A company knowledge base as a service: why it costs more than it looks
A client thinks the work is technical: upload the documents, connect it, done. The technical part is the smaller one. Below is what makes up the main part, why the contradictions found sometimes get valued more than the assistant itself, and how an internal version differs from an external one.
A client thinks the work is technical: upload the documents, connect it, done. The technical part is the smaller one. Below is what makes up the main part, why the contradictions found sometimes get valued more than the assistant itself, and how an internal version differs from an external one.
---
The same questions ten times a day
Every company of more than five people has a set of questions that get asked constantly.
How do I process a return. What are the terms on this plan. What do I write to a client asking to move a deadline. Where is the current price list. Who approves requests like this.
The answers exist. They sit in procedures, in message threads, in the heads of two or three people who get interrupted all day because of it.
A new hire takes months to get going for exactly that reason: not because the tasks are hard but because they do not know where anything is.
The solution looks obvious: an assistant answering from the company's documents. But between "looks obvious" and "works" lies the part of the work people pay for.
What a knowledge base means here
Not a folder of files.
It is a connected source the assistant searches before answering. Somebody asks in their own words, the assistant finds the right place in your documents and answers from it.
The key difference from an ordinary chat: it does not retrieve the answer from the model's memory. It answers from what you put in. When all the material has been supplied by you, there is nothing to invent — the most reliable mode of working that exists.
Hence the other side: what is not in the base, the assistant does not know. And if it is not told what to do in that case, it will improvise.
Why this costs more than it looks
A client usually thinks the work is technical: upload the documents, connect it, done.
The technical part is the smaller one. The main work is elsewhere.
The knowledge is scattered
The price list in one spreadsheet, the procedure in a two-year-old document, the current terms in the manager's messages, the exceptions in a senior colleague's head.
Gathering it is not exporting files but taking an inventory: what exists, what is current, what contradicts what.
Half the knowledge is unwritten
The most labour-intensive part.
What people know has to be extracted by interview: what do you do, in what order, what goes wrong, how do you handle exceptions.
The person does not have to write anything — they have to answer questions. That is many times faster than the request "please describe how you work", which usually gets answered a month later or never.
The documents were not written for retrieval
A forty-page procedure in solid text searches badly. It has to be broken into meaningful chunks, each answering one question.
That is reworking rather than uploading. And it is what determines whether the assistant gives precise answers or general ones.
Contradictions
Almost always turn up: the terms of service say one thing, the procedure another, the staff a third.
Finding the contradictions is a by-product clients sometimes value more than the assistant itself. It is what you bring them for nothing, and it is an excellent argument in the conversation about price.
How this works technically
Briefly, so you can explain it to a client.
The "knowledge base" block in a chain — retrieval across the uploaded documents. The assistant gets a question, finds the relevant chunks and answers from them.
The "document analyst" block — turning a file into text, if the documents are not textual.
Behaviour rules — what to do if the answer is not in the base. A compulsory point: say plainly that the data is insufficient and offer to refer to a person. Without that rule the assistant will fill the gap with something plausible.
An updating rule — who refreshes the base and when.
Two different products
Worth separating, because the prices and the work differ.
An internal assistant for staff
Who uses it: your team.
What it solves: a new person gets going in days rather than months. Experienced people stop being interrupted over trifles.
What is easier: the audience is friendly and an error is not critical — a colleague will ask again.
An external assistant for customers
Who uses it: buyers.
What it solves: enquiries get handled round the clock, standard questions get closed with no human involved.
What is harder: the cost of an error is high. An assistant that gave a customer the wrong price or a non-existent condition creates a problem rather than solving one.
Two rules are compulsory here: do not invent and hand hot enquiries to a human. The sign of a hot one is a question about the price of an expensive item, about timescales, about individual terms.
Starting with the internal one is sensible. It is cheaper, safer, and it is where the base gets debugged before going into the external one.
What gets delivered to the client
A list worth showing in full — it justifies the price:
- The inventory: what exists, what is out of date, what contradicts what
- The gathered and reworked base
- The contradictions found, as a separate document
- The configured assistant with behaviour rules
- The rule for referring a question to a human
- Instructions for staff and training
- An updating procedure: who and when
The last two get skipped most often — and that is exactly why systems die within a month. An employee who did not understand routes around the system in a week. A base nobody updates starts lying within six months.
The price range
By my scrape of 848 job listings over 24 days: a system or pipeline — $490–1 460, against a $30–75 median for one-off work.
A caveat: an explicit budget appears in 6% of listings, with between two and fifteen observations per category.
What determines the price: the volume of documents, how much knowledge has to be extracted by interview, one department or several, an internal or an external assistant, whether support is needed.
What to compare against in conversation. Not the price of software but time: how many hours a month your experienced staff spend answering colleagues, and what a month of onboarding a new person costs.
Where the main money is: in support. Terms change, new questions appear, the base ages. That is a recurring payment and an honest one.
What is available on the platform
AI Workflow — the "knowledge base" block for retrieval across uploaded documents and the "document analyst" for processing files. Plus chains where each step's result feeds the next.
Scheduled runs for regular updating — the Full plan.
More than 140 AI assistants across 14 categories — for the interviews and the reworking of documents. HR will help extract knowledge from people, legal will check the wording. From the Basic plan.
The bot architect from the engineering category — a system instruction with behaviour rules, handover triggers and a restrictions block.
Building your own assistants with a separate base — the Creator+ plan.
Automation templates — if a link to the company's existing systems is needed. The Plus plan, advanced ones Full.
Registration is free and opens three days of full Basic access.
Where to start today
One action. Take a company you know — your own or a client's — and write out the five questions staff ask each other most often.
Then find where the answers are written down. Usually it turns out two of the five are written down nowhere, and another two are written down in two places differently.
Those four points are your first sale. Not "let's build a knowledge base" but "you have four questions the company has no single answer to".
---
What RAG is and why people pay for it
Module 12 covers this as the culmination of the course, and the phrasing there is direct.
A RAG agent is a model working from a company's private base: PDFs, websites, procedures, message threads. It answers from internal documents rather than from the internet — with no invention.
The phrasing from the course: you create an AI that knows the company's whole documentation by heart.
Why that is expensive. Large businesses want their own chat assistant that does not send data outside and does not make things up. Systems like that, by the course's description, fetch tens of thousands of dollars on the market.
The next level: agentic RAG
Ordinary RAG retrieves information and answers. It has one source, one attempt and no re-checking.
Agentic RAG from the automations archive works differently: it dynamically selects several sources, refines the query if the first result is weak, re-checks what it found and makes a second pass.
What that changes for a client. An ordinary knowledge base answers direct questions. This one handles complex ones where the answer gets assembled from several documents.
And the base that shows the scale
132 000 business cases — 260 000 files stitched into a single structure: for each company the pain, the solution and the software used. An example of what can be built on this approach.
Module 12 — the N8N plan; agentic RAG and the base — N8N and Full.
---
What to read next
[What an assistant for a department costs](/en/blog/assistant-for-department-cost) — how to price work like this.
[Anatomy of a system prompt](/en/blog/system-prompt-anatomy) — the technical part of the project.
[A process audit](/en/blog/processes-and-automation-audit) — what gets done before gathering knowledge.
[Local AI models for sensitive data](/en/blog/local-models-sensitive-data) — if nothing can leave the building.
[n8n and Make automation templates](/en/blog/n8n-automation-templates-guide) — the full guide to automations.