Service · Internal knowledge hub

Internally, there should also
only be one version of the answerAn internal knowledge hub and the Q&A agent that answers from it

The customer-facing agent is on the e-commerce page; this one faces inward. Policies, experience and data scattered across chats, drives and laptops get collected into one place and become the single source of truth for every AI answer — staff ask it, the support agent cites it, content production draws facts from it. Built once, used in three places.

Book a diagnostic See what we've built

Knowledge source

Your own policies and data

Permissions

Follow the org chart

Live in

2–4 weeks

Ownership

Knowledge base and code yours

What's actually stuck

It isn't that the material is missing — it's that nobody knows which copy to trust

Four failure modes, and they get more expensive the further down you go.

Three versions of one answer

The policy is in a group announcement, the attachment is on a drive, and the newest revision is on somebody's laptop. Whoever has to follow it guesses, or asks.

Experience leaves with people

How to price this, which customer to avoid, which component fails first — never written down, held by a few individuals.

A handful of people get drained

HR, finance, IT and your best engineers answer the same questions daily. Their real work arrives in fragments between interruptions.

AI exists but nobody dares use it

Fear of a wrong answer, and worse, fear it hands salary or cost data to the wrong person. So only public material gets loaded — and then nobody opens it again.

One root cause underneath all four: there is no single place that settles how a thing is actually done here. Without that, more tools just add another location — and storage was never the problem. Findable, trustworthy and correct is the problem.

One distinction first

This is not a file share, and not another wiki

The three get conflated constantly, but they don't solve the same problem.

Solves “where do we put it”

File share / drive

The file exists, but nobody knows which one is current. Finding something means three folders deep, and then you still have to read it.

Solves “who writes it”

Wiki / doc site

Written by a few conscientious people, out of date within six months. You have to already know the page exists to find it.

Solves “what's the rule”

Knowledge hub

You ask, it answers, and it shows what it's basing that on. What it can't answer becomes a gap list, and the base grows itself.

The difference in one line: the first two send people to look for the answer; this one brings the answer to the person, and tells them what it's standing on.

What this line covers

One body of knowledge, six things using it

You can take the whole line or start with one piece. They share the same knowledge, which is why they know about each other — and why assembling separate tools doesn't get you here.

Knowledge hub

Policies, product facts, positioning and decisions in one place, one version each. It's the source of truth every AI answer resolves against — not another file share.

Internal Q&A agent

Ask a question in the tool your team already uses and get an answer with its source. When there's nothing to answer from, it escalates rather than improvising.

Tiered permissions

Follows your org chart and roles. Restricted content isn't retrievable, permissions move when people move, and every question is logged.

Self-updating loop

Unanswered questions, the human replies to them, and answers flagged as wrong all queue for review. Nothing takes effect until a person confirms it.

Multi-agent handling

Complex questions get split across retrieval, analysis and execution, then come back as one conclusion for a person to check — not a list of links.

An internal front door

Check stock, draft a document, raise an approval, generate a report. It sits on top of your existing systems rather than replacing them.

How each piece is implemented and in what order is written out line by line at the proposal stage, not on a public page — the same reasoning as pricing: proposing before seeing your situation serves neither side.

What it looks like in use

Three exchanges, each proving one thing

The second is the one to read. Without the permission layer, nobody will put anything valuable in— and a knowledge base holding only public documents doesn't get opened twice.

It answers, and shows its source

When the answer exists, the basis comes with it

New hire · day 3

Colleague

What's the accommodation limit for travel? I'm going out of region next week — can I book the one in the city centre?

Internal Q&A agent

Out-of-region travel is capped at 500 per night, rising to 700 in tier-one cities. Anything above that needs your department head's signature on the claim, or finance won't process it. The hotel you're looking at averages 680, which is inside the raised cap — you can book it directly.

Basis: Travel & Expenses Policy v4 · clause 3 · effective 2026-03 (v3 withdrawn)

What he shouldn't see, he can't retrieve

Same question, different answer by permission

Field sales

Colleague

What's our cost on the XX-300? The customer is pushing hard on price and I want to know where the floor is.

Internal Q&A agent

Cost data isn't within what you can see, so I can't retrieve it — and I won't estimate it. Within your authorisation: list price is 12,800 and you can discount up to 8% on your own (11,776). Below that needs regional manager approval. Want me to raise one now?

Restricted content isn't hidden after retrieval — it isn't retrievable · this question is logged

Person responsible

Approved at 11,200, go ahead. Next time flag anything below 12% before the customer pushes you to the wall.

Regional manager approval · pushed by the agent into the tool they already watch

When there's no answer, it says so

It doesn't guess at what isn't there

Admin · a recurring one

Colleague

I have leave left over this year. Can it be paid out instead?

Internal Q&A agent

There's nothing in the current policy documents about paying out unused leave, and I won't guess. I've logged it as a gap and routed it to the relevant person in HR.

Escalated · 5th time this quarter — flagged as a priority gap

Person responsible

Statutory leave not taken is paid at 300%; discretionary company leave expires. I'll add that to the policy.

Human answer · once confirmed it enters the knowledge base, and next time the agent answers it

Illustrative exchanges, rewritten and anonymised. During the diagnostic we run a round against your own policy documents and real questions, so you can see what it looks like in your business.

Permissions follow the org chart

One question, three people, three answers

Restricted content isn't found and then hidden — it isn't retrievable for that person at all. Permissions move when people move, and every question leaves a trail.

The same question

“On this XX-300 deal, how low can I actually go?”

Field sales

Own accounts · within authorised discount

List price 12,800
Discretionary discount to 8%
Below that requires approval
Common objections and how to answer them

Regional manager

Own region · including allowance

Everything above
Regional floor price 11,200
63% of this quarter's discount allowance used
This customer's order history and payment terms

Finance and leadership

Full · including cost

Everything above
Landed cost and minimum margin
Stock turnover and ageing for this model
Approval trail for below-floor deals

How many tiers, who belongs to which, and what never enters the system at all — all of that is yours to set. The example uses a sales scenario; payroll, cost and margin, contracts and legal follow exactly the same logic.

Where the knowledge comes from

Seven kinds go in, and the last one is the hardest

The first six collect things that already exist. The seventh extracts what has never been written down.

Policies and processes

Attendance, expenses, approvals, procurement, safety. The versions scattered across group chats and individual laptops get collected first, then reconciled to one.

Product and technical material

Specifications, configurations, compatibility, common faults and fixes, and which drawing or manual belongs to what.

Positioning and objection handling

How to quote, how to answer price pressure, how you compare against competitors, why deals were won and lost. The part that lives in your best salesperson's head.

Contracts and templates

Standard clauses, what's negotiable and what isn't, document templates and how they're meant to be filled in.

Meetings and decisions

What was decided, why it was decided that way, and who decided it. The things nobody remembers six months later.

Data in your systems

Orders, stock, customers, tickets — these aren't copied into the knowledge base. They're read live, under permission, so they're never stale.

What the veterans know

The judgement that exists in one person's head and has never been written down. We surface it through structured interviews and turn it into retrievable entries — the hardest part of this, and the most valuable.

Knowledge expires

So it has to grow itself — but every entry needs a human nod

A knowledge base built once becomes next year's pile of expired documents. We don't do unattended writes — that contaminates the base with wrong answers, and nobody can tell afterwards when it happened.

Queued automatically

automatic

Unanswered questions, the human replies to them, and anything flagged as a wrong answer collect into a review queue.

A person confirms

required

Someone you nominate looks at it: is this right, and who should see it. Only then does it take effect, and the record shows who confirmed it.

Next time it answers

automatic

Once live, everyone can just ask. A person answers once; the machine learns once.

What that actually costs: usually ten or fifteen minutes a week. It's the one habit this system needs long term, and it's the line between it surviving year two and not.

Ask, and then get the thing done

Once you know how something is meant to work, the next step is usually to go and do it in some other system. The switching back and forth between those two is itself a large amount of wasted time.

Ask, then get the thing done

Look it up, draft it, raise it — in one sentence

Regional lead

Colleague

Which models declined in the eastern region last month? For the worst one, draft me a review summary for the product team.

Internal Q&A agent

Three models declined month on month in the eastern region: XX-300 (-22%), XX-120 (-9%), YY-40 (-6%). XX-300 fell furthest, and after-sales tickets for it went up fourfold over the same period, concentrated on one component. I've drafted the review summary on your template, addressed to the product team and the after-sales lead — send it?

Figures read from your own business systems · as of yesterday 24:00 · draft unsent, waiting on you

Person responsible

Send it, copy me. Put the fourfold ticket thing in a separate meeting.

A person confirms before anything goes out — the machine doesn't press send on outbound

Illustrative exchange, rewritten and anonymised. Which systems can be connected and what can be actioned is confirmed item by item during the diagnostic, against the systems you actually run.

One thing to be clear about: it doesn't replace your existing systems

Inventory, finance and approval systems remain the owners of their own data. We don't touch that and we won't push you to replace them. The knowledge hub is a layer on top — so nobody has to remember which back office this particular thing lives in. How far it can reach depends on whether those systems expose interfaces; we check each one during the diagnostic and say plainly where the answer is no.

Why it's worth building once

The same knowledge, facing inward and outward

This is what buying three separate tools can't give you: their three stores don't know each other, so one change means three edits, and one of them always gets forgotten.

Internal answers

Staff ask about policy, product and process, and get answers with sources.

Customer support

The support agent answers customers from the same policies and positioning, so the wording matches in every language.

See e-commerce automation →

Content production

Articles and product copy draw facts from the same place, so they can't contradict what support is saying.

So a change to one return policy updates what staff are told, what support says and what the articles claim, at the same time. It's also why this is worth doing first: it's the foundation the other two lines stand on.

What stays under your control

Three things that have to be in your hands

Otherwise it's just another black box you don't govern — which is no different from a subscription.

The knowledge itself

Entries live on your servers. Adding, editing or withdrawing one doesn't involve us, and the whole base exports at any time.

Who can see what

How many tiers, who sits in which, and what never enters at all — you decide, and changes take effect immediately.

The boundaries

Which questions always go to a person, and which answers need a human nod before they go out. You set that; it isn't hard-coded.

Timeline and how it's billed

Live in 2–4 weeks, with maintenance and new work billed separately

The diagnostic happens before the project starts and isn't counted in the 2–4 weeks. Whether it's two or four is mostly decided by how smoothly week one goes — whether current versions can be identified and permissions cleanly drawn.

Diagnosticbefore kickoffRemote for 60–90 minutes, or 1–2 days on site: how your process actually runs, who holds what material, who gets asked the most. Produces a deployment list and a proposal, with the fee credited in full against later work.

Week 1your people are neededCollecting material, identifying which versions are in force, drawing permission boundaries, and interviews with key roles. Half a day to a day from one person per function. This week can't be skipped.

Delivery2–4 weeksDelivered in blocks, starting with wherever the questions concentrate. Each finished block is usable immediately — you don't wait for the final day.

Handoverat go-liveSource code into your repository, along with the knowledge base, permission configuration and an operations manual, plus a training session. From that day it runs without us.

Maintenance · in-house or by subscription

Adding entries, adjusting wording, changing permissions, keeping up with versions, routine checks and incident response. Your own technical people can carry this, or you can buy an annual maintenance subscription.

New work · scoped as a new project

A new agent, a new business system connected, a new line of business. That's new development — folding it into a maintenance fee either inflates the fee or means it gets done carelessly.

Project pricing depends on scope and scale, and is quoted after the diagnostic.

What you end up with

A knowledge asset that keeps growing

Even if you stop using our system later, the organised knowledge is still yours — exportable and portable.

An organised knowledge base

One version each, clearly structured, exportable in full. The longest-lasting thing this project produces.

The internal Q&A agent

Wired into the tools you already use, with permission configuration and escalation rules.

The source code

Into your own repository. Another engineer familiar with this class of tooling can take it over.

Run records

Who asked what, which entry an answer cited, who confirmed which new piece of knowledge — all traceable.

Not included

✕Writing your policies for you: we don't create rules, we make your rules findable and traceable
✕Scanning and digitising paper archives — negotiable separately at volume, but not part of this line
✕Replacing your existing business systems: they remain the owners of their data
✕Legal and compliance review of content: correctness of what goes in rests with the people you nominate
✕Fixing management: settled knowledge doesn't mean a settled process
✕Unattended operation: confirming new knowledge and granting permissions are human checkpoints we won't remove for you

When it isn't worth doing

✕Policies still changing weekly, with nobody able to say which version applies: settle the rules first
✕A dozen people in one room who can just turn around and ask
✕Nobody willing to spend the half day on week one's inventory and interviews: then this will fail
✕You just want a chat box over search: an off-the-shelf tool is cheaper, and we'll say so

Questions

What people ask about this one

The first is the most common, so it's first.

01Our material is a mess — group chats, shared drives, people's laptops. Can this still work?+

Yes, and that's the normal starting point. No company arrives with tidy documentation, and collecting and organising it is inside the delivery scope rather than a prerequisite. What you have to supply is different: identifying which version is the one currently in force. Only you can decide that, and it's most of what week one is. The situation that genuinely doesn't work is different — policies still changing weekly, with nobody able to say which version is being followed today. We'll say so plainly and suggest settling the rules first.

02Could it hand salary data or cost prices to someone who shouldn't see them?+

No, and the mechanism is stricter than hiding: restricted content isn't retrievable for that person in the first place, rather than found and then withheld. Permissions follow your org structure and roles and move when people move. Every question is logged — who asked what, when, and what the system said. This layer isn't an add-on, it's the precondition: without it, nobody will put anything genuinely valuable into the system.

03Will it make things up? Can we check where an answer came from?+

Every answer has to cite something in the knowledge base, with the source shown next to it — which document, which clause, effective when. When it can't find an answer it says so and escalates. That makes correctness checkable: either the source material itself is wrong, or the citation is wrong, and both are traceable. Questions with real consequences — money, personnel matters — can be configured to always go to a person, with the machine not answering at all.

04Could our internal material leak, or be used to train someone's model?+

No. The whole system is deployed on your own servers or private cloud, and the knowledge base, retrieval records and question logs all stay inside your boundary. We don't operate it for you and we don't keep a back-door account — after handover we can't get in either; you grant temporary access if you want our help. This is also why we only do self-hosted work and don't sell a subscription: internal knowledge is the most sensitive thing in this kind of project, and it shouldn't live on someone else's servers.

05Where do people actually use it? Another app, another admin panel to learn?+

Neither. The entry point sits inside the messaging tool your team already uses — you ask a question, that's it. Our rule is not to give people a second place to log into. There's a web interface as well for anyone doing heavier research or curation, and it can be wired into your existing office systems; how far is decided during the diagnostic.

06How much of our people's time does launch cost?+

It's concentrated in week one: identifying which versions are in force, drawing the permission boundaries, and a handful of interviews with key roles to get the judgement that only exists in a few heads. Typically half a day to a day from one person per function. After that it's light — ten or fifteen minutes a week confirming new entries in the review queue. Week one can't be skipped. Skipping it means building the knowledge base out of expired documents, and it starts losing people's trust on day one.

Start by finding out who gets asked the most

A diagnostic maps where your material actually lives, who is being interrupted for answers, and which questions repeat often enough to be worth solving. The fee is credited in full against the project.

Book a diagnostic Start with one module instead