BlvkWare.
The arithmetic
AI employee vs hiring someone
Every page about this is written by somebody selling one. We sell them too, so
here is the sum with both sides sourced: what a person actually costs according to the Bureau of
Labor Statistics, what an agent actually costs at published prices, and a calculator that is
allowed to tell you not to buy. It does, quite often.
The short answer. A US private-industry employee costs about
1.43 times their salary once benefits are counted. An agent costs a one-off
build plus what it takes to run. Built from one of our kits, that is $79 plus
the building, then whatever your model and automation providers charge, paid to them.
The agent wins when the work is high volume, rule-shaped and already written down. The person
wins when volume is low, judgment is constant, or nobody has ever written the process down.
Most jobs are a mix, which is why the useful question is never "person or agent" but
"which hours of this job".
What is an AI employee?
It is a marketing term, not a category of software. What is being sold is an
AI agent scoped to one job: software given a job rather than a
list of instructions, which reads what actually arrived, works out the next step, operates the
business software you already run, and escalates what it should not decide alone.
Nobody is employed. There is no contract of employment, no notice period, no employment law
and no unfair dismissal. That cuts both ways, and it is worth being clear about which way in each
direction:
| A person | An agent |
|---|
| Cost shape | Salary plus roughly 43% on top | One build, then what the model and platform charge to run it |
| Ramp | Weeks to months | Days to weeks, then it is either right or it is not |
| Hours | About 2,000 a year, minus leave and sickness | Every hour, including the ones nobody wants |
| Unusual cases | Handles them, often without being asked | Should escalate them, and must be built to |
| Getting it wrong | Notices, and tells you | Repeats the mistake at volume until somebody looks |
| Leaving | Takes the knowledge with them | The configuration is the knowledge, and it stays |
| Scaling up | Another salary | Mostly more model usage, not another salary |
What a person actually costs
Not their salary. The Bureau of Labor Statistics measures this directly, and for March 2026 in
private industry it puts wages and salaries at 69.9% of total employer cost, with
benefits making up the other 30.1%. Divide by 0.699 and you get the real number.
| Salary | True employer cost | Per working hour |
|---|
| $38,000 | $54,400 | $27 |
| $45,000 | $64,400 | $32 |
| $55,000 | $78,700 | $39 |
| $70,000 | $100,100 | $50 |
Hourly figures assume about 2,000 paid hours a year. They are still understated, because they
exclude recruitment, the manager's time, the desk, the software seat and the months of reduced
output while somebody learns the job.
Do the sum for your own case
The number that decides this is not the salary. It is the share of the job an agent
genuinely takes. Every vendor calculator quietly assumes 100%. None of them hit it.
Sixty to eighty per cent is a realistic band for well-defined work, and the remainder does not
disappear: it becomes exceptions somebody still has to handle.
Nothing is sent anywhere and nothing is stored. The arithmetic is in the page, and
you can read it by viewing source.
The statistic everybody misquotes
You will read, on page after page, that "MIT found only 23% of jobs are economically
viable to automate". It is repeated by vendors on both sides of the argument, and it is not
what the study says.
What the 2024 MIT study actually found. It examined tasks that could be
done with computer vision, and concluded that about 23% of the
wages paid for those vision tasks would be economically viable to automate with
computer vision at then-current costs. It was not a finding about all jobs, and it was not
about language-model agents, which barely feature in it.
The transferable point is the one worth keeping, and it is genuinely important:
technically possible and economically sensible are different questions, and
the gap between them is mostly the cost of building and running the thing. That is exactly
the gap this page is asking you to measure.
We are pointing this out against our own interest in one direction and for it in another,
which is roughly the definition of a fact worth citing. If you quote the figure, quote it as what
it is: vision tasks, 2024 costs, wages rather than roles.
When the answer is a person
These are the cases where we tell people not to buy, and we would rather say it here than on a
call in six weeks:
- Low volume. If the task happens a few times a week, the build will
never pay for itself. Ten minutes of somebody's day is not a business case.
- Judgment every time. If no two cases are alike, there is no rule to
encode and you are buying a very expensive way to generate exceptions.
- The process is not written down. Software cannot enforce a process
nobody has articulated. Write it down first. That exercise alone often fixes the problem, and
it costs nothing.
- Nobody follows the process today. A system that enforces an ignored
process does not create compliance, it creates resentment and a workaround.
- The documents contradict each other. Software makes disagreement
permanent rather than resolving it. Settle it first.
When the answer is an agent
- Volume with a shape. The same kind of thing arriving many times a week,
where the right response follows from what arrived.
- Hours nobody wants. Evenings, weekends and the first hour of the morning,
where the cost of a slow reply is a lost job rather than an inconvenience.
- Work that is already documented. If you have an SOP, a template set or a
script, most of the build is done and you are paying somebody to encode it.
- A queue that is quietly costing you. Quotes that were never chased,
invoices nobody followed up, enquiries answered two days late.
What to check before you buy one, from anybody
Six questions. Any vendor who cannot answer them plainly is selling you a demo:
- What exactly does it own, and what does it escalate? A defined job
description, not "handles your email".
- What is it allowed to do without asking? Read, draft, or send. Those are
three very different purchases. Every tool in our kits is marked with which of them it is, and
the levels are covered in permissions.
- How do I see what it did, and what it refused to do? If the answer is a
log file, that is not an answer. See how one is kept
under control.
- How do I stop it? There should be an off switch a non-technical person
can reach in seconds.
- What is the acceptance test? The specific, checkable thing that proves it
works before you pay for it working.
- What is the total first-year cost? Build, monthly, and anything metered.
A kit's price is printed and shown on
the kit page before you decide, and the running costs are your providers'
own prices, paid to them directly.
How we price it, since we are asking you to compare
It would be poor form to publish a cost comparison and hide our own half of it. We sell the
agent's complete specification as a kit: $79 for an agent that owns one
job, $199 for one that owns a whole function. One-off, no subscription.
You, or whoever builds for you, build it with the tools you already use, and you pay the model and
platform providers directly, with nothing added by us.
There is no discovery call and no proposal. You describe the job, the kit and its price are on
the page, you pay, and you download it. If you would rather see the diagnosis first,
Business Scan reads your public website and works out which jobs are costing you
most, and Agent Designer designs the agent for a business you describe. Both are free
and neither asks for an account.
Sources
Prices on this page are BlvkWare's published prices at the time of writing.
Salary figures are illustrative. This is a commercial comparison written by a vendor, which is
why both sides are sourced and the calculator is allowed to reach a conclusion we do not profit
from.