Company knowledge base AI: our weekly reviewed digest

On Monday a new price is agreed in a WhatsApp chat. On Wednesday a colleague writes down changes to a work procedure in their own notebook. By Friday the rest of the team still does not know what has changed.

AI can help prepare a weekly digest of changes from the company's shared notes. Each item gets a link to its source, and the responsible employee reviews the digest before it is published.

This article describes that setup: how it works and where it stops. It gives no results or figures of our own. It describes a routine, and it is not a case study.

Why company notes stop being useful

An internal knowledge base gathers company information in a single location, so anyone who needs it can reach it.² That is a good start, but it is not the whole job.

Putting files in one folder only keeps them as they are. Storing files alone preserves their content as-is without evaluating whether it is fit for retrieval.³ A vaguely named note with decisions buried inside it may be hard to find later.

Decisions can end up in different notes and conversations. When knowledge is scattered across workplace tools, retrieval-augmented generation can be harder to execute.¹ It is harder for people too. They ask the colleague who was in the meeting, and that colleague answers the same question several times a week.

A shared document with a changelog can do the same job if everyone keeps it up. The weekly digest helps when changes sit in many notes and nobody has time to read through them all.

What we keep, and in what shape

Consider keeping separate notes for decisions, how-tos and processes. For decisions, write down, for example, an agreement on a price, a new rule or a change in who is responsible for what. How-tos are the steps for a task that comes back, such as preparing the monthly report. Process notes describe how work moves between people. Give each note a date and name the employee responsible for it, so people know whom to ask if it looks wrong.

Write each decision down in a separate note and add the date. The system builds the digest from that note.

In the setup described here, a person does that copying by hand. Set the system to read only the shared notes folder, rather than chats, inboxes or spreadsheets. For example, when a repeat customer's new price is agreed on WhatsApp, the person who agreed it writes a short dated note, has it approved if needed and updates the shared Google Sheet themselves. The digest then reports the note to everyone. The system does not save that copying. It saves everyone else from asking, because the change reaches the whole team with a link to its source.

The copying is deliberate. Reading chats and inboxes directly would also pull in prices still under discussion and private customer messages, so a person confirms a decision is final by writing it down.

Think about how a system will find information within each note. AI systems that use chunking and semantic search benefit from clear organizational signals in the source material.³ A well-prepared knowledge base can improve retrieval precision, reduce the risk of unsupported answers and strengthen citation quality.³

Put each decision under a clear heading, with a date and the name of the employee responsible for the note. Use a heading that says what the note is about, such as "Discount rule for repeat orders" rather than "Notes from Tuesday".

How the weekly draft is put together

The system first finds the information it needs in the company's notes and then uses it to prepare the digest, adding links to the sources. This approach is called retrieval-augmented generation, or RAG, and it lets AI look up answers from internal sources.¹ Consider limiting the sources to approved notes.

In practice, the digest is made by a scheduled job, a small program set to run at a fixed time. Have the job select notes changed in the past week and pass them to a language model to draft the digest. Plan for someone technical to set up that job and the search tool. Your part is the notes, the reviewer and the weekly check.

Bloomfire's AI search, for comparison, gives answers with clickable citations.⁴ In the digest, the source link lets readers open the note and check whether the summary matches it.

The output is only a draft. The system may count a fixed typo as a change, or merge two related notes into one summary. It cannot report a decision that nobody wrote down.

Who checks the draft, and what they fix

Assign one person to review each digest. Have the reviewer correct vague summaries, remove items that are not decisions and approve the rest. If a source note is unclear, fix it too so the problem is less likely to recur.

Naming one person matters. If "the team" reviews, nobody does. A named reviewer also gives colleagues someone to ask when an item looks odd. Choose a deputy to cover absences.

The source links make the review faster and the result easier to trust. When people have to rely on an automated result, being able to see why it looks the way it does helps them trust it.⁵

Checking the digest can also reveal gaps in the notes themselves. First check items that lack a link to their source, or whose linked source does not support the summary. If the information needed is missing from the notes or unclear, ask the employee responsible to add to the notes.

In the first weeks, keep a short list of what the reviewer corrected, what the draft missed and how long each review took. That list is your own evidence of whether the routine is worth keeping.

Before starting, also count for two weeks how often people ask who knows what changed, and how many order or pricing mistakes trace back to an old note. Compare those counts after a month with the time spent on review. If the review costs more time than it saves, change the routine or stop it.

Keeping notes current without a big cleanup

Not every note needs the same attention. A sensible rule is to set how often a note is reviewed by how quickly its content changes: monthly for fast-moving topics, and every quarter or year for stable policies.³ Prices and stock rules belong in the first group. A holiday policy belongs in the second.

If notes are reviewed only once a month, a price change made on Tuesday can go unnoticed. So the person who agrees a new price should update its note the same day, and the next weekly digest shows the change to everyone. If the change matters before Friday, they should also tell the people affected directly, because the digest is a weekly record, not an alert.

Bloomfire, for example, flags outdated or duplicate content automatically.⁴ Generative AI can also speed up writing and revising content, so keeping notes current takes less manual work.² Leave the final call on changes to the employee responsible for the note.

What you need to set up and run this

The effort depends on how many notes you have, what state they are in and how many places they are spread across. Cleaning up a disorganised knowledge base, outdated documents and inconsistent tags takes real time and effort.¹ You do not have to clean everything first. Start with the notes people use and tidy others as gaps emerge.

A simple first month could run like this. In week one, the owner picks the notes people use most and names a reviewer and a deputy. In week two, the employees responsible for those notes add headings, dates and their names. In weeks three and four, the reviewer works through the first drafts and lists unclear terms for a glossary note.

The system also has to learn your own language: your acronyms, product names and the company's work processes.¹ A short glossary note helps.

If your notes mix Latvian and English, test that before relying on the digest. Take ten recent notes in each language, read the draft summaries against the source notes and list every term the system got wrong. Add those terms to the glossary note, and have the reviewer check Latvian wording closely in the first weeks.

Decide early who may see what. Keep customer details, and any prices or terms not meant for the whole team, out of the notes the system reads, or limit who can open those notes and the digest. Post the digest only in a channel limited to the people allowed to see its items, and have the reviewer remove anything that should not go further.

Plan staff time for the weekly review and for the fixes the employees responsible for the notes make. The pieces are usually a shared notes folder, a tool that can search it, the scheduled job that writes the draft and a place to post the digest. For a small team, consider using a shared space people already use. Each may bring its own setup work and subscription, so ask for them to be listed before you commit. Check the costs of the tools you already use, and ask for a quote for any additional work. Ask for one-off setup and monthly running costs to be shown separately.

Bloomfire offers AI search with clickable citations and automatically flags outdated or redundant content.⁴ You would still need to set up the weekly review routine yourselves.

If you are asked for an AI model or knowledge-tool comparison, first clarify what problem it should solve. An AI model comparison or a wider AI comparison of tools is easier to judge once you know which notes the team relies on. In our experience, what matters more than the choice of model is how well the notes are organised and how regularly they are checked. To plan a build like this, start by discussing the review process.

Frequently Asked Questions

What is a knowledge base in AI?

In general terms, a knowledge base is a body of knowledge kept in a form a program can use.⁵ Retrieval-augmented generation lets AI look up answers from internal sources.¹ Each answer links back to its source note so a person can check it.

Do we need to clean up all our documents before starting?

No. Start with the notes your team actually uses each week. Cleaning up messy, outdated or inconsistently tagged material takes real time and effort,¹ so it is better done step by step. Questions the system answers with low confidence can help you find content that is missing or badly organised.³

What happens if the weekly digest gets something wrong?

Nothing reaches the team until the named reviewer has approved it. Because each item links to its source note, a wrong summary is quick to spot. The fix goes into the source note as well as the digest.

Does this work with notes written in Latvian and English?

If notes mix Latvian and English, test whether the system handles them well. It needs to learn the company's own acronyms, product names and work processes.¹ In the first weeks, the reviewer should pay extra attention to Latvian wording and terms.