Freelancers doing AI work earn 34% more per hour than freelancers who don't (Upwork Future Workforce Index 2026). The average US merit increase budget for 2026 is 3.2% (Mercer, 2025).
That is the whole problem in two numbers. The AI skills premium is being set outside your building, an order of magnitude faster than your internal comp cycle can move, and it is happening inside the same twelve months.
It is not a forecast. It is already being paid โ just not by you, and not to your people while they are still your people.
What Upwork Actually Measured
Upwork released its second annual Future Workforce Index on 14 July 2026: a survey of 2,400 US-based skilled knowledge workers combined with transaction data from its own marketplace (Upwork, 2026).
Three findings matter for anyone running operations at 50โ500 people.
The share of skilled US knowledge workers who freelance went from 28% to 38% in one year. Not a decade โ one year. Alongside it, 58% of full-time employees now say they are considering freelancing, up from 36% the prior year.
And on the marketplace itself, freelancers performing AI work earn 34% more per hour than those who don't, across every work category.
Take the survey numbers as directional and the marketplace numbers as observed. The first is what people say; the second is what clients actually paid. The second is the one that should keep you up.
The Split Is Judgment, Not AI Usage
The finding that gets flattened in every summary is the one that matters most: AI usage does not confer a premium. Judgment does.
Freelancers doing more complex work with AI saw earnings rise 45% year-over-year. AI-augmented professional services โ domain experts folding AI into an established field โ grew 72% in volume with earnings up 22% (Upwork, 2026).
Now the other direction. Generative AI and creative production work saw 90% year-over-year growth in contract starts โ and a 13% decline in per-contract earnings.
Read those two together. Volume up 90%, price down 13%. That is what commoditisation looks like in real time. The same technology that raised the price of judgment-intensive work lowered the price of execution work, in the same market, in the same year.
Upwork names the winning profile the "AI Orchestrator": someone who connects AI tools to domain expertise, applies human judgment, and converts AI-enabled output into a business result.
You already employ these people. They are not your newest hires with AI in the job title. They are your senior operators โ the ones who knew the process cold before the tooling arrived and can now tell in ten seconds whether a model's output is usable. That combination is exactly what the external market just repriced upward.
Your Comp Band Moves at 3.2%
Set the two rates next to each other and the arithmetic is uncomfortable.
Mercer's US compensation planning survey puts the 2026 national average merit increase at 3.2%, with total increase budgets at 3.5% โ essentially flat against 2025 (Mercer, 2025). Merit budgets are designed for stability. They are averaged across a population, negotiated a year ahead, and constrained by internal equity.
None of those properties survive contact with a market that reprices a skill category by double digits inside a single cycle.
This is not an argument for abandoning bands. It is an argument for noticing that a band built to distribute 3.2% fairly across 200 people cannot, by construction, respond to a 34% move in one segment of those 200. The instrument is not broken. It is being asked to do something it was never built to do.
The predictable failure sequence follows from there, and it is worth naming because most teams live through it without recognising it. An operator's market value moves. Nothing internal registers the move, because no internal system is watching that signal. They test the market โ often informally, often via a single inbound approach they would not have entertained a year ago. By the time a retention conversation happens, it is a counter-offer negotiation, which is the most expensive and least reliable version of this conversation available. Counter-offers also reset nothing structural: the band that failed to price the role still fails to price it, now with a documented exception attached.
And the replacement cost is not theoretical. Gallup puts the cost of replacing an employee at one-half to two times their annual salary (Gallup). For a senior operator with deep process knowledge, assume the top of that range and add the months where nobody is doing the thinking.
The Case Against Taking the AI Skills Premium at Face Value
I want to be honest about the source, because the reflex to dismiss it is partly correct.
Upwork is a freelance marketplace. A report finding that freelancing is accelerating and that freelancers are earning more is, commercially, exactly the report it would want to publish. The marketplace data is also self-selected: it describes freelancers who transact on Upwork, not the US labour market. And a survey of 2,400 knowledge workers reporting that 58% are "considering freelancing" measures sentiment, not resignations. Considering is cheap.
So test the claim against sources with no stake in the answer.
Lightcast analysed job postings โ employer-side data, not marketplace data โ and found roles requesting AI skills carry a 28% salary premium over the same roles without them, with demand spreading well beyond the tech sector (Lightcast, 2025). That is employers voluntarily paying more, in their own ads, a year before the Upwork index.
MBO Partners, tracking the independent workforce for fifteen years, counts roughly 73 million US independent workers and a rising trend line that predates the current AI cycle entirely (MBO Partners, 2025).
The vendor incentive is real. The underlying signal survives it. Two independent datasets โ one from employers' own postings, one from a fifteen-year panel โ point the same direction. What none of them tell you is whether your people are leaving. That is a question only your own data answers, and most operations teams have never asked it in this specific form.
Where the Risk Actually Sits in a 200-Person Company
Retention risk from an AI skills premium does not distribute evenly. It concentrates, and it concentrates somewhere counter-intuitive.
It is not your junior execution staff. Their market just got cheaper โ that is the 13% per-contract decline. It is not your AI specialists either; you are probably already paying them attention precisely because their titles are visible.
The exposure is your mid-to-senior domain experts who quietly became AI-fluent in the last eighteen months and told no one. The operations manager who rebuilt the forecasting workflow around a model on their own time. The finance lead who automated half the close. These people carry the exact combination the external market is bidding for, and they are usually sitting mid-band because their formal role description hasn't changed.
They are also, structurally, the least likely to raise it with you. Their leverage is quiet.
There is a second-order effect worth pricing. When one of them leaves, the loss is not a headcount slot โ it is the informal process knowledge that made the AI-augmented workflow work at all. The tooling stays. The judgment that made the tooling productive walks. Whoever inherits the workflow inherits the outputs without the reasoning behind them, and within a quarter the team is running a system nobody fully understands. That is the version of this risk that never appears in a retention dashboard, because the dashboard counts departures rather than what departed with them.
Three Moves Before the Next Comp Cycle
Price judgment separately from execution
Most mid-market bands are built around role and tenure. Neither is now a proxy for market value. Before the next cycle, split your roles into judgment-intensive and execution-intensive, and check whether your bands treat them differently. If a role's core output is a decision under ambiguity, it belongs in a different pricing conversation from a role whose core output is throughput โ even if they sit at the same level today.
Find your AI-fluent operators before the market does
Run one inventory this quarter: who has materially changed how their own work gets done using AI in the last year? Not who has a licence โ who has redesigned a process. That list is usually shorter than expected and rarely matches the org chart's seniority ranking. It is also your actual flight-risk register, and it costs a week to build.
Use the contingent market on purpose, not as a leak
If 38% of skilled knowledge workers now freelance, that is a supply pool as much as a threat (Upwork, 2026). The failure mode is losing an employee to it in April and hiring them back as a contractor in September at 1.6ร the rate, with no knowledge transfer in between. Decide deliberately which capabilities you want permanently in-house and which you are content to rent โ before the market makes that choice for you.
One Decision This Quarter
Open your comp band for your ten most judgment-intensive roles. For each, write down what that person could plausibly earn next quarter selling the same judgment externally, using the AI skills premium as your multiplier rather than last year's benchmark survey.
If the gap is larger than what you can close, you have not discovered a compensation problem. You have discovered which capabilities you are choosing to rent from people who currently think they work for you.
The 3.2% is already approved. The 34% is already being paid. Only one of those numbers is going to move this year, and it is not yours.
