How to run a one-person GTM team in 2026 with AI agents

Look, the honest version of this is not "fire your GTM team and buy agents."

Most founders reading this don't have a team to fire. You have a product, a website you're slightly embarrassed by, maybe a few customers who found you by accident, and no real idea what you're supposed to do on Monday to make next quarter less frightening.

That's the actual problem. Not a million dollars. Monday.

I run Intempt, so I have a dog in this fight and you should read the rest knowing that. Most of what follows works with tools you already pay for. I'll tell you the one place where I think our category earns its keep, and you can skip that part if you want. If you have two or three people rather than one, the GTM team without hiring covers the same split at team size.

Start with the math, not the tools

Everyone starts with tools. Start with the number instead, because the number tells you which jobs matter and which ones you're doing out of anxiety.

Say you want $1M ARR and your average deal is $12,000 a year. That's 84 customers. Call it seven new customers a month.

Work backwards. If one in four real opportunities closes, you need 28 opportunities a month. If half your first meetings become an opportunity, that's 56 first meetings a month. Fourteen a week. About three a day, every working day, forever.

Now the part everyone skips. Where do 56 meetings a month come from?

If outbound is your only channel, the numbers are public and they aren't kind. Instantly's 2026 cold email benchmark, drawn from billions of cold emails across thousands of workspaces in 2025, puts the average reply rate at 3.43%. Top quartile is 5.5%. The top 10% clear 10.7%. And in my experience roughly half of any reply pile is neutral or negative: out of office, not interested, please stop.

So a genuinely good campaign might produce a 1.5% positive reply rate, and maybe half of those book something. Call it 0.75% from contact to meeting, being generous to yourself.

56 meetings divided by 0.0075 is about 7,500 contacts a month. Seven and a half thousand people. Researched, written to, followed up with, every month, by you, alone, while also building the product and answering support.

That's the real reason companies hire. Not ambition. Arithmetic.

What that number is actually telling you

Two things, and neither is "work harder."

The first is that outbound can't be your only channel. Not because outbound is dead, but because being a single person at the top of that funnel means every meeting costs you roughly 133 researched contacts. The second channel doesn't have to be big. It has to compound while you sleep, which means search, or a product that spreads itself, or an audience you're building anyway.

The second thing is subtler and it's the whole argument for agents. Of those 7,500 contacts, the part that needs you is tiny. Deciding who's worth contacting needs you. The judgment call on a reply needs you. The rest, finding them, checking they're real, writing the first version, remembering to follow up on day four, is work that has a checkable output and no judgment in it.

That's the line. Not "what can AI do." What has a checkable output.

The five jobs, and which half you keep

A GTM team, whatever its size, does five jobs every week. Not five tools. Five jobs.

I'm going to go through them in the order they actually happen, and for each one name the half you hand over and the half you keep. The handover half is where agents earn their money. The keep half is where founders who hand it over anyway end up with a pipeline full of nothing.

Job one: deciding who to talk to

You keep: the definition. What a good customer looks like, written down, in specifics that someone else could apply without asking you.

The agent takes: finding people who match it, checking they still work there, checking the company is real and funded and the right size, and scoring the list so the top decile is at the top.

Most people get this backwards. They let the tool define the segment, because the tool has filters and filters feel like thinking. Then they wonder why 7,500 contacts produced four meetings.

The definition is the most valuable hour of your month. Here's the prompt I'd actually run, and it works in any decent model:

Here are my last 20 closed-won customers and my last 40 closed-lost.

[paste both, with company size, industry, what they bought, and why they said yes or no]

Find what actually separates them. Not what I'd like to be true.

For each signal, tell me whether I can get it from a public source or whether I'd have to ask.

Then write the definition as rules someone else could apply without me in the room.

The output you want is not a persona. Personas are decoration. You want five or six rules, each one checkable by looking at a company for thirty seconds.

Job two: getting in front of them

You keep: the offer and the angle. What you're saying, and the reason a stranger should care this week rather than in general.

The agent takes: the volume. Drafting the first version against each person's actual context, sequencing it, following up on a schedule, stopping when someone replies, and never sending twice to the same person.

Here's where most one-person operations quietly break. Not on the writing. On the follow-up. The first email is exciting and the fourth one is admin, so the fourth one doesn't happen. Instantly's same benchmark found follow-ups produce 42% of all replies. Skip them and you give away close to half of what the campaign could have returned.

An agent doesn't get bored on day four. That's most of the value, and it's unglamorous.

What you should not hand over: the decision about what to say. A model asked to write a cold email with no angle will produce something that reads like every other cold email, which is part of why reply rates sit where they do.

Job three: running the conversation

You keep: all of it. Every meeting, every negotiation, every moment where somebody is deciding whether to trust you.

The agent takes: everything around it. The pre-call brief on who you're about to meet. The notes. The follow-up written from what was actually said rather than what you half remember. The CRM update you'd otherwise do on Sunday.

This is the highest return job on the list, and almost nobody sets it up, because it doesn't feel like growth. It just quietly returns hours every week and makes you look considerably more organised than you are.

Job four: keeping the ones who bought

You keep: the relationship, and the decision about what a struggling customer needs.

The agent takes: noticing. Usage dropped. The champion's title changed. Nobody logged in for eleven days. Renewal is in sixty and there's been no contact in thirty.

For most early companies, this job doesn't exist at all, which is expensive in a way that's invisible until a renewal disappears. Churn doesn't announce itself. It shows up in behaviour weeks before it shows up in a cancellation, and behaviour is exactly the thing software is good at watching and humans are bad at watching.

Job five: knowing what happened

You keep: the decision about what to change next.

The agent takes: the weekly read. What moved more than normal variance, what didn't, and the two most likely explanations you can go and check today. Ask for it this way:

Here's our funnel for the last five weeks. [attach]

Compare this week against the trailing four-week average.

Only tell me what moved more than normal week-to-week variance.

For each one, give me the two most likely causes I can check today.

Skip everything that's just noise.

That last line matters more than it looks. Weekly AI reporting fails by telling you everything changed, which leaves you exactly where you started, only now with a dashboard.

Where agents stop working

Here's the part the one-person-business posts don't get to, because they're written by people whose product is a course.

Run the five jobs above for a month, and you'll hit a wall that has nothing to do with the quality of the model. You start every session by explaining your business again. You paste the same customer definition into the targeting prompt. You paste the same customer list into the scoring prompt. You paste the same context into the follow-up prompt, and the agent writes a follow-up that contradicts the email you sent on Tuesday, because it never saw Tuesday.

A prompt holds method. It doesn't hold memory.

By week three, you notice you're doing data entry to avoid data entry, and the agents are only as good as your patience for re-explaining yourself. I wrote up how to build a GTM brain for exactly this wall: it's a retrieval problem, not a storage one.

That's the real dividing line in this category right now, and it's worth understanding before you buy anything. Method is portable. Anyone can copy a good prompt, and you should. Your customer data isn't portable, and it's the thing that makes the same prompt produce a useful answer for you and a generic one for your competitor.

This is where I'll name my own thing exactly once. Intempt is agents sitting on top of the customer data, so the thing writing the follow-up already knows what that account did last week. If you want to test whether that's worth anything before you believe me, run the prompts above by hand for two weeks first. You'll find the wall on your own, and then you'll know what you're actually buying.

What actually breaks

Four things, and I'd rather you hear them from me than find them in month three.

• Agents don't make you good at the job. They execute a method. If your customer definition is wrong, you now get wrong answers faster and with more confidence, and confident wrong is more expensive than slow wrong.
• Everything is a first draft. Read the email before it sends. The founders who get burned are the ones who stopped reading around week three, right at the point where it started working.
• Volume isn't the constraint you think it is. It's tempting to read the 7,500 number and conclude the answer is 15,000. It isn't. Doubling a bad list doubles the damage to your domain and your reputation, and both are harder to repair than a slow quarter.
• You'll over-automate the wrong job. Almost everyone automates writing first, because it's visible and satisfying. Writing is the job that most needs you. Follow-up, research, and noticing are the jobs that don't, and they're boring, which is precisely why they're still undone.

Your first week

Not a quarter. A week.

1. 1Pick the number. One real revenue target, worked backwards to meetings a month, on paper. Ten minutes, and most people have never done it.
2. 2Write the definition. Run the closed-won and closed-lost prompt. Get to five checkable rules.
3. 3Automate exactly one job, and make it follow-up. Not writing. Follow-up.
4. 4Run it for two weeks by hand where you have to, so you learn where it actually breaks in your business rather than where it broke in mine.
5. 5Fix the thing that broke, and only then add the second job.

That's the play. The one-person GTM team isn't one person doing five jobs. It's one person doing the judgment half of five jobs, and being ruthless about which half that is. If your one person is a marketer rather than a founder, the one-marketer stack is the same idea from the marketing side.

Tell me where I've got this wrong. I'm genuinely still figuring out job four.