The short version
- The advertised price is real but partial. Live vendor pages in July 2026 quote $25–100/month on credit plans, $399/month if you do the setup and $999/month for done-for-you, benchmarked against a fully loaded human at roughly $4,375–4,575/month.
- Four costs sit outside that number: the connection work before it does anything useful, the person internally who owns it, the correction load in month one, and the cost of it being confidently wrong in front of a customer.
- I give the first three in hours, not dollars, because your internal hourly cost is the one number I cannot know.
- Budget roughly double the sticker for year one. That still beats the salary comparison. Budgeting the sticker alone is what produces the month-one surprise.
- The only figure that matters is cost per finished task, not cost per month. Rent one for two weeks and measure it before anyone quotes you a build.
Search “how much does an AI employee cost” and page one is almost entirely vendors and agencies publishing their own price lists. I checked it on 26 July 2026. Each result puts a monthly number next to a human salary and declares the argument over. The one piece of independent press on the page answers a different question — what large firms spend on AI per human employee — which is no use to anyone deciding whether to run one.
I run an AI employee in my own business and I have deployed them for clients. The monthly subscription is genuinely the cheapest part. Everything below is the rest of the sheet.
What the vendor pages actually quote
Two representative examples, both read off their live pages on 26 July 2026:
| Vendor page | Advertised price | Human benchmark used | Buyer-side cost on the page |
|---|---|---|---|
| teammates.ai | $0 / $25 / $50 / $100 per month, credit-based | $55K–70K/yr loaded for support, $100K–130K for an SDR | Stated as zero: “no onboarding fee, no implementation charge, no training cost” |
| aiemployee.com | $399/month self-setup, $999/month done-for-you | US receptionist, $3,200 salary, $4,375–4,575 fully loaded per month | 4–6 hours of your setup time on the $399 tier, plus $25–100 to port a phone number |
Credit where it is due: aiemployee.com is the only page I found that puts any buyer-side hours on the page at all. Everyone else prices as though the software arrives already knowing your business.
The pattern across the whole first page is the same. The seller sets the numerator, picks the comparison, and never touches the denominator. That is not dishonesty, it is just what a price list is. It is only a problem when a buyer budgets from it.
The four costs that sit outside the subscription
These are your hours, not the vendor’s. Multiply them by whatever an hour of your team’s time actually costs you.
1. The connection work, before it does anything useful
An AI employee is worth exactly as much as the systems it can reach. Out of the box it can hold a conversation. To do work it needs read and write access to wherever the work actually lives: the shared inbox, the CRM, the invoicing system, the calendar, the stock or job system. Every one of those is credentials, a permission scope, a test, and a failure mode for when the token expires.
Realistic buyer-side hours for a first deployment, from the ones I have set up:
- Deciding what it should actually do, in writing: 2–4 hours. The step everyone skips. “Handle customer enquiries” is not a specification, and an assistant built from it will be wrong in a way nobody can correct.
- Access and credentials: about 30 minutes per modern system (Google Workspace, Slack, HubSpot, Stripe — anything with a proper sign-in flow). Half a day or more per awkward one: an on-premise ERP, a system whose vendor charges for API access, or anything where IT owns the keys and has a change process.
- Supplying the knowledge: 3–10 hours. Your prices, your policies, your exceptions, the answers you actually give rather than the ones on the website. This is the whole difference between useful and embarrassing, and no vendor can do it for you because it is not written down anywhere yet.
- Testing against real cases: 2–5 hours. Feed it last month’s genuine messages, not invented ones.
Call it 8–20 hours of your side for a single-channel assistant with two or three connections, and materially more if any system in the list has no usable API. A done-for-you tier compresses your hours, it does not remove them. Nobody else can tell an assistant what your business does when a customer asks for a discount.
2. The person inside your business who owns it
This line appears on no vendor page and it is the one that decides whether the thing survives to month three. An AI employee with no owner degrades. Prices change. A policy changes. A token expires. Somebody renames a field in the CRM and one integration starts silently returning nothing.
Budget 4–8 hours in month one and 1–3 hours a week after that once it has settled. Not because it breaks constantly, but because the value comes from keeping its knowledge current, and that is a habit, not a project.
The test to run before you sign anything: say the name of the person who owns it out loud. If there is no name, the honest budget line is not zero. It is the cost of it going quietly wrong while everyone assumes someone else is watching.
3. The correction load in month one
It will get things wrong at the start, and it will do it in a very particular way: confidently, in your house style, about the cases your written knowledge did not cover. It does not hedge. That is what makes the first month a real cost rather than a rounding error.
The shape is predictable. Corrections are heavy in week one, fall off sharply through weeks two and three, and are close to background noise by week five or six — provided somebody is writing the corrections back into its instructions as they happen. If nobody is, the correction load does not decay. It stays flat forever, and about six weeks later somebody quietly stops using it. That is the single most common way these deployments die, and it has nothing to do with the model.
The cheapest control I know: run it in draft mode for the first two weeks. It writes the reply, a human reads and sends. You get the same learning curve, and you convert an unknown risk into a known number of minutes per day.
4. The cost of it being confidently wrong in front of a customer
This one is not a monthly line. It is a tail risk, and you should price it by blast radius rather than by probability, because the probability is not zero and you cannot compute it in advance anyway.
| Permission level | Example work | Worst realistic outcome |
|---|---|---|
| Read-only, internal | Daily brief, research, summarising a shared inbox | Wasted time and a wrong impression. Recoverable. |
| Customer-facing, non-committal | Answering questions, taking details, booking a callback | A bad interaction and an apology. Costs goodwill, not money. |
| Customer-facing and committing | Quoting prices, confirming delivery dates, approving refunds, booking slots you must honour | A number. Work out that number before go-live, not after. |
The mitigation is not a better model. It is a smaller permission set. Let it draft anything at all; let it send or commit only in the lane where a mistake is survivable. Most of the deployments I have seen fail on cost did not fail because the AI was bad. They failed because someone handed it the committing lane on day one to prove a point.
Estimate your custom AI project in 30 seconds
Three questions, an instant cost range and timeline based on real shipped projects. After 30 minutes on a discovery call you have a written fixed-price quote.
Or build your own AI system piece by piece and send the design in for a written quote →
The number that actually matters: cost per finished task
Cost per month is the wrong unit, because it is the only unit where the seller controls both sides. Use this instead:
Cost per finished task
(subscription + model usage + hosting + your hours × your loaded hourly rate) ÷ tasks it finished without a human touching them
The denominator is the part no vendor publishes, because it depends entirely on you. A $399/month assistant that finishes 400 things a month costs about a dollar a thing. The same $399 assistant that finishes 40 things and needs a human to check the other 360 is not cheaper than a person. It is a person plus $399.
Two rules that keep the number honest. Count only what was finished without a human touching it, because “the AI drafted it and Sarah fixed it” is Sarah’s task. And measure over at least two weeks, because week one is training and will make anything look terrible.
A worked year, with every input visible
Assumptions you should replace with your own: a $399/month managed assistant, one channel, three connections, and an internal loaded rate of $50/hour. I am showing the arithmetic rather than a conclusion so you can re-run it with your real numbers.
| Line | Month 1 | Steady month | Year 1 |
|---|---|---|---|
| Subscription | $399 | $399 | $4,788 |
| Setup and connection, 14 hrs | $700 | — | $700 |
| Month-one correction load, 6 hrs | $300 | — | $300 |
| Internal owner, 2 hrs/week | — | $400 | $4,400 |
| Total | $1,399 | $799 | $10,188 |
The advertised figure for that year is $4,788. The real one is about $10,200. Roughly double.
Now the part that is against my own interest to point out: doubling the sticker price does not change the vendors’ conclusion. Ten thousand dollars a year is still a long way under the $52K–55K fully loaded human that aiemployee.com benchmarks against. The vendors are not wrong about the direction. They are wrong about the arithmetic, and a buyer who budgets the sticker gets a surprise in month one and blames the AI for it.
Where the comparison genuinely stops working is on the job description rather than the money. An AI employee does not exercise judgement, does not negotiate, and cannot be held accountable. Compare it against a defined slice of work, never against a person.
When a custom build is the right call
Renting is the right start for nearly everyone, including most of the people who email me about custom AI worker development. Commission a build when one of these is true and you can show it:
- The work lives somewhere nothing off-the-shelf can reach. An old ERP, an in-house scheduling system, a database with no API, an industry tool whose vendor charges four figures for integration access. This is by far the most common honest reason.
- It needs the committing lane. Money, contracts, regulated advice. When mistakes cost real money you want the permission model, the approval steps and the audit trail to be yours, not a setting in someone’s dashboard.
- You have measured a ceiling. The rented one finishes its lane well, and you can point at the specific system it cannot reach that would double its coverage.
- Compliance rules out multi-tenant software. Health, finance, legal. If the data cannot leave your infrastructure, customised AI employees on your own server stop being the premium option and become the only compliant one.
If that is your situation, the shape of the work is on my custom AI assistant and agent development page, and the build-side numbers are broken down in the custom AI agent development cost guide. What a deployment actually involves, step by step, is in how I set up AI employees.
What you should not do is commission a build in order to find out. A build priced before you have measured anything is scoped on guesses, and guesses are expensive at custom prices.
The cheapest way to find your own number
No article can price your case, and that includes this one. Every band above is a starting point you are supposed to replace. The denominator — how many things it actually finishes in your business, with your systems and your customers — is unknowable from the outside, and two weeks of running one will tell you more than any amount of reading.
- Pick one lane. Not “everything”. One repeating job with a countable unit: enquiries answered, invoices chased, briefs produced.
- Rent, do not build. Any provider with a real trial and no setup fee will do. Mine is one of them, so discount it accordingly: Nora lists a two-week free trial and unlimited AI employees from $59 a month, which is a useful number to hold next to the $399–999 done-for-you figures above before you accept those as the market rate.
- Draft mode week one, live week two. You will learn the failure modes at zero blast radius.
- Keep a tally. Three columns: finished without a human, needed a fix, had to be redone. Log your own minutes in a fourth.
- Do the division on day 14. That is your cost per finished task, and it is the number every quote you receive afterwards has to beat.
If the number is good, keep renting. Most businesses should. If it is good but capped by a system the rented one cannot reach, you now have a business case with a measured denominator in it, and you can get a build quoted against a real number rather than a hope. Either way you have spent two weeks and no capital finding out, which is the correct order.
Frequently asked questions
How much does an AI employee cost per month in 2026?
Advertised subscriptions run from about $25 a month on credit-based plans, up to $399 a month if you do the setup yourself and $999 a month for done-for-you, based on live vendor pricing pages in July 2026. Those numbers are real but partial. Once you add your own setup hours, an internal owner, and the month-one correction load, the first-year total typically lands at roughly double the advertised subscription.
Can you have AI employees in a small business?
Yes. An AI employee is a software agent that works a channel like WhatsApp, web chat or email, connects to your existing tools, and runs scheduled work without being prompted. Small businesses run them successfully on defined, repeating work: answering product and availability questions, qualifying inbound leads, chasing invoices, producing a daily operations brief. They are not a substitute for judgement, negotiation or anything where being confidently wrong is expensive.
What do AI employee vendors leave out of their pricing?
Four things. The connection work before it does anything useful, which is your hours not theirs. The person inside your business who owns it. The correction load in the first month while it learns the cases your written knowledge did not cover. And the cost of it being confidently wrong in front of a customer, which is a tail risk you should price by blast radius before go-live.
Is an AI employee cheaper than hiring a person?
For defined, repeating work, usually yes, and by enough that doubling the advertised subscription does not change the answer. Vendor pages benchmark against a fully loaded US receptionist at roughly $4,375-4,575 a month. The comparison stops working the moment the job needs judgement, negotiation or accountability, because an AI employee does not carry any of those.
Should I build a custom AI employee or rent one?
Rent first. A build priced before you have measured anything is priced on guesses. Run a managed assistant on one lane for two weeks, count what it finished without a human touching it, and you will have the one number every quote afterwards has to beat. Commission a build when the rented one works but cannot reach the system that would double its coverage, or when compliance rules out multi-tenant software.
Sources
- Vendor pricing and human benchmarks, live pages checked 26 July 2026: teammates.ai AI employee cost and aiemployee.com AI employee cost.
- Trial and subscription terms quoted for Nora, public site at noraclawd.com, 26 July 2026.
- Hour bands, correction-load shape and permission-lane framework: my own deployments and the assistant I run in this business. Stated as operator experience, not survey data.