# How Much Does an AI Employee Cost Versus a VA?

*Comparison — 2026-09-02 — by Mahmoud Zalt*

Published VA rates come from staffing agencies. Here is the honest cost picture, and the software-to-software comparison that actually decides your bill.

**TL;DR.** Almost every published virtual assistant rate comes from a staffing agency that sells virtual assistants, so read it as advertising rather than market truth. We do not price our product against a person's pay, and this article does not either. The comparison worth your time is software against software, where five different billing models are live: per seat, per conversation, per resolution, per session, and a platform fee plus usage. Once seat fees and annual minimums amortise, the sticker unit price on the pricing page and the true unit price on your invoice stop matching.

## What the published virtual assistant rates say

The rates you will find published come in two shapes, an hourly band by region and a monthly retainer band. Hourly, the published figures run $30 to $75 in the US and Western Europe, $10 to $30 in Eastern Europe, $8 to $25 in Latin America, $4 to $12 in the Philippines, and $3 to $10 in South Asia.

Monthly full-time figures follow the same map. US and Canada start around $4,500, Poland runs $1,800 to $3,500, Mexico and Romania both $1,500 to $3,000, Colombia $1,200 to $2,500, Argentina $1,100 to $2,000, the Philippines $700 to $2,000, and India $500 to $1,800. Latin American rates are also published by seniority, from $800 to $1,200 at entry level up to $2,200 to $3,500 and beyond for specialised roles.

| Region | Published hourly | Published monthly, full-time |
|---|---|---|
| US and Western Europe | $30 to $75 | $4,500 and up (US and Canada) |
| Eastern Europe | $10 to $30 | $1,800 to $3,500 (Poland) |
| Latin America | $8 to $25 | $1,100 to $3,000 (Argentina to Mexico) |
| Philippines | $4 to $12 | $700 to $2,000 |
| South Asia | $3 to $10 | $500 to $1,800 (India) |

Now the part most articles skip. Nearly every source publishing these bands is a staffing agency whose business is placing virtual assistants. Those pages are the top of a sales funnel, not neutral research, and the incentive runs toward whichever end of the range closes the deal.

That does not make the numbers useless. It makes them a starting range rather than a finished answer. If you are budgeting, treat them the way you would treat any list price: real enough to plan around, soft enough that the actual quote will differ once scope, hours, and seniority are pinned down.

## Why we do not put our price next to a salary

Because it is a rigged comparison, and readers can smell it. A person you hire brings judgment, accountability, and a working relationship with your customers. A subscription brings a defined set of runs. Putting the two in one table makes the software look like a bargain by quietly hiding everything the software does not do.

There is a practical problem too. A salary is a fixed monthly commitment for a fixed set of hours. Software pricing is usually metered against a unit of work, so the two numbers behave completely differently as your volume moves. A comparison that looks decisive at low volume can invert entirely at high volume, which means the table was never telling you anything durable.

So compare software to software. If you are evaluating an AI employee, put it next to other AI platforms on the axis that actually decides your bill: what unit you are charged for, and what that unit really costs once every fee in the contract is counted. That includes [Sistava](/features), which belongs in the same test as every platform named below rather than beside it.

## The five pricing models you will meet

The AI agent market has settled into five billing shapes: per seat, per conversation, per resolution, per session, and a platform fee plus usage. Vendors mix them, which is exactly why comparing headline numbers across pricing pages tells you so little.

| Platform | Published unit price | What else sits underneath |
|---|---|---|
| Fin | $0.99 per outcome | Charged on resolved outcomes |
| Zendesk AI | $1.50 per resolution committed, $2.00 pay-as-you-go | Suite $55 to $169 per agent per month, plus a $50 per agent per month Advanced AI add-on |
| Salesforce Agentforce | $2.00 per conversation | Service Cloud Enterprise from $175 per user per month |
| Ada | About $1.00 to $3.50 per interaction | Roughly $30,000 per year minimum |
| Gorgias | About $0.60 to $1.27 per resolution | Tiered by plan |
| Freshdesk Freddy | $0.10 per session | Session-based metering |
| Decagon | About $0.50 per resolution | $50,000 per year platform fee |

Read down that middle column and the spread is roughly twenty to one, from ten cents a session to two dollars a conversation. Read the right column and the picture changes again, because several of the cheapest unit prices sit on top of the largest fixed commitments.

## Sticker unit price versus true unit price

Here is the structural point, and it is the only piece of cost analysis in this article that will still be true in two years. A metered price is only the price at the volume where the fixed fees disappear into it. Below that volume you are paying the fixed fee and getting a unit rate as a rounding error.

Take the $50,000 annual platform fee sitting under a $0.50 resolution. At 100,000 resolutions a year the fee adds fifty cents, so your true unit price is a dollar. At 10,000 resolutions a year the same fee adds five dollars, so your true unit price is $5.50, eleven times the sticker. Nothing dishonest happened. The pricing page was accurate and your invoice was accurate and they described different worlds.

The same arithmetic runs through per-seat models. A per-agent monthly fee plus an AI add-on per agent is a fixed floor you pay whether the AI handles four things this month or four hundred. Divide your floor by your realistic volume before you compare anything, and do it at the volume you have now, not the one in your plan.

None of this makes any particular platform wrong. It makes the pricing page an incomplete document. The way through is to stop reading unit prices as if they were comparable and start building one small model of your own: your volume, your fixed floor, your realistic unit count per month. That model takes twenty minutes and it will survive every pricing change the market throws at you, because it is about your business rather than about anyone's marketing.

## What no pricing page covers

Look at the published task list for a virtual assistant and most of it is judgment work wearing an administrative label. Inbox management is really email prioritisation. Calendar scheduling is really travel coordination when the flight is cancelled at eleven at night. CRM updates sit next to customer follow-ups, meeting preparation, operations coordination, and writing an SOP for how this particular company actually works.

There is a technical reason automation struggles with exactly that half of the list. In the OSWorld 2.0 benchmark of 108 authentic professional tasks, one of the three named failure patterns is that agents struggle to maintain hidden state across steps, repeating work or executing from a plan that went stale three steps ago. Holding the thread of a week-long piece of work and noticing when the premise changed is the thing a good assistant does effortlessly and the thing software is measurably worst at.

Two other categories never appear on any AI pricing page at all: anything physical, and anything that needs a legally accountable person to sign it. Neither of those is a gap you close with a better model. They are the shape of the boundary, and knowing where it sits is worth more than any per-unit comparison.

## Work out your own number in five steps

### A twenty-minute cost model

1. **Name the unit you actually buy** — Write down the single thing you want more of: resolved support emails, qualified leads, published posts, reconciled invoices. Everything downstream is priced against this.
2. **Count last month, not next month** — Use the volume you genuinely had, from your inbox, your helpdesk, or your CRM. Planning volume flatters metered pricing and punishes you when the invoice arrives.
3. **Add every fixed floor** — Seat fees, platform minimums, add-on modules, annual commitments. Sum them into one monthly number before you look at any unit rate.
4. **Divide the floor by your real volume** — That figure is what each unit costs you before you pay for a single unit. Add the sticker unit price to it. Now you have a true unit price you can compare.
5. **Subtract the work that is not on the table** — Prioritising, chasing a real person, re-coordinating when plans change, anything physical, anything needing an accountable signature. Those stay with a human, so do not credit them to either column.

Run that model and something useful usually happens: the decision stops being about price at all. You end up with a clear picture of which work is repetitive and well-specified enough to hand to software, and which work was always going to need a person, and the budget question answers itself from there. Once that first job is named, hiring a single AI Employee for it on [Sistava](/hire-ai-employees) is a cheap way to test your arithmetic against a real month of work.

One more habit worth building: re-run the model every quarter. Volume moves, pricing structures change, and a platform that was the obvious choice at 300 units a month is often the wrong choice at 3,000. The model takes minutes to update once it exists.

When you do look at the [Sistava plans](/pricing), look at them the same sceptical way you looked at everything above. Ask what unit you are charged for, what the fixed floor is, and what happens to your bill if your volume triples. Those three answers tell you more than any comparison table anyone can build for you.

## FAQ

### Are published virtual assistant hourly rates accurate?

They are accurate as advertised ranges and unreliable as market truth. Nearly every published rate table comes from a staffing agency whose business is placing virtual assistants, so the figures sit at the top of a sales page rather than in an independent survey. Use them to set a rough budget band, then get a real quote for your actual scope, hours, and seniority before you plan around any specific number.

### How is AI agent pricing structured?

Five models are live in the market: per seat, per conversation, per resolution, per session, and a platform fee plus usage. Published unit prices range from about ten cents a session at the low end to around two dollars a conversation at the high end. Many vendors combine models, charging both a per-seat subscription and a per-unit rate, which is why headline numbers across pricing pages are rarely comparable.

### Why does my AI platform bill not match the pricing page?

Usually because a fixed fee is being spread across fewer units than the pricing was designed for. A fifty-cent resolution sitting under a $50,000 annual platform fee costs a dollar at 100,000 resolutions a year and $5.50 at 10,000. Divide every seat fee, minimum, and add-on by your real monthly volume, add the unit rate, and you get the number that will appear on your invoice.

### Should I compare an AI employee to a virtual assistant's salary?

No, and not because either option loses. A salary is a fixed commitment for a fixed set of hours, while software is metered against units of work, so the two behave differently as volume moves and a comparison that looks decisive at one volume inverts at another. Compare software to software on true unit price, and decide separately which work genuinely needs a person.

### What work can an AI employee not take over?

Anything physical, anything needing a legally accountable person to sign it, and the judgment half of assistant work: prioritising a messy inbox, following up with a real human being, re-coordinating travel when plans collapse, and writing an SOP for how your specific company works. Benchmarks also show agents struggle to hold hidden state across a long task, which is precisely what that judgment work requires.

### What is the cheapest way to start with an AI employee?

Pick one repetitive, well-specified job with a countable output, hire an employee for that job alone, and connect only the tools that job needs. A narrow scope keeps your unit volume predictable, which keeps your true unit price close to the sticker price. Expand once you have a month of real numbers rather than budgeting from a plan. Current plan details are on the pricing page.

Cost questions feel like they should have one number as an answer. They almost never do, because the number depends on your volume, your fixed floor, and how much of the job is repetitive enough to specify in writing. Everything else is someone else's marketing arithmetic borrowed and applied to a business it was never modelled on.

Build the small model, run it on last month's real numbers, and keep the two columns honest: the work that is well-specified enough to hand to software, and the work that needs a person's judgment. Once those columns are clear, the budget stops being a debate and turns into arithmetic, which is a much better place to make a decision from.

**Tags:** ai-employee-cost, virtual-assistant, pricing-models, buying-guide, cost-comparison