.com email address, or a named list you control? Pick per assistant. The HR one and the customer-facing one should not share an answer to that question.

Sources

[1] Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, 2026 — https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026

[2] US Census Bureau, Business Trends and Outlook Survey (BTOS) — 'Large Firms With at Least 20 Employees Biggest AI Users', 2026 — https://www.census.gov/library/stories/2026/05/ai-use-businesses.html

[3] Microsoft, 2026 Work Trend Index Annual Report, 5 May 2026 — https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization

[4] eGain and Deloitte, 'The $9 Trillion Knowledge Exodus', 22 June 2026 — https://www.globenewswire.com/news-release/2026/06/22/3315460/0/en/eGain-and-Deloitte-Publish-Joint-Research-and-Recommendations-on-the-9-Trillion-Knowledge-Crisis-Facing-Enterprises.html

[5] IBM, 2026 Cost of a Data Breach Report, 29 July 2026 — https://newsroom.ibm.com/2026-07-29-ibm-study-one-in-four-malicious-breaches-are-ai-enabled,-costing-companies-6-million-on-average

[6] European Commission, AI Act — Regulatory framework for AI, 2026 — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai