---
title: Document processes before AI: a guide | Blue Door Software
description: AI and new staff both stall on unwritten know-how. How to capture how work is done, key facts and where data lives, so it stays when people leave.
canonical: https://bluedoorsoftware.co.uk/insights/document-business-knowledge-for-ai/
---

Know-how

# Your business knowledge is the bottleneck, not the AI

Blue Door Software · 22 November 2026

If AI is not helping yet, the usual explanation is that you need a better model. In our work with charities, and with small firms, that is rarely the real problem. The work still depends on things nobody wrote down: how a job is actually done, the facts people just know, and which system holds the current numbers. A new starter hits the same wall. So does AI.

When a long-serving caseworker or volunteer coordinator leaves, the process can collapse with them. Small businesses see the same thing when an office manager moves on. The aim is simple: when someone leaves, the know-how stays.

This article sets out what to capture, how to keep it current, and why we start every project by documenting the work. It is written for charity managers and for SME owners. Near the end we credit the essay that prompted us to write this down in public.

## What AI, and a new starter, need to know

There is a difference between the material in front of you and the understanding that makes that material usable.

The material is whatever you have at the moment you do a job: today's referrals, this week's case notes, the figures in a funder report or a finance system. It changes all the time. You do not need a second copy of it. You need a reliable way to get it when the job runs.

The understanding is slower to change, and much harder to replace. It is knowing which report a funder actually wants, why a referral goes to one team and not another, or which spreadsheet is trusted and which one is a leftover. Some of it sits in a policy. A lot of it does not. It lives in a few people's heads, in old emails, and in habits that never made it into a manual.

AI can only work from what you give it. If that understanding is missing, it will produce something that looks finished and is still wrong. A new member of staff will do the same, only more slowly. Capturing the understanding is not a side task. It is the work that makes both people and AI useful.

## Why writing it down now matters more than the tools

For a long time, organisations bought software and then tried to make their work fit it. The better path is the other way around. The people who know how the business runs should spend their time making that knowledge usable, so everyday jobs can be prepared carefully and checked by a person.

That does not mean you must document everything before any work can start. It means you treat the written record as something you build and look after, not as a one-off project that goes stale in a drawer.

We have been doing a version of this for years, because you cannot build a useful system on a guessed process. Documenting standard operating procedures and processes is the first input to our work, and it is something we sell in its own right. See our [process and SOP documentation](https://bluedoorsoftware.co.uk/process-documentation/).

## How the work is done

Write down the actual procedure, not the official version that nobody follows. For each recurring job you need:

- What starts it.
- Who does each step.
- Which systems and files they use.
- What "finished" looks like.
- Where it usually gets stuck.
- What a person must approve, and what they are allowed to change.

This is a standard operating procedure in plain language: short enough to use, specific enough that two people would do the job the same way. A referral that has just arrived, a quarterly funder return, a first reply to an enquiry, or a weekly management report are typical starting points. Our [process library](https://bluedoorsoftware.co.uk/processes/) shows how we set this out for everyday jobs, including the exact point where a person signs off.

If the procedure only exists as "ask Jane", you have a key person dependency. Jane's method is the process, and she is a single point of failure.

## The facts your team relies on

Beside the steps there is a second layer: the facts and rules people use without thinking. What a "priority" case means here. Which funding stream a piece of work sits under. Who the usual partners are. How you calculate a figure that appears in every board pack. The names of products, services and teams, and which of two similar records is the one you should trust.

These facts do not change every hour, but they do drift. A policy is updated and the old wording stays in a handbook. A team is renamed and half the documents still use the old title. Incomplete or outdated facts are often worse than a missing page, because people and AI will use them with confidence.

Write them down in one place, in the language your team already uses. Name an owner. Give them a review date.

## Where your data lives

The third piece is a map of current information. Not a copy of the data itself, but clear guidance on where the live version is, who may see it, and how to get it.

Case notes may live in one system, funder figures in another, and a contact list in a spreadsheet that was never officially adopted. The people who have been there a long time know which source to trust. Everyone else guesses.

Fast-changing data should be fetched when the job runs, not copied into a second store that will be out of date by the afternoon. The written record should say: for this quarter's open cases, use this report in this system. It should not paste last Tuesday's numbers into a document.

This is also where permissions belong. A procedure that tells someone to "check the case file" is unsafe if it does not also say who is allowed to open it.

## Keep it current

A handbook that is not maintained becomes a liability. The way to keep it honest is to treat gaps as a normal part of the work, not as a failure.

When a member of staff, or AI preparing a job, cannot find a step, finds two conflicting rules, or discovers that a field has been renamed, that gap should be flagged. Someone who understands the work then reviews the suggested change and approves it, or rejects it, before the written record is updated.

AI can help you notice what is missing. It should not be allowed to rewrite the official procedure on its own. A person who owns that process approves every change. Give each procedure and fact sheet an owner and a review date, and look at them when the work itself changes, not only once a year.

## Start with one recurring job

Do not try to document the whole organisation first. Pick one job that already repeats: a referral route, a funder report, a handover, an inbox. Write down how it is really done, the facts it depends on, and where the live data sits. Use that record on the next run. See where it fails. Fix those gaps. Then, if it is sound, move to the next job.

That is the same discipline we use on a free call. Bring one job. We will tell you whether it is a sensible place to start. You can also try our two-minute check, [Is this job ready for AI?](https://bluedoorsoftware.co.uk/tools/is-this-job-ready-for-ai/), or read our [12 questions to ask before your first AI implementation](https://bluedoorsoftware.co.uk/insights/12-questions-before-ai-implementation/).

Documenting the process is useful even if you never add AI. It is how a new starter is shown the work, and it is how you reduce key person dependency. If you later want AI to prepare the job, the same written record becomes the brief. That is the path from [process and SOP documentation](https://bluedoorsoftware.co.uk/process-documentation/) into [AI process improvement](https://bluedoorsoftware.co.uk/ai-process-improvement/).

## When someone leaves, the know-how stays

This is the pain we see most often in charities. A long-serving coordinator leaves, and the referral routes, the funder reporting and the unwritten case procedures go with them. The records may still be in the system. The know-how is not.

For a small charity, this is a business continuity problem as much as a staffing one. A business continuity plan for a small charity is incomplete if the only person who can run a vital process is about to hand in their notice.

Writing the process down does not replace the person. It means the next person, and any AI that prepares the work, can start from the same facts. In a case management system such as [Seren](https://serenonline.org/), the records, workflow steps, follow-ups and reports already hold a lot of the operational picture. The written procedure sits alongside that, so the "how" is not trapped in one head.

This piece builds on Harrison Chase's article [How to build a knowledge factory](https://www.langchain.com/blog/knowledge-factory), which argues that better knowledge, not a better model, is what makes AI useful inside an organisation. The examples and the charity emphasis here are our own, drawn from work with UK charities and SMEs.

## Questions

### How do you document business processes?

Sit with the people who do the work. Watch one real run of the job. Write the steps, the facts they rely on, and where the live data is, in the language they already use. Then try the write-up on the next run and fix what was wrong.

### What should a standard operating procedure include?

What starts the job, the steps and who does them, the systems used, what finished looks like, the usual exceptions, and the point at which a person must approve.

### How do you stop knowledge being lost when employees leave?

Write the process down before they go, if you can, and keep it current while they are still in post. Give each procedure an owner. Do not leave the "how" only in one person's inbox or memory.

### What is tacit knowledge, and can AI use it?

Tacit knowledge is the understanding people have but have not written down: why a job is done a certain way, which source to trust, what "good" looks like. AI cannot use it until someone makes it explicit. That is the point of this work.

### Do I need to document processes before using AI?

You need enough of the real process written down for one job. You do not need a complete manual for the whole organisation first. Without that first write-up, AI will guess, and guessing looks confident.

### How do I prepare my business for AI?

Pick one recurring job, capture how it is done, name the sources it may use, and decide what a person must still approve. Our [12 questions](https://bluedoorsoftware.co.uk/insights/12-questions-before-ai-implementation/) article and the [Is this job ready for AI?](https://bluedoorsoftware.co.uk/tools/is-this-job-ready-for-ai/) check are a practical starting point.

### What is key person dependency, and how do you reduce it?

It is when a process can only run because one person knows how. That person is a single point of failure. You reduce it by writing the procedure down, naming an owner, and using the written version in real work so it stays true.

### How often should SOPs be reviewed?

Standard operating procedures should be reviewed whenever the work changes, and on a planned review date so nothing is left to drift. A yearly tidy-up is not enough if a form, a funder or a system has already changed.

Free 40-minute call

## Bring one job to a free call

Bring the job that keeps coming back. A free 40-minute call by phone or Teams. We'll help identify one sensible first step and send a written summary of next steps.

[Ask for a free call](https://bluedoorsoftware.co.uk/#contact) [See what you get after the call](https://bluedoorsoftware.co.uk/what-happens-after-a-free-call/)
