The answer existed. Nobody could find it, including the AI.
A 60-person B2B software company keeps its knowledge in a help centre, a wiki, shared drives, the CRM and years of chat threads. People ask the same questions every week, and the AI assistant answers from guesswork.
A representative example drawn from our hands-on experience. Company details are composite, and the figures are typical of this kind of project rather than from one named client.
The problem
Pricing lives in a slide deck, policies in a wiki, product limits in old chat threads. Three versions of the same answer exist. The AI assistant can't tell which is current, so nobody trusts it.
What we do
Choose one source of truth per topic, give every page an owner and a review date, connect the sources into one searchable layer, and have the AI answer only from it, with the source shown.
What changes
People and AI get the same answer from the same place. Support drafts need little editing, and questions stop bouncing around team chat.
One trusted layer between the sources and the work.
- Help centre
- Internal wiki
- Shared drive
- CRM notes
- Support tickets
- Team chat threads
- One source of truth per topic
- Every page has an owner and a review date
- Existing permissions respected
- Out-of-date pages flagged automatically
- Assistant in team chat
- Support reply drafts
- Sales call prep
- New-hire onboarding
The same question, before and after.
"Do we support single sign-on on the Growth plan?"
Yes, single sign-on is available on all plans.
- No source shown
- Out of date. It moved to the Scale plan in March
- A rep repeats it to a customer
No. Single sign-on is included from the Scale plan up. Growth customers can add it as a paid extra.
- Source: Plans and pricing, reviewed 12 days ago
- Owner: Product team
- If sources disagree, the assistant says so and sends the question to the owner
Sample content for illustration.
How it works.
- 1
Map where knowledge lives
List every source, who uses it and which ones people actually trust.
- 2
Choose one source of truth per topic
Pricing has one home, policies have one home. Duplicates are archived or pointed to the right page.
- 3
Give every page an owner and a review date
Out-of-date pages are flagged to their owner automatically, before anyone relies on them.
- 4
Connect the sources
Everything syncs into one searchable layer that respects existing permissions, so people only see what they're allowed to see.
- 5
Put answers where work happens
An assistant in team chat, drafts inside the support tool, account context inside the CRM.
- 6
Keep a person in the loop
AI drafts, people approve anything a customer sees. Questions the assistant can't answer become a list of gaps to fix.
The change in numbers.
| Metric | Before | After 90 days |
|---|---|---|
| Time to find a trusted answer | about 15 minutes | under 1 minute |
| Repeat questions in team chat | 40 a week | 12 a week |
| AI drafts sent with little or no editing | 20% | 70% |
| First response to customers | 6 hours | 45 minutes |
| Pages with an owner and review date | 15% | 90% |
| New-hire ramp time | 8 weeks | 5 weeks |
Representative figures.
Where this goes wrong
Connecting everything first. Feeding AI three versions of the truth produces confident wrong answers. The clean-up comes before the connection.
Ignoring permissions. An assistant that can read everything will repeat salary or contract details to the wrong person.
Who this is for
Teams who bought an AI assistant and are disappointed by its answers.
Teams answering the same internal questions every week.
Common questions.
Usually not. We start with the tools you already have and connect them. A new tool is added only if there is a clear gap.
That depends on the AI provider and plan. We check each vendor's data terms with you before anything is connected.
In this example, about six weeks: two to map and clean up, two to connect, and two to roll out to the team.
Related case studies
All case studies →Bought an AI tool that gives generic answers?
Show us where your knowledge lives today. We'll show you what it takes to make it usable.