AI Agent vs Virtual Assistant: What Each Really Costs a Small Business

You are paying $30 an hour to copy data between tabs

An owner bills clients at $150 an hour and spends five hours a week on invoices, inbox sorting, lead follow-up and calendar juggling. That is 20 hours a month of low-value work. The usual answers are a virtual assistant or, lately, an AI agent. They cost very different amounts and fail in very different ways, so the choice is not obvious.

This guide compares the two on cost, setup time and the kinds of work each handles well, using published rates and cases documented on this site. It does not pick a winner. For most small businesses the answer is a mix.

What each one costs

Virtual assistant rates first. As of 2026, vendor blogs that sell VA services put full-time Philippines-based assistants at about $5 to $15 an hour, or roughly $800 to $2,400 a month for 160 hours. US-based assistants run about $22 to $65 an hour, which is around $3,500 to $10,400 a month full time. Part-time is proportional: one of those sources estimates $480 to $1,280 a month for 15 to 20 hours in the Philippines. These figures come from companies that sell the service and ranges differ between sources, so use them as a guide and get real quotes. Direct hires are cheaper per hour than agencies, but you take on sourcing, training and replacement.

AI costs depend on how you use it. A chat subscription is the cheapest entry: as of November 2026 Claude Pro is listed at $20 a month billed monthly, or $17 on an annual plan, and Max starts at $100 a month. Automation platforms are cheap too. Make has a free plan with 1,000 credits a month, and n8n can be self-hosted. A fully always-on agent costs more in setup than in running costs. A consultant documented on this site runs an OpenClaw assistant on a $600 Mac mini for roughly $50 to $100 a month in API costs after about four hours of setup (builtwithagents.ai/strategy/openclaw-ai-chief-of-staff-mac-mini-adhd). Another builder reports running lead-response agents for local service businesses at $8 to $40 a month (builtwithagents.ai/strategy/19-openclaw-agents-local-service-businesses-8-dollars-month). Both are self-reported.

So the monthly gap is large. The hidden cost for AI is your time to build and watch it, and for a VA it is your time to train and manage them.

What each one is good at

A virtual assistant is good at anything that needs judgment, a voice, or a relationship. Calling a vendor who has not paid. Reading an odd email and deciding it matters. Booking travel with preferences you never wrote down. Noticing that something looks off. They also handle messy, one-time tasks that would take longer to automate than to do.

An AI agent is good at work you can describe in steps, at any hour, in seconds. Answering the same ten questions. Sorting and labeling email. Drafting replies for you to approve. Pulling data from invoices into a sheet. One accounting builder documented on this site reports a Claude agent posting roughly 2,047 invoices a day, with 94.3 percent auto-posted at launch (builtwithagents.ai/strategy/claude-agent-2000-invoices-day-accounting-firm). Another team reports AI agents removing 96 hours a week of busywork, including support tickets down from 45 hours to 12 a week (builtwithagents.ai/strategy/ai-agents-replaced-96-hours-weekly-busywork). Treat both as self-reported.

Where agents get into trouble: anything with no clear right answer, anything where a wrong action is costly, and anything that needs a human voice on the phone.

Step by step: deciding for your business

Step 1: List your recurring tasks for two weeks

Keep a plain list of every task you do that is not your core work, with minutes spent. Include the small ones: replying to the same question, rekeying an order, reminding someone about an invoice. Two weeks is enough to see patterns.

Step 2: Sort each task into three buckets

Bucket A: you can write the steps on one page and the input is predictable (invoice arrives, extract five fields, add to sheet). Bucket B: it needs judgment but follows a loose pattern (reply to an inquiry in your voice). Bucket C: it needs a human relationship or a phone call. Be honest. Most people overestimate Bucket A.

Step 3: Price each bucket

For Bucket A, estimate build time at 4 to 10 hours for a first agent and a monthly running cost of $10 to $100. For Bucket B, plan on AI drafts with you or a VA approving, and price the review time. For Bucket C, get VA quotes. Multiply your monthly minutes by a VA's hourly rate and compare.

Step 4: Start with one Bucket A task

Pick the most boring, highest-volume task. Build it in Make or n8n with a human approval step. Run it side by side with your current process for two weeks and compare errors.

Step 5: Hire a part-time VA for Bucket C and for supervising the agent

A VA at 10 to 15 hours a week can handle the human tasks and also check the agent's output, fix exceptions and update its instructions. That is the arrangement many agencies end up with.

Step 6: Review at 60 days

Compare hours saved, errors, and total cost. Add the cost of your own time. If the agent is not clearly beating its price, change it or kill it. Do not keep paying out of loyalty to a build.

A worked example

Take a solo consultant with 20 hours a month of admin. She finds that 8 hours are invoice entry and email sorting (Bucket A), 6 hours are drafting replies to prospects (Bucket B), and 6 hours are scheduling calls and chasing late payers (Bucket C). Full-time VA pricing makes no sense for 20 hours, but a Philippines VA at 15 hours a month at $8 an hour would cost about $120. She builds an invoice and email-sorting workflow in Make over a weekend and keeps AI drafts for prospect replies on her $20 Claude plan. Her new monthly cost is about $140 plus the weekend. The same 20 hours with a US VA at $35 an hour would be $700. Those rates are assumptions; replace them with quotes. Her saved time is the real prize: 20 hours at her $150 billing rate is $3,000 of capacity.

The middle option most owners miss

You do not have to choose between a full agent and a human. A chat subscription plus a part-time VA covers a surprising amount. The VA pastes the day's inputs into a saved Claude Skill or a set prompt, reads the draft, fixes it and sends it. You get the human judgment and the drafting speed, and you skip the weeks of building and debugging an automation. It costs more per task than a pure agent and far less than a VA doing everything from scratch. If you are unsure, start here and automate only the steps that your VA says are boring and identical every time.

What goes wrong

Agents fail silently. A VA tells you when something is confusing. An agent will often just do the wrong thing confidently, so build checks and sample its work weekly. VAs fail loudly but expensively: turnover, time zone gaps, training that walks out the door, and the management time that people forget to count.

Security matters for both. Giving either one access to your inbox, bank or client data is a trust decision. Give each the narrowest access that works, use separate logins, and avoid sharing credentials in plain text. For regulated data such as health information, check the legal rules before using either one.

Also watch for the shiny build. Several documented cases show agents with impressive features and no reported revenue effect. Judge a build by hours or dollars saved, not by how clever it is.

When not to do this

Do not automate a task you do fewer than a couple of times a month. Do not hire a VA to do work that an existing tool in your stack already does. And if you cannot yet write down how a task is done, you cannot hand it to either one.

Frequently Asked Questions