Start your knowledge base with the question you are working on today

You need to prepare a proposal for a new client. There was a similar project before: you remember a good solution and have a rough idea of where the presentation is. But the final terms were discussed by email, the calculation stayed in a spreadsheet, and why one of the options was rejected is something only you seem to know. Start with this current task and a few documents from the similar project, not with choosing a platform.

An hour later you have enough material to begin. Along the way it turns out that there are three presentations, the spreadsheet is called “final_2” and an important agreement will have to be confirmed with a colleague.

At a moment like this it is easy to conclude: we need an AI knowledge base so we can ask a question and get an answer straight away. The wish is understandable, but between it and a useful result there are several steps worth taking before you choose a platform.

Start with a question: what do you already know, where has it been kept and in which upcoming task should it prove useful?

The knowledge is already there, but the way to it is rebuilt every time

When people talk about a knowledge base, they picture a neat system: sections, instructions, project cards, search. Their own materials look more modest — folders, correspondence, meeting notes and a few documents that should have been sorted out long ago. But that is exactly where the most useful things may be: a well-worded proposal, an explanation of a difficult decision, the client’s comments, a list of mistakes you don’t want to repeat.

The problem starts when this experience can only be used by the person who remembers what to look for and where. Every time, they have to rebuild the connections between documents and explain to others which one is current. Below is a small practical step: how to make part of your accumulated experience available for a specific piece of work.

Choose a question worth opening the archive for

“Gather all of the company’s knowledge” hardly helps you see where to start: the volume is huge, the criterion for being done is vague, and current tasks have not gone anywhere. It is easier to take a question you actually need to answer. For example:

What from the past project can we use in the new proposal, and what needs to be recalculated or agreed again?

A question like this has a clear result: the proposal has to be prepared, and past experience should help make a few decisions. Now it is clear which materials you will need — not the whole archive, but the documents of one similar project: the agreed proposal, the calculation, the final deliverable and the correspondence containing significant changes.

A good first question limits the scope of the search and lets you check whether it was useful. If after the review you understand what to use as a basis, where new data is needed and what questions to ask the client, you already have a first result.

Put together a small working set

For a first attempt, a few materials that you can review and check yourself are enough. Take what may be passed to the chosen AI tool, given the access rights and the content of the documents. If the correspondence contains the client’s personal data, take into account the requirements of GDPR — the EU regulation on personal data protection — and remove anything unnecessary before uploading.

You don’t need to bring all the documents into a single format in advance. It is more useful to explain briefly what each source is:

Scroll horizontally to read the table.

These explanations are needed by both a person and the model. Two documents may contradict each other because one was created earlier and the other reflects agreed changes. At the same time, the newer file is not necessarily the approved one either. If you don’t know which version is final, say so — that in itself is a useful outcome of the review.

Ask AI to connect the materials to your task

Suppose the documents of the past project and the description of the new request are collected. Now ask AI to compare them, keeping three things apart from the start: what the materials confirm, which conclusions the model proposes and what is missing for a decision. Otherwise a real agreement and a plausible assumption will blend into one smooth answer.

Before the review, make sure the tool has actually gained access to the content of the files: mentioning a folder or giving a link to a restricted document does not guarantee this.

I am preparing a proposal for a new project and want to use the experience of a past one.

New request:
[briefly describe the task]

I have attached materials from the past project:
[list the documents and explain which versions were agreed]

First list the materials whose content is available to you.
If a document does not open or can only be read in part,
tell me.

Then help me work out:
1. Which decisions from the past project suit the new request?
2. What must not be carried over without checking, recalculation or approval?
3. Which pieces of information contradict each other?
4. What is missing to prepare the proposal?

For significant conclusions, give the source and the place in it: the section,
page or a short excerpt that lets me check the answer.

Keep information from the documents separate from your own suggestions.
If the materials do not contain the answer, say so directly.

At the end, draw up a list of next actions.

The prompt sets the way of working, but it does not replace checking: open the source reference too and make sure it really supports the conclusion.

Check the points the decision depends on

The model may suggest using the structure of the past proposal as a basis, recalculating the price separately and clarifying the support terms. Look at what these recommendations are based on.

Perhaps the scope of work really was agreed in the final document, while support was only discussed in the correspondence and never made it into the agreement. For a new proposal, that is a significant difference.

Be especially careful when checking amounts, deadlines, obligations and explanations of reasons. If the materials say that an option was dropped but not why, the model must not turn a guess into a story about the team’s motives. The answer “not found in the documents” can also be very useful: it shows which question to ask a person, instead of searching further through folders or deciding on the basis of an assumption.

Save what will help next time

After a good review it is easy to close the chat feeling that everything is clear now. But in a month you may need the same work again and will have to recall which documents you collected and what you concluded.

So save the result where you will continue the project: on the project page, in the task card or in a separate document next to the source files. It should include:

• decisions from the past project that can be reused;
• links to the supporting documents;
• terms that need to be reviewed;
• open questions and next actions.

Next to it, note the date of the check and a person who can clarify the content: if the terms change, it will be clear whom to contact and what to recheck. This gives you a small working piece of a knowledge base with a specific purpose — to help prepare the next proposal using the experience you have accumulated.

Expand the system once the work starts to repeat

The first task also shows whether you need a large system at all: where AI helps, which materials need to be put in order first and which result is worth saving.

After several reviews like this, you will see which materials are needed all the time, which questions keep recurring and where time is lost most often.

You may need a single place for agreed decisions. Or it may turn out that the documents are accessible, but nobody marks the current version. Or the main difficulty is that the reasons for important changes are not recorded anywhere.

These are different problems, and they will have to be solved in different ways. Such experience gives far more specific requirements for a future system than the wish to “build smart search across everything”. When a team uses the materials regularly, access rights, keeping sources up to date and responsibility for the content become important.

To start: choose one current task that would benefit from past experience. Formulate the question, put together a small set of materials, review it with AI using the prompt above and, after checking, save the result together with links to the sources. You don’t need a model archive. The test of usefulness is simple: is it clearer what to do next, and could another person — or you in a month’s time — carry on with the work without searching all over again?

Read next: where to save the result of an AI conversation so you can carry on working and how to hand over a task so that it doesn’t all stay in the manager’s head.

At a free 30-minute meeting with Simplit you can show where your materials are kept and what you keep having to search for again, and choose a first practical step. Discuss your task