Notion's Head of AI on Building Agents Your Team Can Trust: Sarah Sachs E10

Tokens don't grow on trees. So what are you actually paying for when you pay for AI agents?

In this episode, recorded live in studio, we sit down with Sarah Sachs, Head of AI Engineering at Notion (ex Google, ex Robinhood), to unpack what it means now that Notion is becoming the AI infrastructure real businesses run on. It's a practical conversation for founders, ops leads, chiefs of staff and the accidental systems people who became responsible for how the business actually runs.

What we cover:

• The $500 AI bill: when agent cost is broken, and when it's the cheapest hire you'll ever make
• Use AI for reasoning, code for plumbing: why Notion Workers exist and when a token is the wrong tool for the job
• Permissioned context as the real moat, and Notion as the system of record where humans and agents collaborate
• Agent sprawl, vendor lock-in, and staying model optional as the frontier moves
• Observe in prod or live in a hallucination: watching what your agents actually do and actually cost

Timestamps:
00:00 Cold open: "we're over-indexing on tokens"
00:30 Intro and welcome, first episode live in studio
04:05 Sarah's story: Google, Robinhood, and into AI
07:30 What Head of AI Engineering at Notion actually does
08:55 Why Notion, and AI going from add-on to core
11:23 Two years, shipping fast, surfing the wave
13:30 The $500 bill: tokens don't grow on trees
18:38 Reasoning vs plumbing: stop spending tokens to move data
21:41 Workers vs Zapier and Make
23:00 Managed agents, permissioning and model interoperability
25:44 Agent SDK, API and Dev Day
26:30 Agent sprawl and Notion as the system of record
28:50 Vendor lock-in and model optionality
31:31 Building in the open, primitives and reception
34:53 Staying fresh: the hot yoga rule

Five takeaways for your business:

1. Be deliberate about what you hand to AI. Once agents run real workflows, cost stops being abstract.
2. Reserve AI for judgement and ambiguity.
3. Redesign workflows around trusted, permissioned context instead of bolting AI onto the mess.
4. Stay model optional. Keep your context and workflows in a system of record so you can swap models as the frontier moves.
5. Observe agents in production so you catch hallucinations and runaway spend before they hit your bill.

Connect with Sarah Sachs:
LinkedIn: linkedin.com/in/sarahmsachs
X: x.com/sarahmsachs

Resources:
Notion custom agents and Workers: notion.dev
Notion MCP docs: developers.notion.com/guides/mcp/overview
Claude agents in Notion: notion.com/partners/claude
Notion Enterprise Search: notion.com/product/enterprise-search
Notion AI Starter Kit: smoothops.consulting/Notion-AI-Starter-Kit-387dbd5e653480f8a275c1dc82b691ad

Connect with your hosts:
Jerwin Parker
Tim Jeffries