I spent 9 June at Notion Sessions Singapore. What kept me thinking was watching a company that builds AI for a living show how they actually use it on themselves. They stood up with a model you can self-assess against and a product demo to back it. Here's what they presented.
AI as abundance, not efficiency.
Estie, Notion APAC GTM lead, anchored on one belief: the most meaningful things were never built alone. Making one person ten times faster is just a cost story. What Notion is chasing is more minds on a problem, with AI sitting inside a team's shared memory instead of being stuck behind one person's chat window.
Â
A model to locate your team.
Notion's AI Transformation Model maps four levels of maturity, and the point is to find yourself on it.
Level 1 is AI as a thought partner. Glorified search. You draft in ChatGPT and copy-paste into the tools where the work actually happens.
Level 2 is AI as an assistant. Note-takers, plan drafting. A real productivity lift, but it stays stuck with your power users.
Then there's a big jump, and this is where most teams stall.
Level 3 is AI as a teammate. Value across a whole function, not just individuals. An agent running launch intake, or sitting in a team Slack channel answering with full context.
Level 4 is AI as a system. End-to-end orchestration, synced data, multiple agents, and humans passing work between each other.
Â
Singapore is ahead, but barely anyone is.
Notion surveyed 6,000 knowledge workers on how deeply their teams have actually adopted AI, and Singapore topped it. 21% of teams here are at the most advanced levels, against 12% globally. The flip side is that 88% globally are still at Levels 1 and 2, and fewer than half of teams say their AI projects deliver the outcomes they expected. The messy middle is real.
Â
Why most AI stays shallow.
Notion framed the platform as four layers. The agent layer acts. The collaboration layer adds shared visibility, versioning, and accountability. The context layer holds your docs, wikis, and connected tools. The governance layer covers permissions, security, and analytics. Buy only the agent layer and you get something impressive but shallow. All four together is what makes AI spread across an organisation. Notion's flex: OpenAI, Cursor, and Lovable run on it.
Â
Notion’s Head of Design shipped a feature himself.
Randy Hunt told the story of shipping a real feature himself. He built Presentation Mode, where you drop dividers into any page and it becomes a deck of any content, databases, videos, embeds. He planned it with Claude in a spare 15 to 20 minutes between meetings, a design engineer levelled it up, and it shipped to production within a couple of weeks as side work. His takeaway is the one that stuck: the constraint has moved from "can we build it" to "what should we even build." Roadmaps now run two to three months out, not six, because capability is emergent.
Â
The developer platform.
Nicholas Lui, a Singaporean growth engineer based in SF, demoed Notion's newest launch. At a high level it does three jobs. It syncs any data source in continuously, from Postgres and Snowflake to Salesforce, Gong, Zendesk, ServiceNow, and NetSuite, with OAuth, environment variables, and incremental syncs built in. It lets you build your own custom tools using workers, tiny TypeScript functions that let agents hit any API with no new infrastructure. And it brings multiple agents into one workspace to work side by side.
The standout was watching several agents chained through a single task, each handing off to the next. You tag a PM on a task, she tags Decagon to pull real user quotes, Claude writes the implementation plan and the PR, Codex reviews that PR before any human looks, and a human checks the work at the end. Run that across a whole task board and every item moves through the stages automatically. Notion is also building an external agent API, so any in-house agent (Ramp, Stripe, and Block are building their own) can be wrapped and appear in Notion.
Â
The part that actually stuck: how Notion learns it.
I asked Randy how Notion teaches AI internally, half-expecting a polished onboarding programme. Turns out there isn't one. It's almost all ad hoc. People demo to each other, pass around links to agents they've built, and just sit down and make workflows together. He called it "breathing the air." The only thing close to formal training is the odd 60-minute sprint when a big feature ships and everyone downs tools to play with it. What got me wasn't the system, it was the energy. The learning spreads because people are genuinely excited, not because someone booked a session. You can't manufacture that.
Â
What they sent us home with.
Three things to try: map a recurring workflow worth automating, set up a custom agent for it, and connect the data your team cares about so the agent has context to work with.
That last one is the whole game. Most teams I work with buy the agent and skip the context, then wonder why they don't see any productivity change. Half of my first sessions with a client are just mapping what data the agent needs to see before it can be useful at all.
Act 2: What the practitioners said
The keynote was Notion selling the platform. The panel was three product leaders telling the room what actually breaks when you adopt it. That second half is the part most recaps skip, and it's the part lean teams actually need.
The panel: Randy Hunt (Head of Design, Notion) moderating, with Julie Martin (Head of Product, StashAway), Tamas Gögge (Chief Product Officer, Osome), and Minh Do (COO, Animoca Brands).
The theme across all of them: capability is no longer the constraint. Accountability and team structure are.
Â
The constraint moved from "can we build it?" to "who owns it when an agent built it?"
Tamas put a number on it. Engineers at his company now write roughly 6x more code using AI. So their software development lifecycle was rewritten to say one thing plainly: the engineer still owns their solution and still has to review it, even if an agent wrote every line. "My AI screwed up" is not an acceptable answer. Someone is always accountable. The lawyer for the contract, the accountant for the books, the PM for what shipped.
That principle showed up again when the panel was asked where AI actually fails. The reframe that landed: the problem isn't when the agent fails. It's deciding which calls a human still has to make when the agent is handing you abundant, confident output. Business model, capital allocation, company direction. Those stay human even when the agent advises.
Â
The PRD ping-pong.
This was the sharpest observation of the day, and it's a failure mode nobody markets.
AI made the weak writers strong. PMs who were technically excellent but couldn't communicate verbally now produce genuinely good written docs by iterating with AI, which means non-technical execs can finally absorb what they're saying. Real unlock.
But the documents bloated. A PRD that used to be 500 words is now several times longer. So the engineer feeds that long PRD back into AI to crunch it down to bullets they can actually hold in their head. You write short, AI expands, you compress it back. Accountability leaks out somewhere in that expand-and-compress cycle, because nobody is sure which version is the source of truth.
Â
The playbook fix, the most stealable thing at the event.
Tamas's team preserves ownership without slowing down using explicit written rules. Two of them:
On design: a PM or engineer can build anything that uses only existing components. The moment something needs a new component, it requires a designer. Clean line. Everyone moves fast, design keeps ownership of taste.
On product: the playbook states PMs own what gets built, and prioritisation stays critical whether your throughput is one unit a week or two thousand. More output doesn't remove the need to choose. It raises the stakes on choosing well.
If you take one thing from this event into your own team, it's this pattern. Not the agents. The rules around the agents.
Â
Teams are getting smaller.
Julie described the structural shift. The classic squad (one PM, one designer, three to five engineers, sometimes a researcher) is breaking into groups of one to three who cross-collaborate based on what the project needs. UI-heavy work, design leads. Little UI, a developer leads with light PM and design involvement.
Everyone prototypes now. Customer care, every function, all working off "show, not tell." Which means the scarce skill is no longer producing ideas. It's curating them, so you don't drown the client in twelve options when they needed one good one.
Â
Where it genuinely doesn't work.
The honest part. Tamas named two limits.
Expectation gap. People see LinkedIn posts claiming you can build a whole company in a day, then expect that speed internally. Reality, for an existing company with legacy code and integrations, some work still takes three months even with AI. The tech applies brilliantly in some contexts and poorly in others, and pretending otherwise sets everyone up to be disappointed.
The accounting agent. His team built an end-to-end accounting agent that runs on top of their existing internal UI, no custom interface needed. It performs well at the task. But it still needs a human when it has to phone a client for missing information, and in a regulated, high-stakes domain (investments, company setup) clients want a human to confirm before a major decision. He doesn't expect everything to become an agent. The question is how much can, and where the checks have to stay.
Â
What this means for lean teams.
Most teams I work with buy the agent and stop. They skip the part the panel spent an hour on. The playbook that says who owns what. The rule for when a human reviews. The structure that keeps accountability from leaking out of the system.
Notion sold the platform in the keynote. The panel quietly explained why the platform alone won't save a team that hasn't done that work. That gap is the entire job.
After Notion's developer platform launch, Workers and Dashboards make a credible BI stack for most small teams. Why it's now my default recommendation, and where the warehouse-plus-Tableau setup still has the edge.
We built a contact enrichment pipeline for our Notion CRM using Apollo.io, HarvestAPI via Apify, and Zapier — for roughly $0–5/month. Here's the full architecture, the cost breakdown, and why CRM vendors charging $2–3K/year for the same capability is a broken model.
Notion's new Custom Agents let you automate workflows directly inside your workspace. Here's what they do, how to use them effectively, and where we think Notion is heading next.
