Gusto, Insight Partners, and Leland are doing a TechCrunch Disrupt session about whether your startup’s next teammate might be an AI agent. I read the blurb twice. The teammate word is doing a lot of load-bearing work there, and nobody on the agenda seems to be getting asked the boring follow-up question, which is what day one actually looks like.

Because I’ve done day one. Both kinds.

Day one for a human hire

New person starts Monday. IT sends a welcome doc with fifteen links. Google Workspace, Slack, Notion, Linear, Figma, HubSpot, Stripe dashboard (read-only), Metabase, Zendesk, the AWS console nobody uses anymore, a Shopify admin, the bank portal for expense receipts, Gusto itself, some HR thing for PTO, and a vendor portal with a password written on a sticky note in a shared 1Password vault.

They click through all fifteen in about ninety minutes. SSO handles most of it. By lunch they can see everything. Nobody wrote a line of code to make that happen.

Day one for the agent

You provision API keys for the three tools that have decent APIs. Then you stare at the rest.

The other twelve

This is the part the AI-as-coworker pitch skips. Roughly two thirds of the tools a normal operations person touches either have no public API, have one that costs more than the seat does, or have one that covers 20% of what the UI does. That vendor portal with the sticky note password has an API in the same sense that my toaster has an API.

So the work of “onboarding” an agent turns into an integration project. Someone has to request API access, wait for approval, pick OAuth scopes, decide whether the agent gets a service account or borrows a human identity, then get a security review because a service account with write access to your CRM is a real credential that outlives whoever created it. I wrote about the sprawl that creates in AI agent identity sprawl, and the short version is that every agent you onboard this way leaves behind a permanent set of keys that nobody remembers to rotate.

And then half the vendors want a higher pricing tier before they’ll turn the API on at all. Your teammate costs an extra $400 a month in seat licenses before it has done any work.

A human hire gets a badge that opens every door. An agent gets a lanyard that unlocks the vending machine.

API access and login access are not the same permission

I want to be precise about this because the two get conflated constantly. When you log into Linear as yourself, you get exactly what your role allows, scoped to you, revocable the second HR deactivates your account. When you mint an API token for an agent, you’ve created a second thing that has to be separately scoped, separately audited, separately killed, and that behaves differently from the UI in ways you will discover at 11pm on a Thursday.

One of those is governance you already have. The other is governance you now have to build.

What onboarding looks like when the agent uses your browser

A browser agent skips the provisioning step, and I don’t mean that as a clever workaround. It’s just how the access already works.

Dassi runs as a Chrome extension in the side panel of the browser you’re already signed into. It sees the Linear board because you’re logged into Linear. It sees the vendor portal because that sticky-note password went into Chrome’s password manager three years ago and the session has been alive ever since. There is no OAuth consent screen, no scope list to argue about, no service account sitting in an IAM console with a name like svc-agent-prod-2 that survives two reorgs and a security audit. When you leave the company and IT kills your SSO session, the agent loses access at the exact same moment you do, because it never had access of its own in the first place. That last part matters more than the convenience does. The access boundary is you.

What it costs you: the agent only works while your machine is on and the tab is open. That’s a real limitation and I’m not going to pretend otherwise. Cloud agents run at 3am. This one doesn’t.

But the tradeoff buys you the fifteen tools instead of three. I’ve covered the mechanics of why cloud agents can’t do this in Your AI Browser Agent Can’t See That You’re Logged In. You can try Dassi from the Chrome Web Store and bring your own model key, or just sign in with the ChatGPT subscription you’re already paying for.

The panel I’d actually want to watch

Someone should ask the Gusto person on that stage how they’d handle an AI hire’s access review. Not rhetorically. Gusto literally sells onboarding software, they know what the checklist looks like, and I suspect the honest answer involves a spreadsheet and some swearing.

Hell of a thing to build a category on and not solve first.