A8gent

Calculator

AI Agent vs Virtual Assistant Cost Calculator

Cost comparison for an AI agent against a VA on the same task. Published VA rates by region, real token prices, and the review time most comparisons quietly leave out.

Contents
  1. Fast answer
  2. Quick facts
  3. What this covers
  4. Why it matters
  5. Where to go next
  6. Outcomes
  7. Step by step
  8. The methodology: two cost curves, four inputs each
  9. The review line is the whole argument
  10. A worked example: portal data entry at 800 items a month
  11. The limits: five costs this page names and does not price
  12. When this comparison is the wrong question, and where to go instead
  13. At a glance
  14. Mistakes to avoid
  15. FAQ
  16. Related guides
TL;DRAnswer-first

Fast answer

An AI agent and a virtual assistant are both priced per unit of work once you do the arithmetic properly, and the comparison turns on three numbers: how long the task takes a human, how many tokens it takes a model, and what share of the output somebody still has to check. The calculator on this page takes all three plus your volume and returns monthly cost, cost per item, the break-even volume, and the payback period on whatever you spend building the agent. VA rates come from published 2026 rate reporting by region rather than from anything we made up. The result that surprises most people is that the review line, not the token line, decides the outcome. At a twenty percent review rate and two minutes per review, a reviewer on a sixty dollar hour costs forty cents per item, which is often more than the tokens. An agent nobody checks looks dramatically cheaper and is simply unaudited.

Facts last verified 2026-08

Quick facts

What decides it
Review share, more often than token price
Philippines VA range
$4 to $12/hr, wage only, published 2026
Filipino VA median ask
$6.00/hr, from 1,385 applications, Oct 2025 to Jul 2026
US direct hire range
$28 to $48/hr fully loaded, published 2026
Agent cost shape
Fixed platform and build, plus per-item tokens and review
Rates checked
2026-08-20

On this page

What this page covers

A tool visitor should leave with a decision, not just a number: build now, prepare first, choose another workflow, or follow a course path.

  1. 01The task
  2. 02VA rates by region
  3. 03Agent token cost
  4. 04Review burden
  5. 05Break-even
  6. 06Sources

Why does this matter now?

The AI agent versus VA comparison is usually argued with a token price on one side and an hourly rate on the other, which is not a comparison at all. A VA absorbs judgement, exceptions, escalation and continuity without a separate line item. An agent externalises all four back to you, and if you do not price them the agent wins on paper and loses in the month. Getting this wrong in the direction of the agent is expensive in a way that shows up late: work that quietly needs rework, a portal that changed its markup, a client who noticed before you did. Getting it wrong in the direction of the VA is expensive in a way that shows up immediately on an invoice, which is why that error gets corrected and the other one does not.

Internal path

Where to go next from this page

These links are part of the A8gent learning and conversion path. Use them to move from concept, to diagnosis, to workflow build, to course.

Start with readiness

What you should be able to do after this

  • Monthly cost each way
  • Cost per item handled
  • Break-even volume
  • Payback on the build cost

How do you do it, step by step?

1. Define one task, not a role

Portal data entry is a task. Marketing support is a role. The comparison only works at task level, because an agent handles a bounded, repeated, well-specified job and a person handles the twelve other things nobody wrote down. Pick the single highest-volume repeated item and count how many of them happen in a month.

2. Time the human path end to end

Not the keystrokes, the whole thing: opening the record, finding the input, doing the work, checking it, and handling the case where the input was wrong. Timing the happy path alone is the most common way this calculation gets skewed toward the agent, because the exceptions are precisely what the agent cannot do.

3. Pick the VA rate from published ranges, then adjust

The calculator carries 2026 published rate ranges for the Philippines, Caribbean nearshore, US direct hire and US premium managed firms, and it flags which of those are wage-only and which are all-in. These are different products with similar-looking hourly numbers, and comparing a wage-only figure against an all-in figure is the second most common error here.

4. Estimate tokens per item from a real sample

Run five real items through the model and read the usage numbers off the response rather than guessing. Input tokens are usually dominated by whatever context you attach; output tokens by how verbose you let the model be. Both are controllable, and halving either one moves the per-item cost immediately.

5. Set the review share to what you would actually sign off on

This is the number that decides the answer. If the output touches a customer, money, a record of account or anything regulated, the honest review share is high and the minutes per review are not trivial. Set it to what you would defend to whoever owns the risk, not to what makes the agent look good.

6. Read the break-even volume, not just the monthly total

The agent carries fixed cost in platform fees and build amortisation; the VA carries fixed cost in your management time. Everything else is per item. The calculator solves for the volume at which the two lines cross, and if the agent is more expensive on every single item then volume will never fix it and the model choice or the review design needs to change first.

The methodology: two cost curves, four inputs each

The VA column is production hours multiplied by an hourly rate, plus your management time at your own rate. Production hours are your item count multiplied by the minutes a human spends on each one, divided by sixty. Nothing else. The regional buttons set a starting rate from published 2026 rate reporting and then get out of the way, because the rate you can actually hire at is a fact about your market and not about our page.

The agent column has four terms. Tokens are your item count multiplied by input tokens per item, converted to millions and multiplied by the model's published input price, plus the same for output. Platform is a flat monthly figure you set, anchored to real published entry plans. Review is the share of items a human checks, multiplied by minutes per review, at your rate. Build cost is a one-off divided by the months you choose to spread it over.

The break-even is solved algebraically rather than by scanning volumes. Each column has a marginal cost per item and a fixed cost per month. The VA's marginal cost is minutes over sixty times the rate. The agent's is per-item token cost plus the expected review cost per item, which is the review share multiplied by review minutes at your rate. The crossover volume is the difference in fixed costs divided by the difference in marginal costs, and when that difference is negative the page says so plainly: the agent loses on every item and volume will never rescue it.

Payback is computed against steady-state running cost rather than against the amortised figure, because that is the question people actually mean. It is the build cost divided by the monthly gap between the VA bill and the agent's ongoing costs, excluding the amortisation. If that gap is not positive there is no payback at any horizon, and the page says that instead of printing an implausibly large number.

The review line is the whole argument

Almost every published AI-agent-versus-human comparison omits human review, and almost every one of them therefore reaches the wrong conclusion. Work the arithmetic yourself: at a twenty percent review share, two minutes per review and a sixty dollar hourly rate, review costs forty cents per item. On a task where the model consumes 3,000 input and 400 output tokens on a small model, the tokens cost under a tenth of a cent. Review is not a rounding error on the token line; the token line is a rounding error on review.

This has a design consequence that matters more than any vendor choice. If review dominates, the lever that improves your economics is not a cheaper model, it is a workflow that needs less checking: tighter output formats, a confidence signal the agent emits so you only review the uncertain cases, a validation step that catches the mechanical errors before a person sees them. Sampling instead of checking everything is legitimate for low-stakes work and negligent for high-stakes work, and the calculator lets you model both honestly rather than pretending the choice does not exist.

Setting the review share to zero is available and we would rather you understood what you are claiming when you do it. Zero review means nobody looks at any output before it lands in a record, an inbox or a customer's hands. For classification into a low-stakes bucket that may be entirely reasonable. For anything touching money, personal data, a regulated record or an external party it is not a cost saving, it is a transfer of risk to whoever discovers the error later.

The corollary is that the honest review share usually falls over time rather than starting low. A new workflow needs high review because nobody has evidence yet; a workflow with three months of logged edit rates has earned a lower one. Modelling month one and month six as two separate runs of this calculator is more useful than arguing about a single number, and it gives you a target to manage toward rather than an assumption to defend.

A worked example: portal data entry at 800 items a month

Take the portal data entry preset: 800 items a month, six minutes each for a human, 3,000 input and 400 output tokens for the model, fifteen percent reviewed at two minutes each. The VA side is 800 times six over sixty, which is eighty production hours. At the median published Filipino rate of six dollars an hour that is $480, plus four hours of your management time at sixty dollars, so $720 a month all in, or ninety cents an item.

The agent side on GPT-5 mini: 800 times 3,000 is 2.4M input tokens, at $0.25 per million that is sixty cents. 800 times 400 is 0.32M output tokens, at $2.00 per million that is sixty-four cents. Total tokens: $1.24. One dollar and twenty-four cents for the entire month of model calls. Review is 120 items at two minutes, which is four hours, at sixty dollars that is $240. Platform at twenty-nine dollars. Build of $2,500 spread over twelve months is about $208. Total: roughly $478 a month.

Look at what that breakdown says. The tokens are a quarter of one percent of the agent's cost. Review is half of it. The build amortisation is another 43 percent. The thing everyone argues about, the model price, is the only line that does not matter, and the two lines nobody discusses are the entire bill. This is why the comparison is worth calculating rather than debating.

Now change one input. Push review to forty percent because the portal feeds a billing system and an error is expensive: review becomes 320 items at two minutes, or ten and two thirds hours, which is $640, and the agent column moves to roughly $878 against the VA's $720. The agent has lost, on the same task, the same volume and the same model, purely because somebody insisted on checking the work. That is not a flaw in the calculation. That is the calculation doing its job.

The limits: five costs this page names and does not price

Judgement. A VA notices that the source form has a field in the wrong place, that the customer name does not match the account, that this one is obviously a duplicate. An agent processes all three confidently. The value of that noticing is real and varies from trivial to enormous depending on what sits downstream, and any number we assigned to it would be invention.

Escalation ownership. Somebody has to own the cases that neither party can resolve, and that somebody has a rate. In the VA arrangement this is frequently absorbed silently by the assistant. In the agent arrangement it lands back on you, and it lands as interruptions rather than as scheduled work, which costs more per hour than the rate suggests.

Continuity in both directions. A VA takes holiday, gets ill, and eventually leaves, and replacing them costs recruiting time and ramp. An agent does none of that and instead breaks silently when a portal changes its markup or an API deprecates a field. The failure modes are not comparable in shape, only in cost, and which one is worse depends on how quickly you would notice.

Ramp and rework. A new VA is slow for weeks. A new agent is wrong until the prompt, the tools and the test cases are right, and the calculator's build cost input is the only place that appears. If your build estimate does not include a fortnight of tuning against real examples, it is not a build estimate, it is a demo estimate, and the payback figure inherits the error.

When this comparison is the wrong question, and where to go instead

When the task is not actually repeated. Below a hundred or so items a month, almost nothing survives the build cost, and the calculator will show a payback period long enough to make the point. The right answer at that volume is usually neither an agent nor a VA but a better template and ten minutes of your own time.

When the role is a role rather than a task. If you are comparing a marketing VA against an agent, split the role first: scheduling posts, formatting newsletters, pulling report numbers and drafting first-pass copy have four different token profiles and four different review requirements. Running this calculator once per task produces a hybrid answer, and the hybrid answer is nearly always the correct one. The workflow planner is the faster route to that split than arguing about it.

When the risk has not been scored. Anything that sends external messages, moves money or touches personal data needs its control set decided before its cost is compared, because the controls change the review share and the review share changes the answer. Score it first, then come back and put the resulting review percentage into this page rather than picking one that produces a pleasing verdict.

And when the honest blocker is that nobody has written the process down. An agent cannot be specified against a task that only exists in somebody's head, and a VA will reconstruct it slowly and expensively. Documenting the process is cheaper than either option and improves both, which makes it the correct first purchase far more often than a platform subscription does.

At a glance

Published virtual assistant rate ranges and model token prices used by this calculator, read on 2026-08-20. Note that the regional ranges are not like for like: some are wage only and some are all-in.

ItemRateWhat it includesSource
Philippines VA$4 to $12 per hourWage only; low end direct hire, high end agency managedcallforce.global 2026 rate guide
Filipino VA median ask$6.00 per hourMedian requested rate, middle 50% between $5 and $7ARM Sourcing report, 1,385 applications
Filipino VA median monthly ask$714 per monthAcross all applicants; only 13% quote hourlyARM Sourcing report
Caribbean nearshore VA$12 to $22 per hourAll-in: pay, recruiting, training, supervision, QA, equipment, covercallforce.global 2026 rate guide
US direct hire$28 to $48 per hourFully loaded: wages, payroll taxes, benefits, overheadcallforce.global 2026 rate guide
US premium VA firm$33 to $65 per hourDedicated fractional assistant via managed providercallforce.global 2026 rate guide
OpenAI GPT-5 nano$0.05 in / $0.40 out per 1MStandard tier, no caching discountdevelopers.openai.com/api/docs/pricing
OpenAI GPT-5 mini$0.25 in / $2.00 out per 1MStandard tierdevelopers.openai.com/api/docs/pricing
Gemini 2.5 Flash-Lite$0.10 in / $0.40 out per 1MPaid tierai.google.dev/gemini-api/docs/pricing
Claude Haiku 4.5$1.00 in / $5.00 out per 1MStandard tierclaude.com/pricing
Claude Sonnet 5$3.00 in / $15.00 out per 1MStandard tierclaude.com/pricing
Make Core plan$9 per month10,000 credits, 1 usermake.com/en/pricing
Make Teams plan$29 per month10,000 credits, multiple usersmake.com/en/pricing
n8n Cloud StarterEUR 20 per month2,500 executions, unlimited stepsn8n.io/pricing
Zapier Pro entry$19.99 per month annual750 taskszapier.com/pricing

What mistakes should you avoid?

  • Comparing token cost against an hourly rate while leaving human review out of the agent column entirely.
  • Comparing a wage-only offshore rate against an all-in managed rate as though they were the same product.
  • Timing the happy path and ignoring the exception handling that is the actual reason the human is there.
  • Amortising the build cost over a workflow lifetime you have no evidence for, which flatters the agent in month one.
  • Assuming the agent needs no management, when a changed form field or a moved button can break it silently.
  • Automating a task at low enough volume that even a perfect agent saves less than the build cost.

FAQ

Is it cheaper to automate portal data entry with an AI agent or outsource it to a VA?
It depends almost entirely on volume and review share, which is why this page is a calculator rather than an answer. Portal data entry is unusually favourable to agents because the task is bounded, high volume and low token count per item, so the per-item token cost is often a fraction of a cent against several minutes of human time. The catch is the review share: a data entry error is silent and compounds, so the honest review percentage tends to be higher than people assume, and that is where the saving goes.
What does a virtual assistant actually cost per hour in 2026?
Published 2026 rate reporting puts Philippines VAs at roughly four to twelve dollars an hour for wage-only arrangements, Caribbean nearshore at twelve to twenty-two dollars all-in, US direct hire at twenty-eight to forty-eight dollars fully loaded, and US premium managed firms at roughly thirty-three to sixty-five dollars. A separate report of 1,385 Filipino VA applications collected between October 2025 and July 2026 found a median requested rate of six dollars an hour with the middle half between five and seven. Both sources are linked on the page.
How do I compare a marketing VA against an AI agent?
Split the role into tasks first. Scheduling posts, formatting newsletters, pulling report numbers and drafting first-pass copy are four different jobs with four different token profiles and four different review requirements. An agent may win decisively on two of them and lose on the other two. Running the calculator once per task gives you a hybrid answer, which is almost always the correct one for marketing support.
How long until the build cost pays back?
The calculator divides your one-off build cost by the monthly gap between the VA bill and the agent's steady-state running cost. If the gap is negative there is no payback at any horizon and the design needs to change. Be careful about amortising over a period longer than you can defend, because a workflow that gets rebuilt in six months never earned its thirty-six month amortisation.
Does the agent replace the VA entirely?
Rarely, and the calculator does not assume it does. The realistic pattern is that the agent takes the bounded high-volume portion and the person keeps the exceptions, the escalations and the judgement. That is why the review share input exists: it is where the human hours go after automation, and pretending it is zero is the same as pretending the exceptions stopped happening.
What does this calculation not price?
Judgement, escalation ownership, continuity, ramp time and rework. A VA notices the form is wrong; an agent fills it in confidently. A VA takes holiday; an agent breaks when a portal changes its markup. These are real costs with real magnitudes that vary far too much by situation for us to put a number on, so the page names them instead of inventing figures.

Sources & further reading

Was this page helpful?