You’re staring at your inbox at 11pm, and somewhere in the last hour you also had the thought: I should probably be “doing AI” in this company. Not vaguely. You’ve read the posts, you’ve heard the podcast, you know your competitors are figuring something out. The problem is you don’t know where to start, and every option in front of you looks like a project: a rollout, a policy, a vendor call.
So nothing happens. The inbox stays full, and the AI thought sits there next to it, unresolved, same as everything else you meant to get to this week.
Here’s the move that actually works, and it’s smaller than you think it needs to be.
Why a narrow AI task is the safe place to start
The instinct to build a strategy first feels responsible. It’s actually the thing that stalls you. A strategy is a bet on a dozen decisions at once, which tool, which team, which process, before you’ve tested any of them on real work. You’re trying to design the whole org chart before you’ve hired anyone.
Compare that to picking one task. A single, bounded piece of work, small enough that you could point at it and say “that, right there.” Something with a clear input, a clear output, and a fast way to check if it’s right. If it works, you know in days, not quarters. If it doesn’t, you’ve lost an afternoon, not a rollout budget.
The other reason to start narrow is trust. You don’t hand a new hire the keys on day one, and you shouldn’t hand AI the keys either. A bounded task lets you watch it work before you widen what it touches. That’s not caution for its own sake. It’s how you actually build the judgment to know what this thing is good at and where it still needs you standing over its shoulder.
The pattern: one bounded task, using email triage as the example
Take your inbox. Not because email is glamorous, but because almost every founder has the same version of this problem and it’s easy to see clearly.
Right now, you’re the filter. Every message gets read by you, sorted by you, and either answered, ignored, or left to rot for four days until it becomes urgent, which is attention spent on sorting instead of on the ten messages that actually needed your judgment.
The bounded task looks like this: point Claude at the inbox with one job, read what came in, flag what genuinely needs a reply today, and draft a first-pass response for the routine ones, the scheduling requests, the “just checking in,” the ones you’d answer the same way every time anyway. You read the drafts before anything goes out. Nothing ships without your eyes on it. The task isn’t “manage my email.” The task is “sort and draft,” full stop, and you can check whether it’s doing that job correctly in about thirty seconds a day.
Run it for two weeks. You’ll know fast whether it’s catching the right things, missing anything that matters, or drafting replies that sound nothing like you. All three are fixable, and none of them cost you a client while you find out.
Notice what you didn’t have to do to get here: no policy document, no all-hands announcement about “our AI initiative,” no vendor evaluation spreadsheet. You picked a task you already do every day, handed the sorting and the first draft to a tool, and kept your hand on the send button. That’s the entire setup, and most founders talk themselves out of it because it feels too small to count as doing something about AI. It counts. It’s the only version of “doing something about AI” that actually produces evidence instead of intentions.
This is the same shape as a real gap I closed for myself. I built a teammate, Tess, whose only job is chasing down commitments I make on calls so they don’t quietly die three weeks later. One task, one owner, checkable in a glance. Your version doesn’t have to be follow-through. For a lot of founders it starts right here, in the inbox, because that’s where the daily drain is most visible.
The growth path: from this one job to a seat, and eventually an org
Here’s where this connects to the only framework worth remembering: job, seat, org.
A job is one task. What you just built is a job, sort the inbox, draft the routine replies. Small enough to point at.
A seat is a stack of jobs that adds up to a hire. Once email triage is holding up, you look at what sits next to it. Meeting notes that need turning into next steps. First-draft follow-ups on sales conversations. Tagging inbound leads before they hit your CRM. Stack three or four of those together and you’re not looking at a task anymore. You’re looking at an inbox-and-follow-up seat, doing a meaningful slice of the hire you were about to make.
An org is every seat you’ve built, running around you. You don’t get there by designing it up front. You get there by proving one job, then the next, until the seats start adding up on their own.
That’s the whole model, and it’s the same one whether you’re closing your own gap or building out the first few seats of your company’s AI team. You don’t need the org chart today. You need the first job to work, checked, for two weeks straight.
So which task is sitting in your inbox right now, doing to you exactly what triage used to do to me?
If you want a second set of eyes on which task to pick first, grab fifteen minutes with me. No pitch. Just a look at where the drain actually is and which job to hand off first.
Frequently Asked Questions
What’s the safest first AI task for a small company to automate?
Something bounded, low-risk, and easy to check. Email triage is the clearest example: read the inbox, flag what actually needs a reply today, draft a first-pass response for the routine ones. You can verify the output in seconds and nothing ships without you seeing it first.
Do I need an AI strategy before I start using AI in my company?
No. A strategy document doesn’t move anything. Picking one real task, automating it, and watching whether it holds up for two weeks tells you more than a slide deck ever will. The strategy can come later, built from what actually worked.
How do I know if a task is safe enough to hand to AI first?
Ask if you’d notice a mistake fast and if the cost of that mistake is small. Drafting a reply you still approve, sorting an inbox, tagging a lead. Anything where a wrong answer is embarrassing but not expensive is a safe place to start. Save the higher-stakes work for once you’ve built some trust in the tool and the process around it.
What happens after the first task is working?
You add the next job in the same lane. Email triage plus meeting notes plus first-draft follow-ups starts to look less like a task and more like a seat, an inbox-and-follow-up seat, running most of a role you’d otherwise have hired for. That’s the job to seat to org path, one bounded task at a time.
Isn’t starting this small too slow for a company that needs to move fast?
It’s the fast path, not the slow one. Rolling out AI everywhere at once means you can’t tell what’s working, and a bad rollout costs you more time to unwind than a narrow one ever would. One task proven in a week beats a six-week rollout plan that stalls in a meeting.