AI Employees: What You Are Actually Buying

Products sold as AI employees arrive with a first name, a headshot, a job title and a monthly price that sits neatly below a salary. The framing does a lot of work, and it is worth taking apart before the comparison to payroll gets made. OpenAI published a figure this week that gives a sense of the upper bound: since June, it says, agent runtime across its research organisation has exceeded human labour, at roughly 3.1 agent-workdays for every human workday.

What the OpenAI figure does and does not show

Take the 3.1 seriously and read it carefully. It is OpenAI's own number about OpenAI's own research organisation, which is the most favourable environment for agent work that exists: the staff build the models, the tasks are software and research, and the tolerance for a failed run is high.

It also measures runtime converted into workday equivalents rather than output or value. Agents run in parallel and overnight, so a figure expressed in workdays inflates easily against humans who do neither. A firm doing client work with deadlines and liability sits a long way from that setting.

What it does establish is that agent labour at real scale is happening somewhere, not just in pitch decks. Treat it as the ceiling under laboratory conditions, and your own measurement as the number that matters.

The three products behind the label

Outbound sales. The most crowded category. The agent researches prospects, writes personalised messages, sends sequences and books meetings. It genuinely produces volume. Whether volume helps depends on whether your problem was not enough outreach or not enough good outreach, and for most small firms it was the second.

First-line support. Answers common questions from your documentation, handles simple account actions, escalates the rest. This is the most mature of the three and the easiest to measure, because deflection rate and customer satisfaction are already numbers you track.

Back-office operations. Scheduling, data entry, reconciliation, chasing missing information. Least glamorous, often the best return, because the work is high volume and the correct answer is unambiguous. It is also where verification matters most, which we went into in extracting data from PDFs.

What per-seat pricing hides

These products cost the vendor per token and charge you per seat, and that gap has to be managed somewhere. It usually appears as a usage cap in the contract, expressed in conversations, tasks or credits per month, with an overage rate that is easy to miss.

Work out your real monthly volume before the demo and ask what happens at three times that number, because a busy month is exactly when you want the thing working. The related trap of unpredictable consumption billing is covered in consumption pricing and unpredictable bills.

The second hidden cost is supervision. Somebody reviews the output, handles what the agent could not, and fixes what it got wrong. At five hours a week that is a meaningful fraction of a salary, and it belongs in the comparison rather than in the assumption that the product runs itself.

Four things the label promises and the product does not

Accountability. When a person makes an expensive mistake, someone answers for it and something changes. When an agent does, you have a log and a support ticket. That difference matters most in exactly the regulated and client-facing work people most want to automate.

Knowing when to stop. A junior member of staff who senses a conversation going wrong asks someone. An agent proceeds confidently unless you built the escalation rule, and you will not think of every case that needs one.

Relationships. In small business sales, the person on the other end often buys because they trust someone. An agent can carry the process and cannot carry that, and pretending otherwise annoys the customers you already have.

Noticing that the process changed. When a supplier changes a form or a regulation shifts, a person adapts and mentions it. An agent keeps following the old configuration until someone notices the outputs have quietly gone wrong, which argues for the monitoring approach in watch what it does, not what it says.

The number to ask for

One question separates these products more than any feature list: what fraction of real cases does it finish without a person. Research published this month found that only 2.6 percent of catalogued agent tools complete a recorded work task end to end, with most of the rest assisting instead, which we unpacked in only 2.6% of agent tools finish a whole task.

Assisting is a legitimate and valuable product. It is a different purchase from replacing a role, and the pricing rarely distinguishes between them. Ask for the completion rate on customers who look like you, and ask for a case that failed.

A four-week trial that settles it

1. Pick one workload with a countable volume. Inbound enquiries, invoices, appointment confirmations. Something you can count last month's of.

2. Run it in parallel, not instead. The agent handles the same queue your current process does, for four weeks, with nobody depending on it yet.

3. Record three numbers. Completed unaided, needed a person, and total review minutes. Those three settle the business case, and none of them come from the vendor.

4. Give it a scoped account before it touches anything live. Its own login, access to only the systems the task needs, an approval gate on anything irreversible, and a log. The framework is in three boxes for an AI agent.

Put the three numbers into the ROI calculator alongside the subscription and the supervision hours. Plenty of these products clear that bar comfortably. The ones that do not tend to fail on the review time rather than on the subscription, which is the line item the pricing page never mentions.

Frequently Asked Questions