AI Virtual Assistant Business: Real Earnings, Real Failure Rates

Two founders launched AI assistant businesses at almost exactly the same time. Both had real customers. Both had backing from people who knew the space. One hit $10 million in annual recurring revenue within six months. The other shut down within a year.

Richard Hollingsworth didn't start with AI. He spent nine years running a human executive-assistant agency before building Fyxer — logging, by the time he was done, 500,000 hours of observed assistant work. When he finally built the product, he already knew which tasks burned the most time and where a language model could genuinely help. Within six months of launch, Fyxer hit $10 million in ARR.

Supreet Hegde had a stronger résumé on paper. IIT-Madras graduate. Former VP at a funded AI startup. Backed by Perplexity's founder. His company, Astra, landed two enterprise clients in beta. Then it shut down in 2025 — killed not by a bad demo, but by buyers who wouldn't grant access to Salesforce and Slack, a slow enterprise sales cycle, and a cofounder disagreement about growth pace.

The gap between those two outcomes tells you more about this market than most YouTube channels will in three hours. The question isn't whether AI virtual assistants can make money. Some clearly do. The question is whether you are set up to be on the right side of that gap — and that's something you can actually evaluate before committing anything.

What Actually Separates the Winners

The founders who built durable AI VA businesses didn't just know how to use the tools. They understood, in granular operational detail, the exact work they were automating and where it breaks.

Hollingsworth's pitch to his technical cofounder wasn't a market slide deck. It was a 500,000-hour time-tracking dataset showing exactly which EA tasks consumed the most time and where human judgment couldn't be skipped. That's a different starting point than a weekend demo built on a borrowed API key.

Austin Petersmith's scheduling assistant, Howie, routes uncertain cases to human reviewers — not because the AI failed, but because he understood from the start that a missed calendar invite isn't a minor inconvenience, it's a broken professional commitment. That operational knowledge shaped the product architecture before a single line of code was written.

The main thing that makes this really hard is that mistake tolerance is very low. Building AI assistants that work well is hard.
— Austin Petersmith, Co-founder and CEO, Howie

This pattern isn't limited to software founders. Gartner predicted in June 2025 that more than 40% of agentic AI projects would be canceled by end of 2027, citing two primary causes: unclear business value and inadequate risk controls. Both are downstream failures of not knowing the work before automating it.

The technology is not the constraint. Knowing where it breaks — and planning for that before launch — is.

This mechanism isn't exclusive to startup founders. A freelance VA who has spent three years handling client inboxes understands failure modes that someone who just completed a no-code course does not. That experiential gap is what the research is actually measuring, not job titles or funding rounds.

Before You Build Anything, Answer These Four Questions

The following questions will tell you more about your likely outcome than any income promise will.

The first question: can you describe, from memory, where the work breaks? Hollingsworth's team had logged every edge case before writing AI code. Astra's founders optimized for the enterprise pitch. The one who knew the failure modes built the product that lasted. If you can't list five ways the task you want to automate goes wrong without Googling, you're starting in the wrong place.

The second question: can you absorb one client canceling without collapsing? When Fyxer grew from $1 million to $5 million ARR in just ten weeks, customer support response times ballooned from five minutes to five hours. Even a well-prepared, well-funded team hit a delivery wall under rapid growth. A solo operator with two clients has no buffer. If losing one contract would threaten your ability to keep running, your financial foundation isn't ready — not your product.

The third question: can you define your outcome in a way your client can measure? Gartner's cancellation forecast pins unclear business value as the primary cause of project failure — not bad AI, not bad tools. "I'll help you with your inbox" is a category. "I'll reduce your unread emails by 40% within 30 days" is a measurable outcome a client can evaluate and renew against.

The fourth question: have you confirmed your target client will actually grant data access? Astra's enterprise buyers balked at connecting Salesforce, Slack, and Google Drive. The AI was ready; the trust wasn't established. Before you build an integration, have a real conversation with a real prospect about what they'd actually permit you to touch. Listen carefully to the hesitation.

Failing two or more of these questions doesn't mean the opportunity doesn't exist for you. It means you need a different starting point — smaller scope, a different client type, or a period of doing the work manually before you automate it. These questions apply equally to a customer service manager building an AI triage layer, a recruiter automating first-touch candidate outreach, or a content strategist offering AI-assisted social management. The failure modes differ; the prerequisite logic is identical.

What the Earnings Actually Look Like

No credible survey reports the median annual income of AI VA operators. What exists is enough public data to build an honest model — and that model is more instructive than any claimed average.

Upwork's published median displayed rate for virtual assistants is $13 per hour. The platform's service fee for new contracts, as of May 2025, ranges from 0% to 15% depending on market conditions. Everything below is illustrative math built from those anchors, not a survey median — because none exists.

At 20 billable hours per month, you gross $260 before fees. After a 15% platform fee and a conservative $200 in monthly software costs, you're left with $21. That covers coffee, not rent. At 60 billable hours, you gross $780, net roughly $463 after fees and software — meaningful side income. At 100 hours, you gross $1,300 and net around $905 before self-employment taxes.

That optimistic scenario leaves no time in the model for prospecting, quality review, client communication, or fixing errors. The $905 is before taxes. A specialist charging well above the median rate changes the numbers; a generalist charging below it makes them worse.

The math doesn't say this isn't worth doing. It says the "six figures passive" narrative requires an entirely different rate, client type, and business model than a solo VA charging marketplace rates. If you're evaluating this as a career replacement, the numbers require either specialist positioning with higher rates, recurring retainer clients rather than one-off projects, or an agency model with contractors — all of which require the four prerequisites above to be firmly in place first.

These numbers apply to any reader offering AI-assisted services, not just those calling themselves AI VAs. A marketing freelancer adding AI workflow management, an HR contractor offering AI-screened candidate summaries, or a bookkeeper using AI to scale client capacity faces the same arithmetic of billable utilization versus unpaid overhead.

The Structural Constraint Nobody Talks About

The pitch that kills most AI VA deals — "we'll replace your current assistant" — is also the one customers trust least.

In a Gartner survey of 3,566 B2B and B2C customers conducted in February and March 2026, 50% said AI made their service interactions easier. But 87% said companies using AI must provide access to a human agent. The majority want both — not a replacement.

Klarna announced in early 2024 that its AI assistant had handled 2.3 million customer service conversations — equivalent to 700 full-time agents — in a single month. A year later, the company was actively recruiting human agents again, saying customers should always have the option to speak with a person. The chatbot still handles two-thirds of volume. The humans handle what the chatbot cannot.

We grew from $1 million to $5 million in ARR in two and a half months, and we hadn't planned for that outcome well enough.
— Richard Hollingsworth, CEO and Co-founder, Fyxer

NBER research on AI-assisted customer support agents found a 14% average productivity increase — but the gains were largest for less experienced workers being helped by AI, not for autonomous systems replacing workers entirely.

The operators who charge more aren't selling better AI. They're selling accountable human judgment that AI makes faster and cheaper to deliver. If your pitch is "AI with experienced human oversight," the 87% of buyers who demand human access are your market, not your obstacle.

This reframes the whole opportunity. Human-in-the-loop is not a weakness in your offering. It's the premium product. That structural feature holds across customer service, scheduling, content review, and administrative work — it's not a niche quirk.

Start Here Before You Bet on It

Hollingsworth didn't spend nine years in an EA agency because he lacked the ambition to launch faster. He spent it because that's how long it took to understand the work well enough to automate it reliably. You don't need nine years. But you do need to honestly assess whether you understand the specific workflow you're planning to sell before you build a demo for it.

The AI VA market is real. The failure rate is real. The gap between the two isn't about credentials — Hegde's were stronger than most readers'. It's about whether you've done the work of understanding the work. That's something you can assess before committing anything.

Here's a test you can run this week: pick one administrative task you already handle well — for yourself or someone else. Time every step for five working days. Write down every exception, every judgment call, every moment you had to override the obvious answer. That document is your pilot brief. If you can't fill a page, you're not ready to sell an AI solution for that task. If you can fill three, you might have a real business.

The best AI VA operators didn't start with a tool. They started with a problem they already knew how to solve.

Recommended Tools & Resources