Learning AI

English

Tiếng Việt

English

Tiếng Việt

Appearance

Notion AI — An AI layer that lives inside your workspace, from writing assistant to a team of background agents ​

📝

Hands-on — 30 seconds

You're a product lead, and your whole team runs on Notion as both wiki and project tracker. This morning there's a 45-minute sprint meeting: you type /meet right inside a Notion page, hit Start transcribing → the AI transcribes + summarizes + pulls out action items the moment the meeting ends, with no bot joining the room. In the afternoon, a dev asks "what's the refund policy?" — you open Enterprise Search, type one question, and it scans across Slack + Google Drive + GitHub and returns an answer with sources in seconds. A 30-minute synthesis job shrinks to 3 minutes. Real-world payoff: instead of hopping between 5 apps and copy-pasting, the AI lives right where you already store your data — writing, summarizing, searching, auto-filling databases, and running background agents 24/7. A single Business plan at $20/person/month (or $15 if you pay yearly) wraps all of it together.

"Notion AI isn't a chatbot you open and close — it lives inside your pages, databases, and projects.When Notion is your team's data hub, the AI here wins because it can read the real context. When Notion is just a side app, the math on cost usually doesn't work out."

After this chapter you'll be able to

  • Distinguish Notion Agent (command it task by task) from Custom Agents (run automatically in the background) — and know when to use which.
  • Turn on AI Meeting Notes with /meet to transcribe + summarize + pull action items right inside the page.
  • Use Enterprise Search to ask-and-answer across connected Slack/Google Drive/GitHub/Gmail.
  • Create AI Autofill in a database so the AI auto-classifies / scores / extracts entities for every row.
  • Pick the underlying model (GPT-5.2 / Claude Opus 4.5 / Gemini 3) or enable Auto-select per task.
  • Understand pricing & credits correctly (Business $20/month — $15 yearly, credits $10/1,000 for Custom Agents) and know when NOT to pay for Notion AI.

Note on the "shelf life" of this information

This reflects understanding as of mid-2026, compiled mainly from the official notion.com pages (product/ai, pricing, help center, releases, customers) plus a few third-party reviews. AI tools change very fast — model names/versions, prices, dates, and language counts may have changed by the time you read this. Just head straight to notion.com/product/ai and notion.com/pricing to check the latest. Figures that come only from third-party blogs are flagged as [thin source].

01 · What this tool is & when to use it ​

Notion AI is the artificial-intelligence layer built inside the all-in-one Notion workspace (docs, wiki, database, project). Unlike a standalone chatbot, Notion AI lives right inside your pages/databases and can use context from your own workspace, your connected apps (Slack, Google Drive, GitHub, Gmail…), and the web. Vendor: Notion Labs, Inc. Official URL: https://www.notion.com/product/ai.

Per its positioning on the AI product page (as of mid-2026), Notion AI has evolved from a "writing assistant" (2023–2024) into a programmable agent platform — Notion calls this "Meet your AI team": it includes Notion Agent, Custom Agents running 24/7 in the background, AI Meeting Notes, Enterprise Search, and AI Autofill.

Three things that are easy to confuse (read carefully or you'll mix them up)

  • Notion = the workspace platform (docs/wiki/database/project) — the "place" where you store everything.
  • Notion AI = the AI layer inside that workspace. This chapter is about this.
  • Notion Agent vs Custom Agents = two different AI modes: with Notion Agent you command it task by task; Custom Agents are named agents that run automatically in the background on a schedule/event, without you prompting them. (Details in Sections 02–03.)

What Notion AI does well (confirmed from notion.com/product/ai):

Task groupWhat it doesHow to use
Notion AgentMulti-step tasks using context from Notion + connected apps + the web; create pages, write database formulas, build views via natural languageOpen the AI bar, command it task by task
Custom AgentsNamed agents that run automatically on a schedule or event trigger, operate 24/7, and write results back to a databaseConfigure once, run in the background
AI Meeting NotesRecord + transcribe + summarize in real time right on the page, "no bot needed"; supports ~16 languagesType /meet (desktop app only)
Enterprise SearchSearch across connected apps: "Search across Slack, Google Drive, GitHub & more — in seconds" (also called "Ask Notion")Ask one question, get the answer + sources
AI AutofillDefine one property (category, summary, sentiment, entity…) and the AI fills every row, including rows added laterInside a database
Writing / AI blocksImprove style, fix spelling/grammar, change tone, continue writing, translate, summarize, find action itemsRight inside the page
Research ModeGenerate detailed reports & summaries on a topic, with reasoning capability (a Business AI feature, listed on notion.com/pricing)In the AI bar

Pick the AI model right inside the document (from Notion 3.2)

From Notion 3.2 (Jan 20, 2026), you can choose the underlying model right inside the document: GPT-5.2 (OpenAI), Claude Opus 4.5 (Anthropic), Gemini 3 (Google) — or enable Auto-select to let Notion route by task type. Switching models mid-stream still preserves context. (Model names/versions are per sources as of mid-2026 and may have changed.)

Notable change: the $10 Notion AI add-on has been killed

Previously Notion AI was sold as a $10/month add-on bolted onto a plan. As of May 2025, this add-on was discontinued for new Free/Plus customers — all AI is now consolidated into the Business $20 plan. Existing customers keep access on a grandfathered basis. If you read an old article mentioning "buy the $10 Notion AI add-on," that information is out of date.

Use Notion AI when: Notion is your (or your team's) central workspace; you need ask-and-answer across all your docs + connected apps; you need meeting notes right inside the page; you need to auto-classify/synthesize databases in bulk; you want a background agent doing repetitive work. Think twice when: you don't really "live" in Notion, or you've already paid for another general-purpose chatbot (see Section 09).

Versus other tools — "when to pick which" ​

The key thing to remember: the pro tier of most tools is ~$20/person/month, so price is not the deciding factor. Choose based on the tool you already use and the kind of work you do. The table below is brief and objective, as of mid-2026:

CriterionNotion AIChatGPTClaudeGemini (in Docs)Coda AI
NatureAI inside the workspace; strong when Notion is the data hubGeneral-purpose chatbot, multimodal, web search, codeStrong at writing & reasoning; many praise its text qualityDeeply integrated with Google Workspace (Gmail/Docs/Sheets/Meet)AI woven into formulas & automation, strong at table/row-level
Main weaknessDoesn't freely browse the web, doesn't analyze images, doesn't run tasks outside Notion (per reviews) — whereas ChatGPT/Gemini/Claude all already have web browsing; agents need very precise prompts and must be reviewedNot tied to your internal workspaceSimilar — it's not your "work operating system"Locked into the Google ecosystemSmaller community/templates than Notion; pricing has changed several times
AutomationTriggers at the database/page level (page added / property changed)n/a (outside the workspace; has Projects/Canvas)n/a (has Projects)n/a (has Canvas/Gems)Triggers at the cell/row level — more granular for CRM/inventory
Pro price$15 yearly / $20 monthly (Business)~$20/month (Plus)~$20/month (Pro)~$20/month (AI Pro)Only charges for "Doc Maker"; editors/viewers free
When it winsIt's the central workspace, you run a team, you need workspace-wide Q&A + meeting notesYou need a general-purpose assistant outside NotionYou prioritize writing/reasoning qualityYou already live in Google WorkspaceHeavy data work, a large org with many viewers, a tight budget

Quick summary

If Notion is already your (or your team's) data hub → Notion AI is the natural and strongest choice (workspace-wide Q&A + meeting notes + autofill). If you need a general-purpose assistant outside Notion → ChatGPT; prioritize writing/reasoning → Claude; already live in Google → Gemini; heavy data work, many viewers, tight budget → consider Coda. One Notion advantage few people notice: the deep yearly discount (Business $20 → $15/seat), whereas ChatGPT Plus / Claude Pro / Gemini AI Pro barely discount the annual plan. Per one third-party blog estimate, a 100-person org (20 creators + 80 viewers) costs ~$1,650/month on Notion Business versus ~$600/month on Coda Team [thin source — flagged as "an estimate"].

When NOT to use Notion AI (read before you pull out your wallet)

  • You don't live in Notion. If Notion isn't your central workspace, or you already pay for ChatGPT/Claude/Gemini → the math on cost usually doesn't work out. A community rule of thumb from Reddit: "Do you spend more than 2 hours/week writing or reading in Notion?" — if not, just use ChatGPT and copy-paste across.
  • You need to run tasks outside Notion (freely browse the web, analyze images, run general-purpose code) → use a general-purpose chatbot.
  • Client/team meetings that need speaker identification, or recording on mobile → a dedicated meeting tool (tl;dv/Otter/Fireflies) is better. (Notion does label speakers, but only in English and for 1-on-1/online meetings — see Section 04.)
  • Automation that needs to react per cell → Coda fits better (Notion only triggers at the database/page level).
  • Handling sensitive data (PII/PHI, customer records) while NOT on the Enterprise plan → only Enterprise has zero data retention with the LLM; non-Enterprise plans still let the LLM provider keep data ≤ 30 days (details in Section 04). For genuinely sensitive data, either move to Enterprise or don't feed it to the AI at all.

02 · Sign-up & access ​

Available globally? — Yes. ​

  • No geo-blocking observed — there's no restriction information. You can sign up and use it normally anywhere, with no VPN required.
  • Notion (app + web) has an official localized UI in many languages — for example, Vietnamese shipped on Jul 22, 2025 (alongside Indonesian/Thai). Change the language in Settings.
  • Notion AI Translate supports many languages (a group of ~14 translation languages; AI Meeting Notes covers more, ~16 — the language count differs per feature and may have changed). The writing features (improve writing, fix grammar, change tone, simpler language) work across these languages too.

Important note: AI Meeting Notes only runs on the desktop app

The recording feature of AI Meeting Notes (/meet) only runs on the desktop app (macOS/Windows), per community reviews (Reddit/tl;dv). On web/mobile you can't start a recording session. If you often meet on your phone → this is a limitation to know up front.

Still in beta — don't depend on it 100%

As of mid-2026, per the official pages, both AI Meeting Notes (the help center says "AI Meeting Notes (beta)") and Enterprise Search (pricing says "Enterprise Search beta") are still in beta. That means features may change and can sometimes be unstable — usable for real work, but keep a fallback for anything critical.

Sign up in 30 seconds ​

text
1. Open https://www.notion.com (or download the desktop app for macOS/Windows, mobile for iOS/Android).
2. Sign up with email, or sign in quickly via Google / Apple.
3. Go to your workspace -> open Settings to change the language if you want.
4. Type "/" in a page to call the AI; type "/meet" (on desktop) to turn on Meeting Notes.

Plans & pricing (updated mid-2026 — monthly price and yearly price) ​

PlanPrice/person/month (yearly)Price/person/month (monthly)What AI is included
Free$0$0Limited trial of basic AI features only (chat, generate, autofill, translate)
Plus$10$12Still only a limited trial of AI features
Business (Recommended)$15$20Full set: Notion Agent, Ask Notion / AI search, AI Meeting Notes, Enterprise Search
EnterpriseContact sales—Everything in Business + Zero data retention with the LLM, SCIM, audit log, unlimited history, dedicated CSM

How to read the price (a common mistake): the number people often quote, $20/person/month, is the monthly price of Business; if you commit to yearly billing it drops to just $15/person/month (~$180/person/year). Plus is similar: $10 (yearly) / $12 (monthly). The pricing page says "Save up to 20% with yearly." Source: notion.com/pricing.

🧮 Credits for Custom Agents — the part that's easy to miss

Custom Agents (launched Feb 24, 2026, see Section 03) are free to try until the next billing cycle (on/after May 4, 2026), after which they cost:

  • $10 / 1,000 Notion credits / month, non-rolling (resets monthly), shared across the whole workspace.
  • How far do 1,000 credits go? An estimated ~30–60 runs (≈ $0.17–0.33/run) depending on agent complexity — an estimate as of mid-2026, not an official commitment. Notion 3.4 part 2 (Apr 14, 2026) cut Custom Agent costs ~35–50%, so real-world usage may be lower.
  • This is a fee on top of the Business/Enterprise plan — it's not bundled in.
  • The pricing page also mentions "Workers (Beta)… starts using credits on August 11" — still little material on this [thin source].

A note on the actual "free tier"

Don't get it wrong: the Free and Plus plans only get a limited trial of AI features. To use AI seriously (full Notion Agent, AI Meeting Notes, Enterprise Search) you must move to Business ($20/person/month, or $15 if you pay yearly). If your goal is just "to try it out," Free is enough; but don't expect to use AI every day on Free/Plus.

Pricing & access ​

  • Web: debit/credit card (Visa/Mastercard are accepted), Apple Pay, Stripe Link, SEPA (Europe only).
  • Mobile: iOS via Apple Pay; Android via Google Play (manage the plan through the App Store / Play Store).
  • No PayPal support. Billed prepaid monthly/yearly, processed via Stripe.
  • Pricing is generally charged in USD unless a local-currency option is shown for your region; check the in-app price at checkout.

Note for Vietnam / SEA readers

Payment goes through Stripe via Visa/Mastercard or Apple Pay (no PayPal). The help center doesn't list a VND price, so the charge is most likely in USD (this is an inference — the help center doesn't say so explicitly). Domestic-only cards from some countries can be blocked for recurring international charges — if a payment is declined, use a card that supports international payments (e.g. a virtual card).

Be careful before upgrading the whole team to Business

Because real AI is only in Business ($20/person/month billed monthly, $15 yearly), the cost multiplies by headcount fast: a 10-person team = $200/month (monthly price) or ~$150/month (yearly) — not counting Custom Agents credits. Pilot first with 1–2 Business seats to confirm the AI actually solves your problem (meeting notes, Enterprise Search, autofill) before rolling it out to the whole team.

03 · The hands-on workflow — step by step (with real commands & prompts) ​

Here are the 4 most-used workflows. Each step includes a way to verify that you did it right.

A · AI Meeting Notes — record, transcribe, summarize right inside the page ​

Per the official help center:

text
1. On the DESKTOP app, type /meet in any page.
2. Before the meeting: write your agenda / notes into the "Notes" section -> the AI
   will use them as context when summarizing.
3. Click "Start transcribing" (confirm everyone consents to recording).
4. After the meeting click "Stop" -> the AI auto-generates a summary + action items.

→ Verify: after you click Stop, the page has a summary + a list of action items that accurately match what was said (not made-up content).

A sample prompt to standardize the minutes (from the community/PromptRocket — [thin source]):

text
Meeting type: [STANDUP / CLIENT CALL / STRATEGY]
Attendees: [...]
Date: [...] — Duration: [...]
Raw notes: [PASTE THE TRANSCRIPT]

Produce a summary with:
1. Meeting Overview (1-2 sentences)
2. Key Decisions Made (note who decided)
3. Action Items (owner + deadline)

B · Ask/do things in the page with Notion Agent ​

Open the AI bar → type a natural-language command. These are the real prompts Notion suggests to product teams (from the official guide):

text
Simplify the language        # condense a long PRD
Continue writing             # continue the paragraph in progress
Make longer                  # expand the idea to be longer
Fix spelling & grammar       # fix spelling / grammar
Change tone                  # change tone: friendly / casual / professional
Summarize                    # summarize the whole page
Find action items            # pull to-dos from meeting notes / a doc
Brainstorm ideas             # feature ideas, naming a feature

→ Verify: the result sticks to the content of the page you have open, without "injecting" unrelated information.

✍️ The golden rule Notion emphasizes

Notion recommends writing "detailed and specific prompts" — the more detailed and specific the prompt, the better the result. "Rewrite this paragraph" loses badly to "Rewrite this paragraph 30% shorter, in a professional tone, keeping the 3 figures in the table intact."

C · AI Autofill in a database — let the AI fill in bulk ​

text
1. In a database, create an AI-type property.
2. Describe the task, for example:
   "Classify the sentiment of this feedback as Positive / Neutral / Negative."
3. Turn on autofill -> the AI fills EVERY row, including rows you add later.

→ Verify: check the first few rows to see whether the AI classifies sensibly; add a new row and confirm it auto-fills.

Upgrade: Autofill runs on Custom Agents (Notion 3.4 part 2)

Notion 3.4 part 2 (Apr 14, 2026) upgraded AI Autofill to run on Custom Agents as a background process, with a third-party source describing "sub-3-second autofill" — this speed detail is from the Fazm blog [partly thin source].

D · Custom Agent — a named agent that runs in the background automatically (Notion 3.3) ​

Launched Feb 24, 2026 in Notion 3.3. This is the biggest difference from Notion Agent: you configure it once, and the agent runs in the background without prompting.

text
1. Give the agent a NAME (e.g. "Sales Feedback Categorizer").
2. Assign its skill / task in natural language.
3. Pick a trigger:
   - on a recurring SCHEDULE (e.g. every morning at 8am), or
   - on an EVENT (e.g. when a new row appears / a property changes).
4. Specify which database the OUTPUT writes into.

→ Verify: wait for the trigger to fire (or create a test event) and check the destination database to see whether the agent wrote the correct result.

Notion Agent vs Custom Agents — the rule for choosing

  • Notion Agent = you command it task by task and watch the result. Good for one-off jobs (write, summarize, build a view).
  • Custom Agents = an agent that runs 24/7 in the background on a schedule/event and writes results into a database. Good for repetitive work (classify feedback every day, weekly synthesis…). Ramp reports that one agent was created in 3 minutes between meetings (see Section 06).

Connectors confirmed on the official page

Notion AI connects to: Slack, Google Drive, GitHub, Asana, Gmail. These are the data sources for Enterprise Search and where Notion Agent/Custom Agents pull context from outside Notion.

04 · Pro tips & common mistakes ​

🟢 Tips that pay off ​

7 tips to use Notion AI like a pro

  1. Detailed & specific prompts — Notion says outright this is the #1 factor. Spell out the length, tone, constraints, and the data that must be preserved.
  2. Put knowledge into the prompt and force the AI to "use only the provided documents" → fewer fabrications, especially on niche technical topics.
  3. Use the right mode: one-off work → Notion Agent; daily repetitive work → Custom Agents (no manual prompting).
  4. Lean on Auto-select model, but when you need high writing quality → pick the right model yourself; switching mid-stream still preserves context.
  5. Write the agenda into the Notes section before /meet → the summary sticks to the key points much better.
  6. Treat the output as a "draft," always review — especially with agents: results usually need human approval before use.
  7. Use the localized UI if your language is supported (e.g. Vietnamese since Jul 22, 2025) and use Translate/writing in your language — but double-check domain-specific terminology.

🔴 Mistakes & pitfalls (read carefully — this part saves you) ​

🚨 Hallucination (making things up) — still the #1 pitfall

Like any generative AI, Notion AI can state wrong information very confidently, especially:

  • Very recent news or niche technical topics.
  • When the prompt is vague, the AI "infers" to fill the gap.

→ How to prevent it: treat the output as a draft, always verify; put knowledge into the prompt and require "use only the provided documents; if it's not there, say it's not there."

Notion AI's specific limitations to remember

  • AI Meeting Notes: speaker labels are English-only and only reliable for 1-on-1 / online meetings (it captures both your mic and system audio; if you link a calendar event it can attach the other person's name). For group meetings or multiple people sharing one mic it can barely tell who's speaking. It also doesn't join the meeting room itself and only runs on desktop. For client/group meetings that need to know who said what → use a dedicated tool (tl;dv/Otter/Fireflies). Source: notion.com/help/ai-meeting-notes.
  • Writing style can get repetitive/bland if you don't specify a concrete style.
  • Lacks deep conversational memory compared to a dedicated chatbot; it stumbles on vague prompts.
  • Notion/Custom Agents "need very precise prompts," and results usually require human review; as of early 2026, some reviews call the reliability "not yet stable enough" to depend on completely (Reddit/eesel — [thin source / opinion]).
  • Automation is database/page-level only, with no per-cell triggers (it loses to Coda here).
  • No free web browsing, no image analysis, no running tasks outside Notion (per reviews) — this is not an all-purpose chatbot.

Privacy & data — read carefully if you use it for work

Per Notion's official help center, as of mid-2026:

(a) Does it train on your data? — NO.

  • Quoting directly: "Your use of Notion AI does not grant Notion any right or license to your Customer Data to train machine learning models."
  • Notion has contracts that prohibit AI subprocessors from using customer data for training.

(b) Data isolation: each customer's account is kept separate in the production environment; data isn't mixed during AI processing.

(c) Storage at the LLM provider:

  • Enterprise → zero data retention with the LLM provider.
  • Non-Enterprise → the LLM provider keeps data ≤ 30 days before deleting it.

(d) Encryption & certification: TLS 1.2+ in transit; the vector DB storing embeddings is SOC 2 Type 2 (audited by an external auditor).

(e) HIPAA: Notion AI supports HIPAA compliance thanks to the LLM provider's zero-retention API (for processing PHI).

(f) Even so — be cautious with personal data: putting customers' personal data (names, phone numbers, records) into any cloud service can, in most jurisdictions, touch data-protection law (e.g. the EU's GDPR, and equivalent privacy regulations elsewhere) — carefully weigh the scope of data before enabling a connector that scans the whole workspace.

Reference pages: notion.com/help/notion-ai-security-practices · /help/ai-safety · /help/enterprise-search-security-and-privacy-practices.

FAQ & common errors (click to open)

Typed /meet but don't see it / can't record? AI Meeting Notes only runs on the desktop app (macOS/Windows). On web or mobile you can't start a transcribe session. Install the desktop app and try again.

I'm on Free/Plus — why am I missing Notion Agent / Enterprise Search? Free and Plus only get a limited trial of AI features. The full AI suite (Notion Agent, AI Meeting Notes, Enterprise Search) is only on Business ($20/person/month monthly, $15 yearly) and up.

How much cheaper is yearly billing? Business: $15/person/month yearly versus $20 monthly (~25% cheaper). Plus: $10 versus $12. The pricing page says "Save up to 20% with yearly." If you're sure you'll use it long-term, yearly saves noticeably once multiplied by seats.

How many Custom Agent runs do 1,000 credits get me? Per third-party estimates as of mid-2026, ~30–60 runs / 1,000 credits (i.e. ~$0.17–0.33/run) — an estimate; each agent consumes differently by complexity. Note: Notion 3.4 part 2 (Apr 14, 2026) made Custom Agents ~35–50% cheaper, so real-world usage may be lower. Source: notion.com/help/custom-agent-pricing.

My Custom Agent suddenly says it's out of credits? Custom Agents cost $10 / 1,000 credits/month, non-rolling and resetting monthly, shared across the whole workspace. If many agents run in the background at once, credits drain fast — monitor your usage and add more if needed.

I previously bought the $10 Notion AI add-on — can I still use it? The $10 add-on was discontinued (May 2025) for new Free/Plus customers. Existing customers keep access on a grandfathered basis. New customers who want full AI must move to Business.

Notion AI replies in English even though I wrote in another language? State it explicitly in the prompt, e.g. "Reply in [your language]." Also switch the UI to your language in Settings (Vietnamese, for example, has been officially supported since Jul 22, 2025). Translate supports many languages.

The agent returns wrong / weird results? Agents "need very precise prompts," and results should be human-reviewed. Rewrite the task in more detail, narrow the data scope, and don't fully depend on un-checked output.

How do I pay, and is there PayPal? No PayPal. Use a Visa/Mastercard card, Apple Pay, or Stripe Link (web); on mobile via the App Store/Play Store. Processed via Stripe; the help center doesn't list local currency, so it's most likely charged in USD.

05 · Exercises / mini-projects ​

Actually do 2–3 of the exercises below to turn "I understand it" into "I can do it." Each has clear completion criteria.

Exercise 1 — Meeting Notes + pulling action items (basic) ​

Goal: experience the full /meet lifecycle from recording to action items.

  1. On the desktop app, create a new page and type /meet.
  2. Go to the Notes section and pre-write 3 agenda bullets (e.g. "lock the deadline, assign owners, risks").
  3. Click Start transcribing, read aloud a ~2-minute mock meeting (play 2 people yourself), then click Stop.
  4. Cross-check: does the summary cover all 3 agenda items? Do the action items assign the right person + task?

Done when: you've confirmed the summary sticks to what was said and pulls out at least 1–2 sensible action items (or you catch where it made something up). Write one sentence: did writing the agenda beforehand make the summary tighter?

Exercise 2 — AI Autofill for sentiment classification (important) ​

Goal: practice bulk AI fill + the reflex to verify.

  1. Create a simple database with a "Feedback" column and paste in ~8 mock feedback lines (positive/negative/neutral mixed).
  2. Add an AI-type property named "Sentiment" with the description:
text
Classify the sentiment of the "Feedback" column content as exactly one of three:
Positive / Neutral / Negative. Base it only on the text in that row.
  1. Turn on autofill, then add a new feedback row and confirm it auto-fills.
  2. Manually cross-check all 8 rows: are any of the AI's classifications wrong?

Done when: most rows are correct, the new row auto-fills, and you can point out which rows (if any) the AI got wrong. This is the "don't blindly trust the output" reflex you must keep forever.

Exercise 3 — Custom Agent running in the background (advanced) ​

Requires the Business plan + credits

Custom Agents need the Business plan and consume credits ($10/1,000). If you don't have it yet, do a "paper draft": write out the agent config (name, task, trigger, output) as if you were about to create it — you'll still train your agent-design thinking.

Goal: build a background agent for one of your repetitive tasks.

  1. Pick one task you do repeatedly in Notion (e.g. every morning, synthesize the "In progress" tasks into a status paragraph).
  2. Create a Custom Agent: name it, describe the task, choose a schedule trigger (e.g. 8am), and specify the output database/page.
text
Name: Daily Standup Summarizer
Task: Every morning, read the tasks with status "In progress" in the [Project]
database, write a 4-6 bullet summary (what, who's on it, what's blocking),
and write it into the "Standup [today's date]" page.
Trigger: scheduled, 08:00 every workday.
  1. Wait for the trigger (or trigger it manually) and check the output.

Done when: the agent writes the correct summary into your chosen location without you commanding it each time — or (if you did the "paper draft") you've written an agent config clear enough that someone else could build it.

06 · Case studies & real use cases (from the community & official sources) ​

This section gathers real examples from Notion's official case studies, partner announcements, and third-party reviews as of mid-2026. The point: to show you how Notion AI runs in the real world — both when it shines and when its limits show.

Read carefully about source reliability

  • Figures in Notion's own case study (Ramp) are vendor claims — read them with measure.
  • Figures from a partner (Decagon) are published by the partner — flagged "per Decagon."
  • The [thin source] parts are third-party reviews/blogs or community discussion, not independently verified.
  • A company's valuation/revenue figures (e.g. Ramp) are business context and should not be directly attributed to Notion AI's effectiveness.

CS1 — Ramp: running 300+ Notion Agents per day (official case study) ​

  • Background: Ramp (a fintech company) is an official Notion case study at notion.com/customers/ramp.
  • What they do: Ramp runs 300+ Notion Agents/day, each named for a specific job: Product Q&A Oracle (connected to Slack, answers product questions), Sales Feedback Categorizer (maps feedback to the roadmap), Referral Bonus Roy, Enablement Eddie, Customer Advocacy Miner, The Underwriter…
  • Results / figures (per Notion): ~70% reduction in productivity-tool costs; the team "moves 3× faster." Ramp says one agent can be created in 3 minutes between meetings.
  • Lesson: the real power of Custom Agents is cloning repetitive work — instead of one "do-everything assistant," Ramp built hundreds of agents each dedicated to one task, clearly named, running in the background. (Ramp's valuation / revenue figures are company context and should not be directly attributed to Notion AI's effectiveness.)
  • Source (official Notion — vendor claim): https://www.notion.com/customers/ramp

🎧 CS2 — Notion + Decagon: up to ~34% reduction in customer-support ticket handling time ​

  • Background: Notion deployed Decagon's AI customer-experience solution for its customer-support team.
  • Results / figures (per Decagon): up to ~34% ("up to 34%") reduction in ticket handling time; an ask-for-human rate of only ~3.4%; handling roughly 1 million requests/year. Smart routing increased first-touch resolution and lightened the load on human agents.
  • Lesson: even a company "born in Notion" bolts on a dedicated tool for customer support — illustrating that Notion AI is strong at knowledge/workspace work, while large-scale ticketing still needs a specialized layer.
  • Source (partner Decagon — read with measure): https://decagon.ai/case-studies/notion

CS3 — Product team: upgrading PRDs/RFCs, synthesizing user research (official guide) ​

  • Background: Notion published an official guide on using Notion AI for product teams.
  • What they do: use the AI to upgrade PRDs/RFCs (simplify language, make longer, fix grammar), synthesize user research, brainstorm features, scope projects — using exactly the prompts in Section 03B.
  • Lesson: for product people, the most immediate value isn't the "background agent" but the AI blocks right inside the document — editing, summarizing, pulling action items.
  • Source (official Notion): notion.com/help/guides (the guide series for product teams).

CS4 — Surfacing insights from a database with AI Autofill (official guide) ​

  • Background: Notion's official guide "5 AI prompts to surface insights from your databases."
  • What they do: use AI Autofill + prompts to classify, score, and extract entities for many records at once (e.g. classify sentiment, tag topics, extract a company name from a description).
  • Lesson: this is the "sweet spot" of Notion AI for data-operations people — you describe the criteria in plain language, the AI fills every row, and every row added later.
  • Source (official Notion): notion.com/help/guides.

CS5 — Solo/creator: meeting notes handy for individuals, but weak for group meetings ​

  • Background: reviews from tl;dv, a Medium post "60 days test," and Saner.ai ([thin source — blog/review]).
  • What they do / takeaways: AI Meeting Notes is useful for individual users working in Notion (taking your own meeting notes solo). But it's weak for group/client meetings in languages other than English because: speaker labels only work in English and for 1-on-1/online meetings (with many people sharing one mic it's hard to separate who's speaking), it doesn't join the room itself, and it only runs on desktop.
  • Lesson: choose the tool by meeting context. Solo + you already live in Notion → handy; multi-person meetings that need to know who said what → a dedicated meeting tool wins.
  • Source (third-party reviews): tl;dv, Medium "60 days test," Saner.ai.

CS6 — The "~35% productivity boost" figure deserves a big question mark

A few blogs claim "Notion's own data" shows AI boosts productivity "by up to ~35%," and the "~34%" figure (Decagon) also gets re-quoted a lot. The problem: no original official page can be found for the 35% number. → Use these figures with a strong hedge or drop them entirely; don't put them in a report as verified fact. This is exactly the kind of "pretty number that spreads" you should be skeptical of yourself.

07 · Summary & official sources ​

6 things to take away

  1. Notion AI = an AI layer living inside the workspace — strongest when Notion is your (or your team's) data hub.
  2. Two agent modes: Notion Agent (command it task by task) vs Custom Agents (run 24/7 in the background on a schedule/event, launched Feb 24, 2026).
  3. Real AI is only on Business ($20/person/month monthly, $15 if yearly); Free/Plus get only a limited trial. Custom Agents cost extra credits at $10/1,000.
  4. /meet (desktop only) for meeting notes; Enterprise Search scans Slack/Drive/GitHub/Gmail; AI Autofill fills databases in bulk.
  5. Pick the model GPT-5.2 / Claude Opus 4.5 / Gemini 3 or Auto-select; switching mid-stream still preserves context.
  6. Usable globally (localized UI, e.g. Vietnamese since Jul 22, 2025, with Translate support for many languages); pay by card/Apple Pay/Stripe, no PayPal. It doesn't train on customer data; non-Enterprise keeps it ≤30 days, Enterprise is zero-retention.

These are the official pages to check the latest information yourself — always trust these links over third-party summaries:

Reliability notes (research as of mid-2026)
  • Certain (official): plan pricing (Business $20 monthly / $15 yearly; Plus $12 / $10); add-on discontinued May 2025; credits $10/1,000; Custom Agents launch date (Feb 24, 2026) & Notion 3.2 (Jan 20, 2026); Research Mode (listed on the pricing page); connectors (Slack/Drive/GitHub/Asana/Gmail); security/retention policy; localized UI (Vietnamese Jul 22, 2025); payment (card/Apple Pay/SEPA, no PayPal, via Stripe); the Ramp case (>300 agents, ~70% cost reduction); AI Meeting Notes & Enterprise Search still in beta (as of mid-2026).
  • Hedged "per sources as of mid-2026 / ~": language counts (~16, varies by feature); model names/versions (GPT-5.2 / Claude Opus 4.5 / Gemini 3 — Notion may have updated these underlying models to newer versions); the "sub-3s autofill" speed; the ~30–60 runs / 1,000 credits estimate; "Workers Beta starts using credits Aug 11."
  • [Thin source — re-check before quoting]: the ~34% productivity figure (Decagon, an "up to") and ~35% (blogs citing "Notion data," no original page found); the Coda 100-person cost comparison; agent-reliability assessments (Reddit/blog).

Figures (prices, models, features, dates) may have changed — always re-check at notion.com/product/ai and notion.com/pricing.