What an AI Agent Actually Costs (The Honest Math)

Most owners price an AI agent the way they price a software subscription. They open two tabs, compare monthly fees, and pick the cheaper one.

That is the wrong comparison, and it flatters us. An agent is not a seat licence. It is a worker with a job description. So the number it should be measured against is not another vendor’s monthly fee. It is what it costs you to put a person in that seat.

The number you are actually comparing against

A salary is not a cost. It is the visible part of a cost.

Add employer payroll taxes. Add health cover and any retirement contribution. Add the software seats that person needs, the phone line, the share of rent and equipment, the recruiting fee spread across their tenure, and the manager hours spent hiring and training them.

Across small B2B service firms, that loading typically lands between 1.25x and 1.4x base salary. A $50,000 role is a $62,500 to $70,000 commitment. A $75,000 role is a $93,750 to $105,000 commitment. Firms that skip this arithmetic consistently underestimate their own headcount cost by a fifth or more.

So when you evaluate an agent, put it beside the loaded figure. Not the salary line.

A worked example

Take a customer-support role in a 20-person service firm. Base salary $48,000. Apply a 1.3x load and add the seats and space that come with a desk. These are illustrative figures, not a quotation, and your own loading will differ. Run yours.

The loaded human hire comes to roughly $4,000 a month in salary, about $1,000 a month in payroll tax and benefits, and around $200 a month in software seats, phone and allocated space. That is $5,200 a month, or $62,400 a year. An AI support agent at $499 a month is $5,988 a year, with the seats and space included. The gross annual difference is about $56,412.

Hold that number loosely. It is the gap before the costs that never appear on a pricing page. Those come next, and they are the honest part.

The costs people forget

The loading period. An agent arrives knowing nothing about your business. Somebody has to feed it your policies, your product list, your refund rules, your tone, and a body of past tickets so it learns how you actually answer. For a tidy business with documented processes, that is a few days. Where the knowledge lives in three people’s heads, budget a few weeks. This is real internal time, usually from your best person, and it is the single most common line buyers leave out.

Approval time. For the first stretch, a human reads outbound messages before they go. Say ten seconds per message on a queue of eighty a day. That is roughly fifteen minutes a day, or five to six hours a month, of someone’s attention. It shrinks as trust builds, and on low-risk categories it can go to zero. It rarely goes to zero everywhere, and you should not want it to.

Integration work. If you run mainstream tools, connection is configuration. If you run a bespoke practice-management system, an old on-premise database, or a CRM somebody’s nephew built in 2013, that is a project. Get it scoped before you sign, not after.

Governance. Setting spend and action limits, defining what the agent may never do unsupervised, keeping a reviewable audit trail of what it did and why. This costs a few hours to establish and a small ongoing habit to maintain. It is also what makes the difference between an agent you can defend to a client and one you quietly switch off after an incident.

Put a realistic figure against those. In our worked example, call it 30 hours of internal loading time and six hours a month of review. At a $45 loaded internal hourly rate, that is roughly $1,350 in year-one setup plus $270 a month ongoing. The agent still comes in around $10,500 all-in for year one against $62,400. The gap narrows. It does not close.

When the math does not work

There are four situations where I will tell an owner not to buy.

Volume is too low. If the task happens eight times a month, the loading period will never earn itself back. Automation rewards repetition. Below a few hundred instances a year, the arithmetic is usually against you.

The work is mostly judgment. If every case is genuinely different, and the value is in reading a situation rather than executing a known response, an agent will produce plausible answers to the wrong question. Judgment work can be assisted. It should not be delegated.

Nobody has written the process down. If you cannot describe the task as a sequence of decisions, the agent cannot either. This is fixable, but the fix is writing it down, and that has to happen first. An agent will faithfully reproduce a process you never agreed on.

The owner will not delegate approval. If you intend to read every message forever, you have bought a drafting tool, not a worker. That may still be worth $499. It is not worth $62,400 of avoided salary, and you should price it accordingly.

The five-minute payback formula

Sit down with a pen. Four numbers.

One. Hours reclaimed per month. Instances per month multiplied by minutes each, divided by 60. Be conservative. Halve your first instinct.

Two. Your loaded hourly cost. Annual loaded cost of the person doing this work, divided by 2,000. A $62,000 loaded salary is $31 an hour.

Three. Monthly value reclaimed. Line one times line two.

Four. Monthly cost. Agent fee plus your monthly review time valued at line two.

Then: payback in months equals one-off setup cost divided by monthly value minus monthly cost.

Worked through: 90 hours a month reclaimed, at $31, is $2,790. An agent at $499 plus six review hours at $31 is $685. Net monthly gain $2,105. Setup cost $1,350. Payback: under one month.

If your payback comes out under six months, the case is strong. Six to twelve, it is worth doing but sequence it carefully. Beyond twelve months, the honest answer is usually that you have picked the wrong process to start with. Pick another one.

Where to go next

If you would rather not do this on paper, the free AI ROI and Payback Calculator on our resources page runs the same arithmetic with your numbers, in three cases, and shows the assumptions it is using so you can argue with them.

And if you are not yet sure which process to point it at, start with the AI Readiness Assessment. It takes about fifteen minutes and it will tell you plainly whether your processes are documented enough for this to work, which, as the section above says, is the thing that actually decides it.

Both are free at www.agentic-ai-adoption-now.com/resources. No call required.