Everyone has ChatGPT. Almost nobody has given it a job.
I’ve hired 69 AI agents since April. They keep books, send invoices, chase late payments, answer inbound leads, fix code and draft this blog. In September they produced 120.3 million words of work. The number of people on payroll the whole time was one.
What made it work was treating each agent the way you’d treat a new employee: a job, the context to do it, the keys it needs and no more, and a probation period.
Why hasn’t ChatGPT changed how my business runs?
ChatGPT hasn’t changed how most businesses run because nobody gave it a job. It sits in a browser tab and waits for someone to think of a question. A new employee treated that way would do the same: answer when asked, then go quiet until somebody remembers they exist.
The Federal Reserve’s 2025 Small Business Credit Survey shows the pattern. 46% of employer firms said their business or employees were using AI for work. Among those firms, 83% used it for writing or marketing, and 21% for process automation.1 Writing came first by a wide margin.
I think of it as a team of PhDs sitting in the corner of the office. They know accounting, law, logistics, pricing and copywriting. You talk to them five hours a week, whenever someone thinks to ask. A week has 168 hours.
The same team, given jobs, works all 168. Something starts the work: an email lands, an error fires, the clock hits Friday at five. The only difference between the two is whether anyone gave them a job.
Why does ChatGPT give smart answers to the wrong question?
ChatGPT gives smart answers to the wrong question because it doesn’t know your business. It doesn’t know your clients, your systems, your pricing, or what you promised a customer last week. A brilliant new hire on day one has the same problem, and nobody blames the hire for it.
You fix it the way you’d fix it for a person. Tell it who you are, who the client is and how the job is done here. Give it the logins the job needs. Then it can do the job instead of guessing at one.
Most of the work of hiring an agent is this part. The model is the easy bit.
What is an AI employee?
An AI employee is an AI agent with one job. Something starts it without you asking, it knows how that job is done at your company, it can reach the systems the job needs, and it has written limits on what it does alone. Plenty of software is sold as an “AI employee.” The job is what makes one.
Microsoft’s 2025 Work Trend Index uses the same frame. Its advice to leaders is to “treat these digital employees like you would any team member: onboard them, assign ownership responsibilities, and measure their performance.” The same report found 82% of leaders confident they’d use “digital labor” to expand workforce capacity within 12 to 18 months.2
How do you hire and onboard an AI agent?
Hiring an AI agent takes the same four things as hiring a person: what starts the work, what it knows, what it can touch, and what it’s allowed to do on its own. Write those four down for one job and you have a job description an agent can follow.
NVIDIA CEO Jensen Huang made the same comparison in his CES 2025 keynote: “the way that you would bring these specialized agents into your […] company is to onboard them just like you onboard an employee.” He added that “in a lot of ways the IT department of every company is going to be the HR department of AI agents in the future.”3 For a 50-person company without an IT department, that HR job lands on the owner.
What starts it. A person shows up at nine. An agent needs a trigger: an email lands, something breaks, the first of the month arrives. If the answer is “when I remember to ask,” you’ve hired a chatbot. My invoicing agent starts on the first of the month. My collections agent starts before I’m asked to chase a payment.
What it knows. Who you are, who the client is, and how this job is done at this company. The definitions matter more than you’d expect. The weekly booking report at a 53-person travel company applies one agreed definition of “booked,” so the number means the same thing every Friday.
What it can touch. The books, the inbox, the calendar. Only what the job needs. My collections agent reads the bank account, because it has to see money arrive. It has no permission to move money, because that isn’t its job. The booking report reads its database and can’t change it. The rest of my access rules are in AI agent security: how to give agents access safely.
What it’s allowed to do. On day one, it drafts and you approve. Once it has proven itself on that job, it runs that job alone. Everything outside the job still waits for you.
If you can’t fill in all four for a job, the job isn’t ready for an agent. It usually isn’t ready for a person either.
How do you know when to trust an AI agent to work alone?
You trust an AI agent to work alone the same way you trust a new hire: after it has done the job right, in front of you, enough times that checking it stops finding anything. You don’t hand a new employee the company card on day one. An agent climbs the same ladder, one job at a time.
My invoicing agent is the clearest example I have. On the first of the month it pulls last month’s sales from the billing sheet, works out what’s owed, checks for stale drafts, then creates the invoice in Stripe.
| Month | Who sent the invoice |
|---|---|
| June | It drafted, I checked and sent |
| July | It drafted, I checked and sent |
| August | It drafted, I checked and sent |
| September | It sent on its own |
| October | It sent on its own |
Three months of drafts before one invoice went out alone. Now it sends when the numbers check out and holds the draft for me when they don’t. I can take the job back any day I want.
Anthropic measured the same climb in Claude Code, its coding agent. In a February 2026 study, newer users let the agent run with full auto-approve in roughly 20% of sessions; by 750 sessions, it was over 40%. Anthropic’s read: “This shift is gradual, suggesting a steady accumulation of trust.” Experienced users also interrupted more often, about 9% of turns against 5% for new users. They stopped approving every step and started watching the work, stepping in when something went wrong.4
OpenAI’s practical guide to building agents puts the rule plainly: actions “that are sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent’s reliability grows.” Its examples include large refunds and payments. Invoices belong on that list.
The ladder is per job. My invoicing agent sends alone; my collections agent, a rung lower, still drafts every reminder for me to send. Trust earned on one job doesn’t transfer to the next one, for an agent or for a person.
What jobs can an AI agent actually do?
AI agents can do any job you can describe end to end: what starts it, what it does, and what comes out. Mine send invoices, chase late payments, post weekly reports, write blog posts and lay out direct mail. Here are six I run, from simple to complicated.
Invoicing. Starts on the first of the month. Pulls the sales, does the math, checks for stale drafts, creates and sends the invoice. It has sent on its own since September.
Collections. Starts before I’m asked to chase a payment. It reads the bank account, not a spreadsheet, and matches every incoming payment to an open invoice. When one is overdue, it drafts the reminder in my voice and asks whether to send it. It also tells me when nothing has arrived, so I’m not chasing money that already landed.
The weekly booking report. At the travel company, it starts Friday at five. It queries the booking database read-only, applies the agreed definition of “booked,” and posts the week to the team’s chat with every order listed. Nobody asks for it. It just shows up.
The blog writer. For a fitness app I co-founded, it starts every weekday morning. It picks a topic people search for and writes the post, a second agent checks every fact against its sources, and it publishes in three languages. That’s about 150 posts since February. Google’s own numbers show visits from Google at 33 times what they were in February, and 1.25 million appearances in Google results over the last three months. ChatGPT and similar tools read about 1,380 of its pages a day, and 80% of the app’s downloads now come from Google and ChatGPT.
The direct-mail team. Starts when a campaign is approved. One agent writes the copy and another lays it out print-ready for three locations, with compliance lines, QR codes, bleed and trim marks. A person signs off before anything is printed or mailed.
The apartment. Starts with one sentence and a folder. An agent read the floor plan, the 3D scans, the photos and the receipts in my housing folder, modeled every room and every piece of furniture to size, and wired the model to the home’s sensors so the temperatures are live. I include it because it shows how far “describe the job” goes.
At the travel company, 13 agents run 17 jobs that start on their own. Last month they finished 290 tasks, half of them after hours. When something breaks on the site, a fix is under way before anyone has seen the error. When someone asks for numbers in the team chat, the answer is back in thirty seconds. On the first of the month, reports are posted and invoices matched to the bank before anyone logs in. I was gone for two weeks in September. Nobody noticed.
How much does an AI employee cost compared with a person?
The thinking an AI employee does costs almost nothing; the setup is where the money goes. Writing 1,000 careful emails to real customers takes a person about six weeks at fifteen minutes each. ChatGPT writes the same thousand for about $3.5 My collections agent costs me less than $200 a month to run. The real cost is plugging it into your business, one job at a time.
That’s why my own agents’ words of work look the way they do. In July they produced 8 million words of work. August was 44.6 million, September 120.3 million, and the first seventeen days of October 180.7 million. One person on payroll the whole time.
Growing used to mean hiring. Now it can also mean giving the intelligence you already pay for a job.
How fast should a small business add AI agents?
Add one AI agent a month, each with one job. A year from now, twelve jobs run without you. Hiring one at a time keeps each agent on its own probation, and each job you write down makes the next one easier to describe.
A year of hires for an owner-led company could run in this order: quote requests answered by morning, invoices sent when the numbers check out, late invoices chased, the Monday numbers posted, the inbox sorted every two hours, reviews answered in your voice, supplier price changes flagged, the month-end close drafted, new-hire paperwork prepared, website updates made, errors triaged before anyone sees them, and the gap in the schedule filled.
The timing matters more than it looks. Hire one agent a month for three years and you get 666 months of work done. Wait a year to start and you get 300.6 Waiting a year doesn’t cost you a year. It costs more than half the work.
When is an AI agent the wrong hire?
An AI agent is the wrong hire when the job can’t be written down. If nobody can say what starts the job, what a good result looks like, or what the agent may do alone, the agent will guess, and a confident guess is worse than no work at all.
A few other cases where I’d keep a person on it, or keep it in ChatGPT:
- The job happens twice a year. Writing it down costs more than doing it.
- A wrong answer is expensive and you can’t check it cheaply. Agents earn trust by being checked. Without a cheap check, they can’t climb the ladder.
- The systems have no way in. If the job lives in software with no login an agent can use and no export, someone has to build that access first.
- Nobody owns it. An agent needs a person to bring its questions to. A job nobody owns stays stuck, whoever does it.
Where the line sits between a ChatGPT task and an agent is its own question, and I wrote it up in AI agent vs ChatGPT in 2026.
You may not need outside help either. If someone on your team can write the four things down and owns the job, start there with the tools you already pay for.
Which job still needs you every week? Start with that one. If you’d like help hiring for it, that’s what Fifth Nine’s AI automation work does.
Sources
How this was written: drafted with AI from my notes and Fifth Nine’s operating records, then edited and approved by me.
Footnotes
-
The 2025 Small Business Credit Survey was fielded September 3 to November 14, 2025 and published in March 2026 as the 2026 Report on Employer Firms (pages 26 and 27). The AI-use question had 5,248 employer-firm responses; the question about uses had 2,225 responses from firms using AI. ↩
-
Microsoft, 2025 Work Trend Index Annual Report, April 2025: a survey of 31,000 knowledge workers in 31 markets, February 6 to March 24, 2025. The 82% figure is on page 4; the “digital employees” advice is on printed page 18 (PDF page 11). ↩
-
Transcribed from NVIDIA’s own video of the January 6, 2025 keynote, at about 39 and 40 minutes in. NVIDIA’s written recap of the keynote doesn’t include these lines. ↩
-
Anthropic, Measuring AI agent autonomy in practice, February 18, 2026. It covers Claude Code users who signed up after September 19, 2025. Anthropic notes that the default settings require approving each action, “so part of this transition may reflect users configuring the product,” and that in fields like finance, checking an agent’s work “may require significant effort, which could slow the development of trust.” ↩
-
1,000 emails at fifteen minutes each is 250 hours, a little over six 40-hour weeks. The $3 is my figure. As a check at OpenAI’s published API prices, assuming about 2,000 tokens in and 500 out per email, the same thousand emails cost $0.45 on GPT-6 Luna and $9 on GPT-6.1 Sol, before any reasoning tokens. ↩
-
One agent hired a month for 36 months: the first runs 36 months, the second 35, and so on down to 1, for 666 agent-months of work. Starting a year later leaves 24 months: 24 + 23 + … + 1 = 300. 300 is 45% of 666. ↩
Put AI to work with your team.
A free session using one real workflow from your business. Start with a 30-minute planning call.