Most small businesses I look at use Notion like a garage. Everything's in there: SOPs, client notes, the content calendar, a hiring checklist somebody started two summers ago and abandoned somewhere around step four. Nobody can find anything, and nobody has a free Saturday to clean it out.
Until now, Notion's AI could help with that garage about five minutes at a time. You asked a question, got an answer and moved on, and anything bigger ran out of road.
On October 19, Notion rebuilt its agents so one can stay on a single job for more than 24 hours. It also added a model router called Auto that it says gets similar results to the expensive models at about 90 percent lower average cost. And if you're an existing Business or Enterprise customer, your AI usage moves to credits on Monday, November 2.
So you're getting a worker who doesn't sleep and a meter that's about to start running. I'd get both sorted before November 2, and below is exactly how I'd do it.
What actually changed
The plain English version, minus the launch party and the confetti.
Agents can run long. Jobs used to top out around five minutes. Now an agent can stick with one task for more than a day, and because sessions are queue backed and survive restarts, a hiccup at 2 a.m. doesn't wipe out the work.
Agents got their own computer. Each one has a browser and can write scripts that run as code, which Notion says is cheaper and faster than paying tokens to grind through the same work step by step.
Auto picks the model. Easy requests go to fast models and harder multi step work goes to more capable ones. If a model stalls, Auto swaps it out. Notion says that across 20+ use cases it delivered results similar to Opus 5.5 and Astra at about 90 percent lower average cost, so credits go roughly ten times further.
There's a second set of eyes. A mode called "Review for me" uses a separate model to check actions before the agent takes them. More on that below, because I want you to turn it on.
It plugs into more. Hundreds of tools through MCPs and connectors, up from a few dozen earlier this summer. Custom Agent runs cost about 30 percent less in Notion's internal testing.
Every one of those numbers comes from Notion, by the way. Keep that in your back pocket.
The credits math
Per Notion's announcement, each Business and Enterprise seat gets two buckets a month.
1,500 Standard credits. Notion picks the model. This is Auto's territory.
500 Select credits. You pick the model, or you run Custom Agents. You can start building Custom Agents with these monthly credits, no extra purchase needed.
Admins can buy more. All AI usage lands under one credit system, including outside AI tools working inside your Notion, which will matter when we get to ChatGPT and Claude.
New customers are already on credits. Existing customers switch November 2, and admins get an email about it. If your renewal hasn't come up yet, you still move November 2 with the same credits, and your final allocation gets set at renewal. Notion estimates the vast majority of customers won't run out.
Notion didn't publish a dollar price per credit, or say anything about plans below Business, so I can't tell you what running over costs.
Watch the Select bucket. It's only 500, and it pays for both premium model picks and Custom Agents. Notion itself flagged Fable and Astra as high credit consumption. Picture your ops manager pinning three overnight agents to the fanciest model on the menu because it felt safer, then everyone finding out midmonth that the bucket's empty.
Before November 2
Find out who's using AI now. Send one message to the team asking who uses Notion AI, for what, and how often.
Pick your builders. Decide which one or two people can create Custom Agents. Everyone else can use them.
Write a house rule. One paragraph in your team wiki: Auto is the default for everything. A named premium model is allowed only for client facing writing, complex analysis, or a job where Auto clearly fell down. Fable and Astra need a reason.
Watch for the admin email and forward it to whoever owns the budget.
After November 2
Set caps when AI Command Center arrives. Notion says it's coming soon. Admins will see usage by person, feature and agent, set spending caps and alerts, choose allowed models, and control who can create, pause or archive Custom Agents. When it lands in your workspace, turn on every bit of it that day.
Check usage weekly for the first month, then monthly. Any agent whose output nobody reads gets archived.
Three overnight jobs for a small business
One caveat up front. Notion's post doesn't spell out how triggers and schedules work for these long runs, so check what your workspace offers for kicking an agent off. The setups below work either way. Worst case, you start the job on your way out the door and read the results with your coffee.
Job one: weekly client status pages
I'd build this one first. It can save a real hour a week, and you'll know in thirty seconds whether the output is any good.
Build two databases. Clients (name, owner, status) and Meeting Notes (date, client relation, summary, action items). Link every note to a client.
Create a Status Pages database with one page per client per week. Give it a title, client relation, week, and a Draft or Approved status.
Build the Custom Agent with the instruction block below. Set the model to Auto.
Turn on "Review for me."
Run it Thursday night. Read the drafts Friday morning, fix what's off, flip to Approved, send.
Paste this into the agent's instructions and swap the bracketed bits.
You compile weekly client status pages for [Company Name]. Every run, do this for each client in the Clients database with Status = Active: 1. Find all pages in Meeting Notes linked to that client from the last 7 days. 2. Find any open items in the client's Tasks view. 3. Create one new page in Status Pages titled "[Client Name] week of [Monday's date]" with Status = Draft. Each status page has four sections, in this order: What we did this week. Plain sentences, no jargon, max 6 lines. What's next. The next 3 to 5 actions with an owner name for each. What we need from you. Only real asks with a due date. If there are none, write "Nothing this week." Flags. Anything that sounds like a risk, a complaint, scope creep, or a missed deadline. Quote the exact line from the meeting note and link the source page. Rules: Never invent a task, date, or commitment that isn't in the source notes. If something is unclear, write [CHECK] and say why. Never email, message, or share anything. You only create Draft pages. Never edit or delete existing Meeting Notes. If a client has no notes this week, create the page anyway and write "No meetings logged this week" under What we did. |
Don't skip that last rule. An empty status page usually means someone stopped logging calls, and you want to find that out this week instead of in three months.
FROM THE AI NEWSROOM The AI Workflow Blueprint The exact systems behind everything in this issue. The audit sheets, the routing logic, the templates and the review cadences, built out step by step so you can copy them straight into your own stack. One time, forty seven dollars. Get the Blueprint for $47 |
Job two: clean and tag a messy database
Every CRM or content database I've seen in Notion has the same rot: duplicate contacts, three spellings of the same tag, and a "Misc" status nobody remembers creating.
Duplicate the database first. That's your backup, and no, you can't skip it.
Write the tag list. Put the exact allowed values for each select field on a page the agent can read.
Add three columns. Proposed Tags, Proposed Merge (a link to the suspected duplicate), and Agent Notes.
Tell the agent to propose only. It fills the Proposed columns and leaves the real fields alone.
Spot check 20 rows in the morning. If 18 or more look right, let it apply the changes on the next run. If not, tighten the tag list and run again.
Auto is the right model here. This is volume work, not poetry.
Job three: first pass SOPs from transcripts
You already explain how you do things, on Loom or on calls or while training the new hire. That's raw SOP material sitting there doing nothing.
Make an SOP Inbox page per process. Paste in every transcript that covers it. Two or three is plenty.
Create an SOP template. Purpose, when to use it, steps, tools, common mistakes, who owns it.
Have the agent draft one SOP per inbox page, using only what's in the transcripts, and mark every gap with [CHECK].
Status stays Draft until the person who actually does the job signs off.
The drafts will be rough, and that's fine. Fixing a lumpy draft beats staring at a blank page for an hour.
Turn on "Review for me" for anything that writes
My rule is short. If an agent writes, edits, sends, or deletes anything, "Review for me" is on.
Having a second model check actions before they happen is smart design. It'll catch the obvious stuff, like an agent about to overwrite the wrong field or reply to the wrong thread.
It won't catch everything, though. For anything a client sees, a human still reads it before it leaves the building.
Bring your own AI
If your team already lives in ChatGPT or Claude, you have options.
ChatGPT plugs into your personal agent. ChatGPT Plus, Pro and Go accounts can connect to your personal Notion Agent, though not to Custom Agents, so the overnight jobs above still run on Notion's side.
Notion MCP is the bigger deal. Notion overhauled its MCP so ChatGPT, Claude, Grokbot and Muse can use Notion as a shared knowledge base. Memory sync is live, and skills through MCP are coming soon. What that means for you: your SOPs and client notes become the memory every AI tool on your team reads from.
Mind the credits line here. Outside AI tools working in Notion draw from the same pool, and the chat happening in another window doesn't make it free.
One more split I'd keep. Notion is great for processes and project notes, but relationship history (who you met, who introduced you, when you last talked) lives better in Clay, which was built for it. Not everything belongs in the garage.
Where it falls short
The big numbers are vendor claims. Ninety percent lower cost, ten times further credits, 30 percent cheaper Custom Agent runs, 50 percent faster. All of it is Notion's own testing. Maybe it holds up. I haven't run this at scale, and I'm not going to pretend I have.
Credits can surprise you. No published dollar price, a small Select bucket, and pricey models a click away. Until AI Command Center shows up with caps, your only guardrail is the house rule you wrote and people actually following it.
Fable and Astra burn fast. Notion said so itself. Save them for jobs that actually need them, and ignore the urge to pick them because the names sound expensive.
Notion still can't reach everything. When data has to move between Notion and a tool it doesn't talk to, that's a job for Make.com.
Garbage in, garbage overnight. A long running agent working from sloppy notes will produce a mountain of sloppy output, very efficiently. Clean up the inputs first.
My verdict: if your business already runs on Notion, this is worth a week of setup, and the status page job alone can pay for the effort. Start with one agent on one database, Auto as the default and review switched on. Add the second agent once the first one has earned it.
Two smaller tools worth a look
Both launched or relaunched on Product Hunt the week of October 19.
The Library by Querri. It organizes your company data around the questions you actually ask, and a Librarian agent builds the library and answers from it. It has 125+ connectors plus spreadsheets, and business rules you set in chat get saved as facts. Monitors flag which region, store or rep changed, and answers show the rows, calculation and sources. There's a free option and a 14 day Pro trial. On security, it claims SOC 2 Type II, HIPAA and ISO 27001 certification and says it doesn't train on customer data. My caveat: any tool like this is only as good as the data you hook up, so verify a few answers by hand before you trust the monitors.
Needle. This one's an outbound sales engine built for agents like Claude, ChatGPT or Grok to drive. Describe your leads in one prompt and it finds them on LinkedIn, Instagram and Google Maps, enriches them, and sends from your own connected accounts with sending limits and a human approval step before anything goes out. Two worries. A reviewer noted token costs on premium models add up, and outreach volume and platform rules are entirely on you, so keep the approval step on and the volume sane. There's a 7 day free trial.
FROM THE AI NEWSROOM The AI Business Accelerator For operators who would rather build it with someone than build it alone. The full stack, installed alongside you, with the decisions made in the room instead of guessed at. Ninety seven dollars. Join the Accelerator for $97 |
This week
Ask your team who uses Notion AI and for what. Ten minutes, one message.
Pick your one or two agent builders and write the house rule: Auto by default, premium models only with a reason.
Watch for Notion's admin email about the November 2 switch and forward it to whoever owns the budget.
Build the client status page agent, turn on "Review for me," and run it Thursday night.
Duplicate your messiest database and let an agent propose tags on the copy.
Put a reminder on your calendar to set caps the day AI Command Center shows up.
Jordan
The AI Newsroom | Practical AI for people with a business to run.