Notion AI review (2026): agents, connectors, and what's actually working

Notion AI is no longer the inline writing helper it shipped as in early 2023. By May 2026 it is closer to a full agent platform that lives inside the Notion workspace, with autonomous Personal Agents, scheduled Custom Agents, AI Connectors that read across Slack, GitHub, Google Drive, Salesforce, Box, and Notion Mail, plus Research Mode, AI Meeting Notes, and a credit-based billing model that started this week.

This review covers what shipped, what each feature actually does, what real users say after living with it, what the full plan structure looks like, and where the honest limits are.

What is Notion AI in 2026?

Notion AI is the umbrella name for every AI feature inside Notion: agents, search, writing, database tools, image generation, and the standalone Notion Mail and AI Meeting Notes products. It is bundled into the Business and Enterprise plans, with Free and Plus users getting a capped trial.

The positioning is agent-first. Notion's own product page leads with this customer line:

"If it's something that's repetitive, we have a Notion Agent for it." Ben Levick, Head of AI & Operations, Ramp (Notion AI)

Ramp has built more than 300 custom agents on Notion, which Notion claims cut their productivity-tool spend by 70%. That is the use case Notion is now selling: not "draft this paragraph", but "have an agent triage support tickets every Monday morning while you sleep."

It is also a real revenue line. Notion hit $600M ARR in 2025 with 50% of that from AI products, which is what justifies the $11B valuation the December 2025 secondary sale settled at. Whatever you think of Notion AI as a product, it is paying the bills.

Notion 3.0 and the agent era

The pivot moment was September 18, 2025. Notion 3.0 launched Agents, and the framing shifted from "AI inside docs" to "agents that use Notion." The launch blog describes the original Personal Agent as software that can run "across hundreds of pages simultaneously" for up to 20 minutes of autonomous work. That capacity has only grown since.

Then in Notion 3.3 (February 2026) Notion split the agent product in two: Personal Agents (one per user, your assistant) and Custom Agents (team-level, schedule-driven, runs while everyone is asleep). It is the second category that has driven most of the 2026 conversation.

Personal Agent

Your Personal Agent pulls context from your workspace plus any Connectors you have authorised: Slack, Google Drive, GitHub, Figma, Linear, and so on. You can rename it, give it custom instructions on a dedicated Notion page (tone, formatting rules, working hours), and pick a model per request from GPT-5.2, Claude Opus 4.5, Claude Sonnet 4, or Gemini 3, with the model context preserved when you swap. Per-prompt model picking arrived with Notion 3.2 in January.

What it actually does in practice: search across connected tools and the open web, draft documents, build tables and databases on the fly, schedule meetings via the Calendar connector, and update hundreds of pages in one autonomous run. The 20-minute autonomous limit is the one to remember; if your task can be expressed in 20 minutes of Notion work, the Personal Agent can probably handle it.

Custom Agents

Custom Agents are the ones a workspace admin sets up once for the whole team. The official guide on building one walks through the pattern: click "+ New agent" in the sidebar, describe what it should do in natural language, pick a trigger (a Slack mention, a database row change, a daily schedule), grant it permissions to specific pages and connectors, and test it manually before turning it loose.

The use cases Notion documents are mostly about internal operations:

• A Q&A agent that auto-answers questions in a #product-ask Slack channel using existing docs.
• A task-routing agent that captures incoming requests and tags them by severity or component.
• A reporting agent that builds weekly summaries from project databases and Slack threads.

The 2026 cost story matters here. Notion 3.4 part 2 cut Custom Agent runs by 35-50% by adding cheaper models (GPT-5.4 Mini, Haiku 4.5, MiniMax M2.5), with up to 10x fewer credits per run. That release also shipped Skills (save a workflow as a reusable command), AI Autofill in databases (continuous enrichment of database rows), and inline text editing. If you tried Custom Agents on the original pricing and stopped, the math has changed.

The credit billing started May 4, 2026: $10 per 1,000 Notion credits, available as an add-on for Business and Enterprise. Workspace admins buy a credit pack that matches the team's expected usage. It is real spend; you should calibrate against actual agent runs in your trial period before committing.

AI Connectors

Connectors are how Notion AI escapes the Notion workspace. As of May 2026 the live integrations are Slack, Google Drive, GitHub, Figma, Linear, HubSpot, Notion Mail, and Calendar. Salesforce and Box landed with Notion 3.4 part 2, and Outlook Mail is queued. There is also an n8n MCP integration for routing agent work into other systems, and an MCP audit-log for Enterprise customers worried about external AI tool sprawl.

The permission model is the part to understand: agents only see what you already have access to in the source app. If your Salesforce role does not let you see closed-won opportunities, the agent will not surface them either. On May 1, 2026 Notion added private Slack channel access, gated behind Settings → Notion AI → AI Connectors → Enable access to private content. Agents only see private channels they are explicitly invited to.

This is also where Notion AI shows its scope. Connectors read across these tools, but they read for internal team consumption. There is no Zendesk, Freshdesk, Front, or Gorgias connector on Notion's side, and there is not really a notion (sorry) of "answer a customer ticket". If your goal is a customer-facing AI agent that uses Notion as one knowledge source among many, that is a different product category, and tools like eesel AI's Notion integration are built for that helpdesk-side workflow.

Q&A, Enterprise Search, and Research Mode

The three search-flavoured features all live under the same chat surface but answer different kinds of questions.

Standard Q&A is the everyday one. Open Notion AI from the bottom-right icon, the search sidebar, or Shift+Cmd+J, and ask a question in plain language. It searches your workspace pages, databases, and connected tools, and returns "an easy-to-digest AI overview along with a breakdown of all the sources" with citations back to specific docs (per the help docs). It is fast, and on a tidy workspace it is the feature most teams use first.

Enterprise Search is the variant that adds connected-tool sources (Slack, Google Drive, GitHub, JIRA, Microsoft 365) into the same answer. The enterprise search guide describes results "organized by source type" and lets you pick between Notion AI, ChatGPT, or Claude as the answering model. It is the feature that the Q&A complaint thread is really about: answers are only as consistent as your underlying data hygiene.

Research Mode is the deep one. From the help docs: toggle Research in the search bar, ask a question, and the system will run "deep, multi-step searches across your workspace, connected tools, and the web, following every lead until it has the complete picture." It runs up to five minutes while you keep working, then synthesises a hyperlinked report. Notion suggests it for things like drafting a PRD grounded in customer feedback, prepping for a sales call with full account context, or building a market entry plan from sales data. It is the closest Notion gets to the Deep Research pattern that started showing up in dedicated AI products in 2024-2025.

How to actually get good answers

Notion's own prompting guide is unusually specific about this. They recommend adding context across four dimensions: topic ("Q1 2025 APAC sales performance" beats "sales numbers"), author ("Sandra's brand guidelines"), time ("from last quarter"), and document type ("the database, not the wiki"). They also recommend @ mentions for people and pages, since Notion AI uses those as direct context handles.

This is real advice. On a workspace with a thousand pages and three connectors firing, generic prompts will return generic answers, which is most of the gap between the praise reviews and the criticism reviews.

AI Meeting Notes

AI Meeting Notes is the feature most reviewers praise unprompted. From the help docs, it can transcribe a live meeting, generate a summary, surface action items "as a to-do list," and reorganise raw notes into tables. Notion 3.4 part 2 added custom summary instructions (define your team's preferred format) and exposed transcripts and summaries via API.

The Notion 3.2 release in January shipped Mobile AI Notes, with one-tap transcription that keeps recording when you lock your phone or switch apps. For consultants, sales teams, and anyone who lives in airports, that is the killer feature. From the Capterra reviews:

"The AI tools are excellent. The AI can take meeting minutes, answer questions about your notes" Eric J., CEO (Food & Beverages), Capterra

Notion Mail

Notion Mail is the standalone email client Notion launched in 2025. The pitch is "the inbox that thinks like you," and the introduction blog is worth reading if you want the design rationale.

What the AI actually does in Mail:

• Auto-labels and sorts incoming email based on what you tell it matters (tag senders, key phrases, "Urgent / Needs reply" detection).
• Drafts replies grounded in your Notion workspace, so a customer-success-style email about your refund policy can pull from your wiki rather than hallucinating.
• One-click snippets for repeated responses, with calendar integration that drops scheduling links into a draft without opening another tab.

Per the April 17, 2026 release notes, Mail and Calendar are now connected through a dedicated Notion AI settings tab, so a Custom Agent can both schedule the meeting and draft the email confirming it.

The Icebox honest review is worth reading if you are evaluating Mail specifically: their take is that the AI sorting works well on newsletters, receipts, and high-priority threads, but "falls short in reply generation," which matches what we have heard from teams using it in pilot.

AI writing, image generation, and database setup

The original Notion AI surface is still there.

Writing: in any page, hit space and @Notion AI, or use the AI block, to draft, summarise, translate, change tone, expand, shorten, or extract action items. The help docs describe matching an existing style guide, repurposing templates, and generating outlines from prompts. This is the bread-and-butter use case for solo Notion users.

Image generation and editing: ask Notion AI for an image in chat, iterate by describing what to change. It is fine for cover images and quick concept visuals. It is not a Midjourney replacement.

Database setup: describe what you want to track, and Notion AI will generate a database schema with properties and views you can edit. It is one of the better quality-of-life features for new users; building a CRM-lite or a content calendar from a sentence is genuinely fast.

File analysis: drop a PDF, image, or doc into the chat and ask questions about it. Useful for contract reviews, design feedback, or pulling numbers out of a deck.

Pricing in 2026

Notion's plan structure flattened this year. The big change: the standalone "Notion AI" add-on (formerly $8 / $10 per member per month) is gone for new customers. Full AI is now bundled into Business and above.

Custom Agents are billed separately on top of Business or Enterprise, at $10 per 1,000 Notion credits as an add-on, after the trial that ended May 3, 2026. The Notion 3.4 part 2 release cut credit consumption by 35-50% by adding cheaper models, so the practical cost per agent run is lower than the headline number suggests.

A few non-obvious pricing details worth flagging:

• Page history doubles each tier: 7 days on Free, 30 days on Plus, 90 days on Business, unlimited on Enterprise. If you treat Notion as long-term institutional memory, this matters.
• Private teamspaces, granular database permissions, and the premium connectors (GitHub, Asana, Salesforce) are all Business+.
• External guests: 10 on Free, unlimited above. Guests are always free and only see pages they were invited to.
• Data retention with LLM providers is 30 days on Free, Plus, and Business, and zero on Enterprise. If you are in a regulated industry, that is the line that often forces the upgrade.

For the official details and the current promotional pricing, see notion.com/pricing.

What users actually say

The Capterra reviews on page 1 through page 5 are the most consistent third-party signal we found. The product carries a 4.7/5 across roughly 2,300 reviews, but the AI sentiment is split.

The praise:

"Notion AI adds significant value by making it easy to surface existing knowledge quickly" Jiahao H., Deployment Strategist, IT Services (Capterra)

"With the introduction of Notion AI, I get to write my 1 pagers much more efficiently" Deniz K., Senior Manager, Insurance (Capterra)

"the Q&A AI feature helps me find information fast, saving time during hectic days" Alexander Z., Owner, Media Production (Capterra)

The criticism:

"Notion's AI is significantly less capable than ChatGPT. The AI is slow, and when used on pages with extensive data, it experiences severe latency, often freezing for several minutes" Kerdjou T., Founder, Management Consulting (Capterra)

"I have found the addition of AI to the app to be kind of annoying, especially when a note-taking banner would automatically pop up" Stephanie P., American Accent Coach, E-Learning (Capterra)

"The only issue is the AI: it is overdue, poor for searching and context" Andrew H., Software Developer, IT (Capterra)

The consistent themes:

1. Meeting Notes is the most-loved feature, hands down. Anyone who runs a lot of meetings finds it pays for itself.
2. Q&A is loved when the workspace is clean, and frustrating when it is not. The same Notion AI gets 5-star and 1-star reviews depending on how messy the underlying data is.
3. Performance breaks at scale. Several reviewers mentioned freezing on pages with extensive databases. This is a known scaling pain point that surfaces in third-party reviews more often than Notion's marketing acknowledges.
4. Writing assistance is competent but not differentiated. If you already pay for ChatGPT or Claude, the Notion writer is a convenience, not a reason to upgrade.
5. The AI banner is intrusive for users who do not want it. Several reviewers in recent Capterra reviews describe the AI being pushed into note-taking and editing flows even when they have not opted in.

Where Notion AI falls short

A few honest limits worth naming:

It only knows what is in Notion (plus connected tools). That is a feature, not a bug, but it means Notion AI is genuinely useful only if Notion is your central system of record. If your team writes specs in Linear, ships from GitHub, supports customers from Zendesk, and uses Notion as an occasional wiki, Notion AI will keep telling you things you already know.

It is a wrapper, mostly. Notion AI runs on GPT-5.2, Claude Opus 4.5, Claude Sonnet 4, Gemini 3, and Notion's tuned variants of those. There is no proprietary model. The differentiation is the workspace context and the agent runtime, not the LLM itself. That is fine, but it is also why "Notion AI is just a wrapper around models I already pay for" is a common Reddit complaint.

Custom Agents are powerful but expensive in real workspaces. The credit math gets meaningful fast. If you have a Custom Agent triaging hundreds of Slack messages a day plus a daily standup writer plus a database autofill, you can blow through $100 a month per workspace easily. Set a credit ceiling in the admin panel and audit usage weekly until you have a sense of your run rate.

Latency on large pages and large databases is a real issue. Multiple Capterra reviewers describe pages that "freeze for several minutes" when AI is invoked on data-heavy content. If your databases are over 10,000 rows, performance test before committing.

It is not a customer-facing AI. This is the scope thing again, but worth being explicit. Notion AI is built for internal team workflows: workspace Q&A, drafting, meetings, agents that automate ops. It does not connect to helpdesks, ticketing systems, chat widgets, or e-commerce tools, and it does not do live customer chat. If you want AI to handle customer support tickets while pulling Notion content as one knowledge source among many, that is what platforms like eesel AI are for, with the Notion integration acting as the bridge into your support stack alongside Slack and Google Drive as additional knowledge sources. They solve different problems.

Who Notion AI is for (and who should pass)

Buy Notion AI on Business if you fit most of these:

• Your team already runs on Notion (wiki, docs, projects, OKRs).
• You hold a lot of meetings and want transcripts plus summaries automatically.
• You have at least one repeating ops workflow (status reports, ticket triage, Q&A) you would offload to a Custom Agent.
• You can stomach the Custom Agent credits as a separate line item.

Skip it if:

• Notion is a side tool you check occasionally; the AI surface area cannot be built on a thin workspace.
• You already pay for ChatGPT Team or Claude for Work and your bottleneck is not "search across Slack and Drive."
• You run a heavily database-driven workspace with 10,000+ row tables and have hit Notion's existing performance ceiling.

For most readers, the honest test is: do you spend more than two hours a week reading or writing in Notion? If yes, Notion AI on Business is probably worth the upgrade. If no, the trial on Free or Plus will tell you what you need to know without committing.

Bottom line

Notion AI in 2026 is the strongest version of itself yet. The agent platform is real, the connectors actually work, the Mobile AI Notes feature is best-in-class, and Notion 3.4 made Custom Agents 35-50% cheaper to run. It still has the limits of being a workspace-first product: if your work does not live in Notion, the AI cannot help you much, and the LLM under the hood is the same one your existing tools probably already use.

If you are evaluating it: start the Business trial, build one Custom Agent against a real recurring workflow, watch the credit consumption, and see whether Q&A search returns answers you trust on the first try. That is the question worth answering, and it only takes a couple of weeks.

If you are evaluating AI for customer-facing work specifically (support, sales chat, ticket triage), Notion is not the right shape; an AI support agent platform that can pull Notion as a knowledge source while integrating with your helpdesk is closer to what you want. We have written separately about how to evaluate Notion as a workspace if that is the comparison you are running, and the eesel AI pricing page if you want a side-by-side cost picture.