You are probably weighing the same question many firm owners ask: AI agent vs virtual assistant, which one do you actually hire for the next block of admin work? The org chart keeps growing, and every new seat adds management time you do not have.
This is a framework, not a verdict. The numbers below are illustrative, not a guarantee. They come from stated assumptions you can swap for your own. I have not invented a study or a client, and no engagement runs on the same clock.
The assumptions behind the comparison
To keep this honest, here are the inputs. Change them and the answer changes.
- Full-time human VA: about $60,000 a year all-in. That is a market range, not a quote. Onshore assistants often run higher, offshore ones lower.
- AI agent: about $400 a month in model usage and tooling for a working agent. Real usage costs vary with volume.
- Setup and oversight: agents are not free to build. Someone has to scope them, test them, and review the output. I include that below, because most comparisons skip it.
A $400 agent is not a $400 outcome. The cheap part is running it. The expensive part is the design work that makes it reliable.
The 11-task comparison
Here is how I score agents against a human VA on eleven common operational tasks. A score of "agent" means the agent can do the bulk of it with light review. This is my judgment from first principles about how these tasks are structured, not measured data.
| Task | Better fit | |---|---| | Inbox triage and categorising | Agent | | Drafting routine replies | Agent | | Meeting scheduling and reminders | Agent | | Data entry between systems | Agent | | Weekly reporting and summaries | Agent | | Lead and intake follow-up sequences | Agent | | Document and invoice processing | Agent | | Research and briefing notes | Agent | | Status chasing across the team | Agent | | Sensitive client conversations | Human | | Ambiguous judgment calls and exceptions | Human |
Nine of eleven lean toward the agent. That does not mean nine jobs vanish. It means nine categories of work are structured enough to hand off.
Where agents score 9 out of 11, and why
The reasons are specific, and they repeat across the nine tasks.
The work is rule-shaped. Triage, data entry, scheduling and reporting follow patterns. If you can describe the rule to a new hire in a page, an agent can usually follow it.
The work is high-volume and repetitive. A person gets tired and inconsistent at the four-hundredth invoice. An agent runs the same steps every time.
The work is always on. An agent handles the 11pm inquiry without overtime. For a firm with clients across time zones, that matters.
The work has a clear check. You can verify a reconciled entry or a booked meeting. Where output is checkable, you can build evals around it. If that is new to you, see how to build a production-grade AI workflow with real evals.
The honest failure mode: agents are confident when wrong. A misfiled document or a bad summary looks fine until someone relies on it. That is why review steps and permissions are part of the build, not an afterthought.
The 2 tasks where human VAs still win
Sensitive client conversations
When a client is upset, grieving a bad outcome, or negotiating a fee, tone and relationship carry the message. An agent can draft the first pass. It should not send it. Look for a VA who reads a room, writes with warmth, and knows when to escalate to you.
Ambiguous judgment and exceptions
The unusual case, the client who breaks the pattern, the request that does not fit any category: this is where a person earns the salary. Look for someone who asks a clarifying question instead of guessing, and who knows your firm's unwritten rules.
Notice what these two have in common. Both are about accountability and context, not speed. In a professional-services firm, that is where your reputation sits.
The hybrid pattern: one VA plus eight agents
The comparison most owners need is not human versus machine. It is three full-time VAs versus one VA supported by a set of agents.
The hybrid keeps one person for the two human tasks, plus oversight of the agent output. Eight narrow agents cover the rest: triage, drafting, scheduling, data entry, reporting, follow-up, document processing and research.
A caution on the number eight. Research from Vals AI, reported by The Decoder, found multi-agent teams showed measurable quality gains in only one of four tests while costing up to 5.1x more. More agents is not automatically better. Start with one agent on one task, then add. We cover that approach in price one agent first.
Illustrative output comparison, not a guarantee:
- Three VAs: three people's worth of hours, bounded by working hours, with handoffs, sick days and training between them.
- One VA plus agents: one person's judgment, with the repetitive volume handled around the clock. The VA's time shifts from doing to reviewing.
The hybrid wins when your work is mostly the nine agent-shaped tasks. It loses if your work is mostly relationship-heavy, where you need the people.
The 3-year cost comparison
Here is the arithmetic using the assumptions above. I add setup and oversight to the agent models so the comparison is fair. These are illustrative figures built on stated inputs, not a guarantee of savings.
Model 1: three full-time VAs
- $60,000 x 3 = $180,000 a year
- Three years: about $540,000
- Excludes recruiting, training, management time and turnover, which are real but vary widely.
Model 2: one VA plus eight agents
- One VA: $60,000 a year, or $180,000 over three years
- Eight agents at $400 a month each: $3,200 a month, or $38,400 a year, about $115,200 over three years
- Assumed build and review allowance: say $30,000 to $60,000 up front, then modest upkeep. Your number depends on scope.
- Three years: roughly $325,000 to $355,000
Model 3: agents only, no human VA
- Eight agents: about $115,200 over three years
- Same build allowance: $30,000 to $60,000
- Three years: roughly $145,000 to $175,000
- The catch: you or a partner absorb the two human tasks and all review. That time has a cost, and it usually lands on you, the person you are trying to free.
On paper Model 3 is cheapest. In practice, it often recreates the bottleneck you started with. Model 2 usually fits a firm that wants relief without losing the human layer.
Treat the $400 as a placeholder. If usage runs heavy, or you need larger models, it climbs. Price a single agent against your real volume before multiplying by eight.
How to decide for your firm
Run a simple method on your own time:
- List every task your assistant or team touches in a week.
- Mark each one rule-shaped or judgment-shaped.
- Note which have a checkable output.
- Price one rule-shaped, checkable task first.
If custom builds versus off-the-shelf tools is part of your decision, the compound edge argument covers the tradeoff.
A next step
If you want a second set of eyes on which of your tasks are agent-shaped and which need a person, the free 30-minute AI Opportunity Audit walks through exactly that. It is honest scoping, no pressure and no lock-in, and you keep everything we map out either way.
Visuals in this article are AI-generated.
Sir Tay Jackson
Founder, Air Flow Automations
Founder of Air Flow Automations. Builds custom AI systems for agency owners, e-commerce founders, and service operators who want their time back without adding headcount.
