--- title: "Notion Integration: AI Agent Knowledge Base | JustAutomate" description: "Notion as the source an AI agent answers from: procedures, prices and client decisions, each answer linked to its page. Rollout in 2 to 6 weeks." url: "https://justautomate.ai/en/integrations/notion/" lang: en --- Data and back office # Notion integration knowledge your AI agent can answer from Notion is where your company knowledge lives: procedures, price lists, what you agreed with a client, notes from meetings. We wire it in so an AI agent answers from those pages and shows which page the answer came from. Along the way you find out which documents are out of date, because the agent repeats every mistake somebody wrote there, and which tables have outgrown Notion. Book a free consultation [See how the workshop works](https://justautomate.ai/en/services/process-workshop/) - 5 to 10 secondswhat BOND needs to answer from your knowledge, with a link to the Notion page it took the answer from - 100 recordswhat one query to a Notion database returns at once, at an average of 3 requests per second per integration. Limits from the Notion API documentation - 180 daysa typical review threshold: a page with no edits for longer than half a year goes on the list to confirm or retire - 2 to 6 weeksfrom the workshop to a working integration, depending on how many pages enter the index and what state they are in The problem ## Why does an AI agent repeat an out of date note from Notion? A new joiner asks about the discount for a large client. The answer sits in Notion in three places: on a price page from March, in a meeting note from June, and in a comment under the project page. Each one gives a different number. The one somebody finds first wins. Pointing an AI agent at that fixes nothing on its own. The agent does not know the price page has stood untouched since March. It reads what is there, turns it into a full sentence and passes it on with more confidence than the content deserves. An out of date document in Notion becomes an out of date answer, only faster and in nicer language. The second half of the problem is tables. Notion has databases with relations and rollups, so it is easy to keep projects, clients and orders in it, and up to a point that genuinely works. Then it starts to grind, and it meets the limits described in the [Notion API documentation](https://developers.notion.com/): one query to a database returns at most 100 records, and one integration gets an average of 3 requests per second, above which the API answers with a 429 error. A report built from three related tables is then several hundred calls and a few minutes of waiting. The third thing surprises most people during the first pilot. An integration in Notion does not see the whole workspace. It sees only the pages and databases somebody explicitly shared with it. That is good news for security and bad news for the first week of testing: the agent says it does not know about things that have sat in Notion for years, because the page with the answer was never connected. - The same procedure lives in three places and reads slightly differently in each - A price page from March answers a question asked in December - Nobody in particular owns the page, so nobody updates it - A report from Notion tables is assembled by hand, because the export cannot cope - The agent says it does not know, though the answer sits in Notion, just never shared Why now ## What changed, so Notion can be handed to an AI agent as a source? Notion exposes REST API v1, and that is enough to work in both directions: a query to a database returns records with their properties, POST /pages creates a page, PATCH /pages/{id} changes a property, PATCH /blocks/{id}/children appends content. Authentication goes through an integration token, and access is granted per page and per database. There is no all or nothing switch here. The second piece is the index. We pull pages with a query, split them into chunks, compute vectors for them and store those next to the text along with the address of the source page. That gives every answer a signature: not according to the documentation, but a link to a specific page and the date it was last edited. A person has something to check against, and whoever owns the content knows what to fix. The third piece is keeping up with changes. New and edited pages are picked up by a query filtered on last\_edited\_time, checked every few minutes, and only those go back into the index. You do not have to recompute the whole knowledge base once a day for an answer to keep up with what somebody added after lunch. The fourth is a way out for data that has outgrown Notion. Tables are pushed once a day to [BigQuery](https://justautomate.ai/en/integrations/bigquery/) or to a spreadsheet, and Notion stays the place people work and the view onto that data. The team works in the tool it knows, and the report is computed outside it, in seconds. What changes ## What does a team already living in Notion get? - ### An answer from company knowledge in 5 to 10 seconds Instead of hunting through pages, somebody writes the question as a plain sentence: what discount do we give above one hundred thousand, how does the return procedure work, what did we agree at the kick-off. The AI agent answers from your pages, not from the internet. Under every answer, a link to the Notion page and the date it was last edited - ### You can see which documents aged in silence Indexing produces a list of pages with no edits for half a year, pages with no owner, and pages that say the same thing as another one, only differently. It is usually the first time anybody sees that list in full. Owner and review date as page properties, a list to confirm after 180 days - ### One version of a procedure instead of three copies In the workshop we settle which page is the source of truth for each topic. The rest becomes a redirect or goes to the archive. The agent indexes only sources, so there is nothing to stitch two contradictory answers out of. Only pages named as sources enter the index, not the whole workspace - ### Notion fills up without manual copying A meeting transcript, what was agreed over email and a document from Google Drive land through the API on the right client or project page. Nobody pastes it in the evening, so knowledge does not stay in one person's inbox. Written through PATCH /blocks/{id}/children on a page that already has an owner - ### Notion tables in step with the CRM and with tasks A won deal in the CRM creates a project page in Notion with the properties filled in: client, value, owner, deadline. A status change in the operations tool updates the property in Notion, so the manager's view does not lie. Property updated through PATCH /pages/{id}, with no second system to log into - ### A report outside Notion, when the table stops coping Status history, month on month totals and reports that join several tables go to a warehouse or a SQL database. Notion stays the input for people and the view onto the result, not the engine that computes it. Export to BigQuery or Google Sheets once a day, the same view inside Notion And in your company ## What if half our Notion is out of date? Four sentences we hear most often about Notion, and what we answer to each of them. - Half the knowledge in our Notion has expired. That is the rule, not the exception, which is why a content review is part of the rollout rather than a precondition. The agent will not fix a document for you, so first we settle which pages are the source of truth, who owns them and what goes to the archive. If too little content is left after that conversation for an agent to answer from, we say so plainly. - We do not want AI seeing everything we keep in Notion. It will not. An integration in Notion is given access to named pages and databases, not to the workspace, so the scope is a decision rather than a side effect. Salaries, performance reviews and private notes stay outside the index, and the list of shared pages is one of the documents that closes the workshop. - We keep a project database with relations in Notion and it is enough for us. While it is enough, we leave it alone, and we will tell you so. The ceiling starts to show when a report joins three tables at once, when one of them grows by hundreds of records a month, or when you start measuring how long something sits in each status, which Notion does not record. Then history goes to a database or a warehouse, and Notion stays the view. - Notion has its own search, why would we need an AI agent? Search gives you a list of pages and leaves you to read them. The agent gives you a sentence with the answer in it plus a link to the page it came from. The difference is largest on questions whose answer sits in two documents, for example a procedure and the current price list. If the real problem is the structure of your Notion, we say so and start there. Knowledge base and documents Notion, Google Workspace, Microsoft 365 Team work and tasks Slack, ClickUp, Jira, Asana CRM HubSpot, Pipedrive, Salesforce, LiveSpace Data and reporting BigQuery, Google Sheets, GA4, SQL databases Content sources Meeting transcripts, email, Google Drive Automation layer n8n, Make, Zapier, custom API integrations ## What stays in Notion and what moves on The split we start from in the workshop. Notion is a good place for content a person reads, and a poor place for history somebody has to compute. | Kind of data | Notion is enough | When to move it on | | --------------------------------- | --------------------------------------------------------- | ---------------------------------------------------------------------------------- | | Procedures and instructions | yes, this is what it is built for | almost never, this stays the source for the agent | | What you agreed with a client | yes, one page per client, with an owner and a review date | when the agreement affects an invoice, then the CRM or the contract is the source | | Project database with relations | up to a few hundred records and a few views | when a report joins three tables or the table grows by hundreds of records a month | | Status history and time in status | no, Notion holds state, not the run of events | straight to a database or a warehouse, if you want to measure it | | Numbers for monthly reports | for simple totals inside a view | when you compare month on month, then BigQuery or a SQL database | | Personal and HR data | only what is necessary, and outside the agent index | when you need access control at record level | | Attachments and files | yes, as the place the team works | when a file is meant to be a source of answers, then its text enters the index | The rows are a starting point. In most companies the first rollout leaves everything in Notion and adds only the index plus one export to a warehouse. How we do it ## Five steps from pages in Notion to an answer with a signature Each step ends with something you can judge before you decide on the next one. 1. ### Workshop We go through your Notion and the questions the team asks most often. We settle which page is the source of truth for each topic, what is a copy, what is not in Notion at all, and which tables are quietly acting as the company database. 2. ### Cleanup and owners Duplicates collapse into one page, and every source page gets an owner and a review date as a property. Out of that comes a list of pages to confirm and pages to retire. This is work on content, not on code, and you can do it without us. 3. ### Sharing and the index We create the integration and share only the pages from the list with it. Content is split into chunks, vectors are computed, and both are stored with the page address and the edit date. Changes are pulled in by a query filtered on last\_edited\_time, every few minutes. 4. ### Pilot on real questions We collect 20 to 30 questions the team actually asks and check the answers one by one. A bad answer almost always points at a document rather than at the model: a missing page, two versions of a procedure, or a price list with no date. 5. ### Production and measurement The agent goes into the tool the team already sits in: Slack, Teams or the browser. After 30 and after 90 days we look at questions that got no answer and at pages nobody has touched in a quarter. That produces a queue of content fixes. Before and after ## How the day changes for the person who knows the answers ### Before - Answers the same questions again, in chat and at their desk - Searches Notion to work out which version of a procedure is current - Hears about an out of date price list from a client, not from a system - Builds the project report by hand, because the export loses the relations ### After - The team asks the agent and gets an answer with a link to the source page - Every topic has one source page, with an owner and a review date - Pages with no edits for 180 days land on the confirm list on their own - The report is computed in the warehouse, and Notion keeps the same view The first answers show up in the week the index is built. Quality is judged after 30 days on questions the team actually asked, and after 90 days on whether the content is being maintained. For the person signing this off ## What the person who owns company knowledge gets out of it You stop being the bottleneck. Questions about procedures, prices and past decisions stop landing on you, because the agent answers them from the same pages you signed off anyway. Onboarding a new joiner stops depending on whether you have time for them in the first week. The second thing matters more in a conversation with the board. After indexing you have it on paper: how many pages nobody has touched in half a year, and how many topics carry two contradictory versions. That is an argument for time spent on documentation, and today you have nothing to back it with. - A list of source pages with an owner and a review date, not in your head - The access scope for the agent is a decision written down in the workshop, not a side effect - A pilot on 20 to 30 real questions before anything goes to production - Questions with no answer come back as a queue of content fixes, after 30 and 90 days What it costs ## The workshop has a published price, the integration is quoted after it The process workshop costs **EUR 3,500 net per week**. On a knowledge base one week is usually enough, two when Notion serves several departments and each has to be walked through separately. We quote the integration after the workshop, once we know how many pages enter the index, what state they are in and which systems Notion has to talk to. Tool costs you pay directly to the vendors: Notion licences, the language model, hosting for the automations. No commission from us. Model queries cost the vendor EUR 0.005 to 0.07 each. With an index it is the number of pages that counts rather than the number of questions, so the usage bill does not grow with the team as fast as you expect. - ### Process workshop EUR 3,500 net per week, on a knowledge base usually one to two A self-contained piece of work. It ends with a list of source pages, named owners and a plan you can execute without us. - A review of your Notion and the questions the team asks most often - A list of source pages, duplicates and gaps, with owners - The data split: what stays in Notion and what goes to a database or a warehouse - The access scope for the integration, written down before any permission is granted Start with the workshop - Most popular ### Workshop and Notion integration Quoted after the workshop scope follows from the number of pages and the systems in play The option most companies choose. A working integration within 2 to 6 weeks of the workshop, on your pages and your data. - Everything in the workshop - An integration with permissions to named pages and databases only - Feeding Notion from the CRM, tasks, email and Google Drive through the API - Change capture on last\_edited\_time and a table export to a warehouse - Training for the person who maintains the content and the source page list Talk about a rollout - ### Notion as the knowledge base for BOND Quoted after the workshop a company AI agent rollout usually takes around 2 months Everything in the integration plus BOND, the company AI agent that answers from your knowledge and names the source page. - Everything in the Notion integration - An index of pages, with a link and an edit date under every answer - The agent in Slack, Teams or the browser, in English and in Polish - A report of unanswered questions as a queue of content fixes - Measurement after 30 and 90 days, with a list of pages untouched for a quarter Ask about BOND The workshop ends with a list of source pages and a cleanup plan that stands without us, including when you do not order the integration. If it turns out your Notion knowledge is too far out of date for an agent to answer from, we say so, and the first step is a content review rather than a rollout. ## Frequently asked questions Can an AI agent answer from our Notion? Yes, and it is the most common reason we wire Notion in. Pages are pulled through the API, split into chunks and indexed, and the agent answers only from what is in the index. Only pages named in the workshop as the source of truth enter the index, not the whole workspace. How do we know which Notion page an answer came from? With every chunk we store the page address and the date it was last edited, so under the answer there is a link and a date. That is a requirement rather than an extra: an answer with no signature cannot be checked, and with company knowledge being checkable matters more than being convenient. What happens if a document in Notion is out of date? The agent repeats what is written, because there is no source of truth other than your content. That is why a page review is part of the rollout: every source page gets an owner and a review date, and pages with no edits for 180 days go on a list to confirm or retire. Questions the agent got wrong come back to you as a queue of content fixes. Does the integration see our whole Notion workspace? No. An integration in Notion has access only to the pages and databases somebody explicitly shared with it, and that is written into the Notion API documentation. We settle the scope before any permission is granted, and salaries, performance reviews and private notes stay outside it. When does Notion stop being enough as a database? Usually with three symptoms: a report joins several related tables, one of them grows by hundreds of records a month, and you start measuring how long something sits in each status, which Notion does not record. On top of that come the API limits: 100 records per query and an average of 3 requests per second per integration. Then history goes to a SQL database or a warehouse, and Notion stays the input and the view. How do you pick up changes in Notion? With a query filtered on the last\_edited\_time field, checked every few minutes, so only pages touched since the last run go back into the index. On larger bases there is also a full recompute once a day, overnight, to catch deletions and permission changes. How long does a Notion integration take? A week of workshop and 2 to 6 weeks of rollout. The short end is a knowledge index for the agent across a few dozen pages, the long end is Notion wired to the CRM, tasks and a warehouse at once. Rolling out BOND as the company AI agent usually takes around 2 months, because further sources outside Notion come into scope. Does content from Notion leave our company? The chunks of pages needed to answer the question go to the language model, and we settle that in the workshop together with which vendor we use and where the index sits. Pages outside the list go nowhere, because the integration cannot see them. Personal and HR data stays outside the index by default. ## Let us check whether your Notion is ready for questions The free consultation takes 30 minutes. You tell us what you keep in Notion and what the team asks about most often, even if half the pages have not been touched in six months. We show you which topics can be indexed right away and which need one version named first. You do not have to share anything at this stage. We run at most two rollouts at a time, so that someone is actually sitting on each of them. [Book a free consultation](https://justautomate.ai/en/contact/) **PS.** If somebody at your company asked this week which version of a procedure is current, that is exactly this problem. An AI agent answers it in 5 to 10 seconds and names the page it took the answer from, but only when there is one such page. That is where the workshop starts. ## Related - [How to build an AI knowledge base](https://justautomate.ai/en/guides/how-to-build-ai-knowledge-base/) - [BOND, business AI agent](https://justautomate.ai/en/services/bond-business-ai-agent/) - [BigQuery integration](https://justautomate.ai/en/integrations/bigquery/) - [API integration](https://justautomate.ai/en/services/api-integration/) - [Slack integration](https://justautomate.ai/en/integrations/slack/) - [Process workshop](https://justautomate.ai/en/services/process-workshop/) --- **JustAutomate** · contact: contact@justautomate.ai · form: https://justautomate.ai/en/contact/