How to create an AI knowledge base that improves your average resolution time

It's a Friday afternoon, a customer asks about your refund policy for a discontinued product line.

The support rep knows the answer exists somewhere. They begin going through the sources that are front of mind. "I'm sure it was asked a few months back. Didn't we discuss this on Slack at some point? Or was it in that product email thread?". They open four tabs. They ask a colleague. Five minutes later the customer is now clearly frustrated. What could have been handled by a simple query to an AI knowledge base has suddenly butchered your Average Resolution Time (ART) and stressed your rep.

This isn't a new problem - you've got great reps, good documentation and excellent training procedures. Unfortunately your company's knowledge grows and updates faster than reps can learn it, fragmented team knowledge can feel as if hunting for a needle in a haystack. It can harm your key metrics if reps take slightly longer to answer customers. It's even worse if they give outdated answers.

We now have the exact tools to fix the problem - building your own AI knowledge base that reps use to discover answers to nuanced questions in a matter of seconds. One controlled and audited endpoint for them to query built on your company's knowledge .

Humans are at the heart of good support, AI assistance gives them superpowers

We've all heard the pitch, AI-powered support flows can handle a large volume of tickets so you can stay on top of volume and even reduce staff count. Unfortunately the promise doesn't pay off. Take Klarna as the epitome of poor "AI driven" decisions.

Sebastian Siemiatkowski declared in 2023 that they were going to cut support staff in their hundreds in order to focus on AI first support, only to roll that back in 2025 because "As cost unfortunately seems to have been a too predominated evaluation factor when organising this, what you end up having is lower quality." AKA AI isn't a human, and humans want to talk to humans, not robots. Maybe it has something to do with the soul.

“41% of global companies name data access and integration as the single biggest obstacle to AI progress, more cited than any other barrier across every function surveyed”

Bain & Company

Bain Automation and AI Pathfinder Survey 2026

If you over estimate AI's ability to solve problems, you'll miss the open goal of utilising it as a copilot, human led, AI powered. Don't overfit the solution, focus on making the process better with AI. Building a cited, on-topic AI knowledge base to share with your internal team can massively boost your ART, giving your human support better velocity and leading to much stronger customer satisfaction.

One of the biggest barriers to get started with an AI knowledge base is knowledge that is scattered across your company. A knowledge base is only as good as the data it's fed with and solving this problem has to be the top of your agenda.

The stat we always come back to - 41% of global companies name data access and integration as the single biggest obstacle to AI progress, more cited than any other barrier across every function surveyed [1]. Support just feels it first, because support is where the questions arrive in real time and the clock is visible.

There's a few simple steps you can take to begin to fix this problem.

1. Audit where your answers actually live

Look at your tickets from the last month that are above your goal ART and identify where the answers came from. Normalise into a category "Another ticket", "Slack", "Knowledge base", "Undocumented knowledge" for example, start to build a picture of where the answers to the tickets live that take the most amount of time.

This audit gives you a great picture of what exists outside of your knowledge base and therefore the additional sources that your need to collect in order to give the knowledge base enough context to answer questions.

2. Bring the answers into a shared format

Most of you will have a knowledge base that is well maintained and easy to use. From the previous steps, you now understand where the gaps are and it's time to bring all of these sources into a format the knowledge base can understand.

Markdown is great for this as it's very well recognised by a large language model, can decipher importance based on heading size and cuts down token consumption because it's a text based format. They key is to know how to go from source to markdown, for example if you're exporting a list as a CSV of tickets you could get ChatGPT or Claude to parse the list and output markdown documents for each ticket with a good reference.

Once you have the documents, you'll want to create a spreadsheet that tracks each document so you can easily audit and check if knowledge needs to be switched out as it's outdated, or more information needs to be added.

To highlight the importance of this, Gartner recently published the first-ever Magic Quadrant for Knowledge Management Systems for Customer Service — formally recognising that retrieval, not authoring, is now the distinct piece of infrastructure worth evaluating separately.

“A European telecoms provider covering 10,000 contact-centre agents improved first-contact resolution by 37%, lifted NPS by 30 points, and cut new-hire ramp time by 50% after unifying its knowledge management.”

eGain and Deloitte Insights

The $9 Trillion Knowledge Exodus (2026)

3. Add your documentation to shared environment

Both Anthropic and OpenAI give you the ability to create projects and upload documentation into them, now that you have the markdown ready to go it's as easy as creating a project and dropping that folder of information into that project ready for users to query.

Once it's all set up, you have a couple of watch-outs:

• You'll have to pay a license fee for each user that utilises the chat, this starts at $15 per user on Anthropic and OpenAI (Stark Chat has unlimited users)

• Information is stored on American infrastructure - not necessarily a problem unless you have strict EU processing and GDPR requirements (Stark Chat has strict data processing rules and is ISO27001 certified)

• You can't limit number of queries or how many messages a rep can send on a query, this means that reps can rack up tokens costs inadvertently (Stark Chat solves this with rate limits)

• You're not able to control whether answers are cited or the LLM finds alternative answers (Stark Chat solves this with cited only answers)

• Updating is a manual process that takes human time to repeat step one and two (Stark Chat solves this with source integration and importing)

A shared AI knowledge base can be a powerful addition to your tools, build it right from the ground up and you'll see your ART on those trickier to answer questions improve materially.

Sources

[1] Bain & Company, Bain Automation and AI Pathfinder Survey 2026 — https://www.bain.com/insights/your-ai-budget-is-growing-your-returns-arent-heres-why/

[2] eGain and Deloitte Insights, The $9 Trillion Knowledge Exodus — 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