“AI agents for business” is what everyone searches for, so it’s the phrase in the title. The honest answer behind the search: most of what’s sold under that phrase will make you busier, not better off. The ones worth paying for hold a job — prospecting, bookkeeping, follow-up — start themselves, finish the work, and report on a number to a person. Some people call them agents. We call them digital employees, because the word changes what you end up buying.
This piece answers one buyer’s question fully: what the word means, the seats worth filling first, what it costs, what results look like, how to buy without getting burned, and where not to put one.
What “AI agent” means right now — and why that’s a problem
“Agent” is the most abused word in AI. Peter sat on a panel at DC Tech Week with the founders of five companies that all called themselves agentic, and asked each one to define the word. None could.
A word with no definition can be attached to anything. A category with no edges can’t go on an org chart, can’t be priced against a salary, and can’t be held to a number. That’s why so much of it never gets deployed: nobody knows where it goes.
So when you go looking for AI agents for your business, you get everything from a browser extension to a six-figure platform, all wearing the same word. (Peter has written more on why the word itself causes the trouble in Stop Calling It an Agent.)
Agents do tasks. Digital employees do jobs.
The line we draw: agents do tasks. Digital employees do jobs.
A task is “draft this email.” A job is “keep every lead under 48 hours from first contact to booked call, and tell me Friday how many you booked.” The task needs you at both ends — you start it, you check it, you carry it somewhere. The job needs you at neither.
Run anything you’re being pitched through the four tests. Does it start itself? Does it finish? Does it have one number? Does it report to a person when it can’t hit it? Most of what’s sold as an “AI agent” fails the first one. (The full definitions and tests are in What Is an AI Employee? — read that one first if you’re new here.)
The five seats worth filling first
Where does one pay for itself? In the seat nobody filled, doing work that makes or saves money the moment it’s done. Repeatable work, with a number attached:
- Prospecting and follow-up. Find the leads, write the outreach, chase the ones who went quiet. Number: booked calls per week.
- Bookkeeping and reporting. Watch the books, flag what’s off, produce the Friday numbers. Number: a clean close, anomalies caught early.
- Inbox and intake. Triage what comes in, answer what it can, route what it can’t. Number: share of same-day responses.
- Content. The newsletter, the LinkedIn posting, the podcast outreach you keep meaning to start. Number: shipped per week.
- Monitoring. A market, a competitor, a portfolio. Number: changes caught before they cost you.
One job each. Single-purpose beats the one-box-does-everything bundle every time — you’re buying a hire, not a toolbox.
What it costs to run
About two dollars a day per digital employee on the open-source frontier models PSI runs. Discount the run cost to zero in your head: what you pay for is the expertise in the job description, not the electricity. For scale, Lisa Murphy’s fleet ran 33 jobs unattended over a weekend at about $1.25 a day.
Compare that with the seat you’d pay a person five figures a year to fill, if you could find and keep them. That’s the buyer’s math.
What it looks like in real businesses
Three results from PSI’s work, cleared for publication:
- Ruke at Game Changers: one digital employee supporting the sales team, call volume up 25%.
- Matt Shepherd: a digital employee watching the books full time across a short-term rental portfolio — overhead down 22%.
- Ashley Lucero runs Ashdon, a property management business with a few hundred doors. Two virtual assistants at about $19,000 a year each were replaced by a stack of digital employees at about $200 a month. As the business grew, the employees absorbed the extra work.
These are individual results, not a forecast of what you’ll get. What they show is the shape of the change: work moves off a person’s desk and onto a seat that reports back.
How to buy without getting burned
- Ask for the job description before the demo. Role, goal with a number, who it reports to. If the vendor can’t produce one, they’re selling tasks.
- Run the four tests on the demo. Watch what starts the work. If the answer is “you do,” keep your card in your pocket.
- Buy one job, not a platform. And hold the access line: the smallest set of permissions that lets it do the job. Removing a permission is a hard guardrail; an instruction is only a request.
- Set the review before it starts. Who checks its work, how often, and what happens when it’s wrong — decided before it is.
- Give it probation. Two weeks doing the full job where nothing leaves the building without a human saying yes. Then loosen the reins as the number moves.
- Hold this line everywhere: AI can be wrong, confidently. Anything that matters gets checked by a human until the employee has earned less checking.
Where not to put one
Knowing where not to put one is the skill. Keep a human where:
- The process is broken. It amplifies a process; it doesn’t fix one. If a competent person did the job exactly as written and you wouldn’t be pleased, fix the process on paper first.
- The relationship is the product. A billing problem with your biggest client. A moment that needs trust. A person.
- A wrong answer costs you before anyone can review it. Regulated, financial, medical, reputational. Build the review step, or keep the human.
- It’s the owner’s job. What the business is, who it’s for, what to refuse. AI produces the average of everything it has seen. The idea that makes your business isn’t in the average.
If you run a small business
Same rules, smaller budget. One job done properly beats a platform license you’ll open twice. The United States has 33.2 million small businesses (SBA Office of Advocacy, 2023), and nearly all of them are being told to learn AI. McKinsey’s January 2025 Superagency in the Workplace report found 92% of companies plan to increase AI investment while about 1% of leaders call their company mature at deployment. The gap between those numbers is owners who bought tasks and got busier.
Don’t learn it. Deploy it. (Peter made the small-business case at length in While You Were Sleeping.) IDC projects 2.5 billion agents in daily use by organizations by 2030 — the direction is not in dispute. Peter’s view is that every business will be AI-enabled within five years, and the ones that do best won’t be the ones who learned the most prompts. They’ll be the ones who hired first.
Frequently asked questions
Is this just ChatGPT or Zapier with extra steps?
Those are tools you operate — applied AI. It made you faster, and it kept you at both ends of every task, which is why you’re busier than before you bought it. A digital employee can be built on the same models. What changes is the unit: a job with a number and a reporting line, not a workflow you trigger.
Will it replace my staff?
It replaces tasks, and some owners do use one to avoid a hire. Most of the value we see comes from filling seats that were never filled: the follow-up, the newsletter, the reporting. Your people move to the work that needs a human — trust, judgement, relationships.
How do I know whether a vendor is selling me a job or a task?
Run the four tests: does it start itself, does it finish, does it have one number, does it report to a person when it misses. If it fails any of them, you’re buying a task — and you’re the one who will be running it.
What about my data?
Treat its access like a new hire’s: the smallest set of permissions that lets it do the job, reviewed by you, removed when the job changes. Ask the vendor exactly what the employee can touch and where your data goes. Those are operational questions; the rules for your market are a question for your attorney.
Want one deployed in your business? We diagnose the job, scope it, build the digital employee and hand it over. See how deployment works →