Brookings Found AI Cuts Freelance Earnings. Upwork Says It Raises Them. Both Are Right.

Here's the honest verdict on AI freelancing: it's a real opportunity, but probably not the one being sold to you. The tools are genuine. The demand is real — Upwork reports that AI-skilled freelancers earn a 40% hourly rate premium, and demand for AI-referencing skills grew 109% year-over-year on their platform. And yet Brookings found that freelancers in AI-exposed occupations saw contracts drop 2% and earnings fall 5% in the short term after generative AI arrived. Both of those things are true, simultaneously, for the same technology.

Kesar Rana is a useful place to start. She's a content strategist who'd spent five years building a freelance business writing for SaaS and technology companies. In 2023, the work slowed, then stopped. Clients said they were "experimenting with AI." She tried competing on price — $100 for a 1,500-word article, $50 to clean up an AI draft. By her own account, she was barely breaking even. Meanwhile, someone with an IT background was packaging AI automations for businesses, charging $800 to $3,500 a month per client, and claiming $47,000 in 90 days. Same technology. Very different trajectories.

The difference wasn't the tools. It was what was underneath them.

What the Numbers Actually Prove — and Don't

Before accepting any income figure in this space, there are three things worth knowing about the data.

First, there is no published median income specifically for AI-enhanced freelancers. The most credible proxy is Upwork's 2025 Future Workforce Index, which found that full-time skilled freelancers reported a median income of $85,000. That survey covers skilled knowledge workers broadly — writers, developers, marketers, analysts — not AI-specific operators. It's a useful benchmark, not a promise.

Second, the aggregate numbers hide catastrophic individual outcomes. Brookings' 2% contract decline is a short-run average across many occupations. But James Presbitero Jr., a freelance writer, reported a 90% income drop after ChatGPT arrived — while the aggregate statistic looked modest. The average can be fine while individuals are devastated, depending entirely on how exposed their specific service is to substitution.

Late 2022, ChatGPT wiped out a chunk of my freelance writing income overnight.
— James Presbitero Jr., AI Marketing Specialist

Third, some of the most-cited success figures don't hold up under scrutiny. Anuj Bhalla's Medium post claims $47,000 in 90 days and describes 14 clients at $800–$3,500 per month — roughly $23,000 monthly. A separate post claims $30,000 per month. Those figures don't reconcile cleanly, and neither is independently audited. Bhalla is a real person with a real professional background in IT and customer service analytics, and the recurring-automation model he describes is legitimate. But treat specific revenue claims as a possible outcome with no denominator, not a forecast.

This isn't an argument to distrust all data. It's an argument to ask the right questions: which population, what time period, gross revenue or owner income? The numbers cited in the rest of this article hold up to those questions. Most of what circulates online doesn't.

What Realistic Earnings Actually Look Like

The earnings picture splits into three tiers — defined not by which AI tools you use, but by what your service is actually selling.

The global freelancer average from Payoneer's 2023 survey of 2,000+ respondents across 122 countries sits at $23 per hour. That's the broad lower-to-middle market baseline. Upwork's specialized AI, technical, and consulting categories show directional ranges of $75–$150+ per hour. That's the premium ceiling — achievable when the buyer can't self-serve the outcome.

Most people aren't at either extreme. What separates the tiers isn't the tool stack. It's the layer of the value chain where the service sits.

Kesar, at her lowest point, was selling something AI could generate cheaply: a document. The $50 cleanup rate was the market's honest assessment of that value. Bhalla — self-reported figures and all — was selling something different: a maintained business process. His clients weren't paying for output. They were paying for a system that kept running and someone accountable when it didn't. That's a fundamentally different product.

The math matters here too. Gross revenue isn't owner income. Upwork fees run 0–15%; Fiverr takes 20%. A $2,000 project that actually requires 35 total hours — including sales calls, revisions, client questions, and support — yields roughly $57 per hour gross before expenses. Not the $200 hourly rate the project price implies. This is worth calculating for any service before assuming the premium tier is where you'll land.

The diagnostic question every reader needs to answer before buying a tool subscription or watching another tutorial: does your current service sell something AI can generate cheaply, or something AI can help you do more reliably at higher quality? Writers face this most acutely, but it's the same for designers (generic AI image versus brand system), developers (boilerplate code versus architecture ownership), and analysts (AI-generated dashboard versus interpreted business decision). The question is universal.

What Actually Kills These Businesses

Knowing which tier you're targeting is necessary. It's not sufficient. The most common reason AI freelance attempts fail isn't picking the wrong tier. It's four failure modes that have nothing to do with AI capabilities.

The first is the acquisition problem. Payoneer's 2023 survey found that 73% of freelancers name finding new clients as their top challenge. That problem predates AI and is not solved by tool access. Producing output faster doesn't generate qualified leads. If you don't have a reliable way to reach buyers, the automation doesn't matter.

The second is selling a tool rather than a business outcome. "I build AI agents" is an implementation description. A buyer needs a result: fewer missed leads, faster document review, lower support backlog. When the offer is defined by a tool, it's easily compared with templates and competing freelancers. When it's defined by a measurable process change, the comparison gets harder.

The third is ignoring implementation risk. RAND's 2024 report, drawing on interviews with 65 experienced data scientists and engineers, found that by some estimates more than 80% of AI projects fail — with root causes including unclear success metrics, insufficient data access, and ignoring monitoring and fallback requirements. A freelancer selling automation inherits those failure modes. A workflow that works in the demo can fail silently in production when client data changes or an API updates.

The fourth is confusing a successful demo with a viable product.

Carlos Rucker's first AI-agency client came from a Facebook comment — a business owner venting about administrative chaos. He responded, sold a follow-up automation for $150, and delivered it in two days. The client gave him a testimonial and two referrals. The lesson isn't the $150. It's that Rucker heard a specific operational pain before building anything, then delivered something testable fast enough to generate proof. He validated demand before building infrastructure.

I help local service businesses with follow-up, scheduling, and admin chaos.
— Carlos Rucker, Founder, Rucker Tech

A useful failure audit has four questions: How will you reach buyers without a marketplace or referral network? Can you name the specific business metric your service improves? Have you priced support and QA, not just build time? Has a real buyer paid for a narrow pilot, or only responded positively to a demo?

Where the Opportunity Is Actually Moving

The market isn't static, and two recent shifts are changing what "differentiated" actually means — in ways that create real opportunity for the right operators and real risk for everyone else.

The first shift is demand moving from novelty toward applied integration. Upwork's 2026 data shows AI integration grew 178% year-over-year, AI video generation and editing grew 329%, and AI data annotation grew 154%. The growth is concentrated in applied categories, not prompt-selling. Buyers who once asked "can you use AI?" are now asking "can you make AI work reliably in my business?" Those are different questions with different price points.

The second shift is accountability becoming a scarce, valued skill. An NBER field experiment tracking 5,179 customer-support agents found roughly a 14% productivity gain when workers had AI assistance. But the gain was shared between worker output and organizational efficiency — the worker captures it only when the client is paying for accountability, not just throughput. When an AI system is wrong, breaks, or produces something subtly off-brand, someone has to own that. Right now, that someone is expensive and rare.

Masato Hagiwara, an NLP specialist who billed $200 per hour in 2020, still earned less than full-time employment in his first independent year after accounting for utilization, insurance, unpaid admin, and cash-flow gaps during a move. High rate doesn't equal financial security. But his case also shows the floor that scarce expertise creates — and that floor is rising as applied AI work becomes more demanding.

The defensible position isn't "I can produce AI outputs faster." It's "I can tell you when the output is wrong, integrate the system with your data, and own what happens when it breaks." That's a different service, and it's the one the market is actively pricing upward.

What Kesar Finally Got Right

Which brings us back to Kesar — and the specific moment that explains why her recovery worked when competing on price didn't.

She eventually stopped selling articles and started selling editorial workflows, content audits, and AI-optimized strategy for SaaS clients. When she showed one founder how her edits made a draft "actually sound like us," the client renewed. That moment — a human recognizing what AI got wrong about their voice — was the service. The tools were infrastructure, not the product.

AI-enhanced freelancing isn't a new career. It's a new layer on expertise you already have. The freelancers gaining ground aren't the ones who learned the most prompts — they're the ones who figured out what their clients still need a human for, and made that the explicit product. The commodity layer is getting cheaper. The accountability layer is getting more valuable. Those are different opportunities, and only one of them requires a career reinvention.

Before evaluating tools, courses, or niches: write one sentence in this format — "I help [specific buyer] avoid or reduce [specific operational outcome] that AI alone gets wrong because [specific domain knowledge I have]." If you can complete that sentence specifically, you have a starting offer. If the sentence stays vague, that's the actual work to do first, and no automation platform will do it for you.

The opportunity is real. The prerequisite is being honest about what you're already good at — not what you're hoping AI will make you.

Recommended Tools & Resources