More than half of businesses using freelancers in 2022 had stopped entirely by 2025. The data is clear on which categories are gone. Here's the honest category by category breakdown and how to make the call for your own arrangements.
The email sits in the inbox. It's from someone you've paid for two years. Good work, reasonable rate, reliable enough that you stopped thinking about it. Their contract renews next month. You've been meaning to reply.
The reason you haven't is sitting in another tab.
That's where most of these decisions actually live right now. Not in a strategic meeting. Not in a formal review. In the small, uncomfortable gap between a renewal reminder and the tool you've been quietly using instead.
Ramp's economics lab tracked exactly this pattern using real spending data from their business customers, actual card transactions, not surveys. They measured what companies spent on freelance platforms like Upwork and Fiverr versus what they spent on AI providers like OpenAI and Anthropic, from 2021 through to 2025.
The numbers are unambiguous. Freelance marketplace spending fell from 0.66% of total business spend in Q4 2021 to 0.14% by Q3 2025. Over the same period, AI provider spending went from zero to nearly 3%. More than half of businesses that were using freelance platforms in 2022 had stopped entirely by 2025. The companies that spent the most on freelancers switched the fastest.
The substitution rate among high-exposure firms was $1 of freelance spend replaced by $0.03 of AI spend. Ninety-seven cents in the dollar, kept.
Ramp's lead economist named the common thread across everything being replaced: "well-scoped tasks that don't require a lot of institutional knowledge." The examples given were specific: blog writing, logo design, code fixing. That framing is more useful than any category list. It's the diagnostic that tells you which of your current arrangements are actually at risk, and which aren't.
The Category Breakdown
Written content.
Commodity written content is gone. Freelance writing postings have fallen 33% since ChatGPT launched, according to Brookings research. That number maps cleanly to what's actually happening: the brief-following, format-repeating, output-volume work that a lot of content budgets were built on. Four blog posts a month to a standard brief. Weekly social copy to a template. Email sequences written to a spec. AI does all of that now, faster and cheaper.
What's still worth paying for is harder to brief. A writer who understands your audience's specific psychology. Someone who can take a position your competitors won't take, defend it without hedging, and make it sound like you said it. The test is simple: if you could hand the brief to a capable AI tool right now and get something acceptable, you were buying output. If what made the work good was how that particular person thought about your problem, you were buying thinking. Thinking is not gone.
Logo and basic graphic design.
Basic visual work is largely substitutable. The work that lived here was always well-scoped by nature: make a logo from these elements, create social templates in these dimensions, produce banner ads following this format. That work has moved to AI design tools. Graphic design postings are down 17% overall, and the steepest declines are in exactly this kind of task-based output.
What's not gone is visual identity work that requires real creative direction. A brand system that holds together across every context your business operates in. Campaign creative that needs to be unexpected enough to actually stop someone scrolling. The more your design brief was "make this look professional," the more that relationship has been replaced. The more it was "make this feel like us in a way we haven't figured out how to articulate yet," the more it's still a human job.
Data entry, research and admin.
Gone. This isn't a category where there's a nuanced line to find. Data entry was always a task sold by the hour for its tedium. Basic research compilation, document formatting, spreadsheet work, summarising reports — AI handles all of it faster, at any hour, with fewer errors. Data entry postings are down 35%. If you're still paying a person for this work in 2026, that's a deliberate choice with a real cost attached.
Translation.
Basic translation between common languages for standard content is largely solved. Postings are down 19 to 28% depending on language pair and platform. A product description, a blog post, a marketing email — if the goal is accuracy and readability, AI translation has crossed the threshold where it's hard to justify the human cost for standard content.
The case for a human gets stronger as the stakes get higher. Legal documents where a mistranslation creates liability. Technical content where precision matters more than fluency. Localisation work, which is different from translation: cultural adaptation, idiomatic choices, knowing what lands differently in a specific market versus what the words technically mean. The bigger the reputational or legal exposure, the stronger the argument for a specialist. For everything else, the math changed.
Basic development.
Down 21% in postings, and the Ramp data specifically names code fixing in its list of replaced categories. Entry-level development work: fixing bugs, building standard components, implementing something well specified from a template — has been absorbed into AI coding tools faster than most people expected. A Stanford study in 2025 found entry-level software engineering jobs declining as AI tools took on the boilerplate and the routine fixes.
What's still a real job is the work that requires understanding why, not just how. Architecture decisions. Figuring out what the solution actually is before anyone can implement it. Integration work that requires understanding how systems behave inside a specific business, with its specific constraints and history. If you were hiring someone to implement a solution you'd already figured out, that work is increasingly gone. If you were hiring someone to think alongside you about what the right solution was, that relationship has more durability than almost anything else on this list.
Short-form video editing.
Moving fast. AI video tools have crossed the quality threshold for social content. Instagram Reels, LinkedIn clips, short YouTube content, the routine editing work that goes into this has largely been automated. The signal is in the platform data: AI video generation and editing is up 329% on Upwork year-on-year as of early 2026, which means the demand has shifted from "edit my footage" to "build me an AI video system." Businesses aren't hiring less video help overall. They're hiring different video help.
What's not gone is narrative video with real brand purpose. Anything where the editing is the creative decision, not a technical execution. A documentary-style piece where the structure tells the story. A brand film where the pacing is the argument. If the video was "make this footage presentable," that work has moved to tools. If the video was "make an argument," it hasn't.
Customer support scripting and chatbot content.
Gone, or close enough that the distinction doesn't matter for most businesses. Scripting FAQs, writing canned responses, building support flows, this was always task-based work with a clear spec. AI handles it directly now, often not just writing the scripts but running the conversations. Customer support gig postings are down 16%. What's still a job is designing the support system strategy, handling escalations where the customer needs to feel heard by a person, and the situations where something went genuinely wrong and the relationship needs repair. Those require judgment that AI doesn't have. The rest doesn't.
What's not gone.
The work that's holding isn't a category so much as a characteristic. It's work that's hard to scope precisely because it requires understanding how a specific business thinks. A brand strategist who has worked with you for three years isn't a content factory. A conversion copywriter who knows exactly how your customers talk about the problem you solve isn't a blog factory. An AI systems builder who can wire your tools together into something that actually works for your specific stack is now one of the fastest-growing freelance categories on Upwork, up 109% year-on-year as of early 2026.
Ramp's economist put it plainly: the replacement pattern lives in tasks that are well-scoped and don't require institutional knowledge. The inverse holds. Work that's hard to scope without deep context, work where the output is inseparable from the judgment behind it, work where a new person would need months to understand what they were actually being asked to do — that work has gotten more valuable, not less.
The Decision Framework
Two questions. Not ten.
Is the task well-scoped? Can you write a clear brief for this work in under thirty minutes that would produce an acceptable result from a capable AI tool right now? If yes, the economic case for a human is weak. Not zero, quality still varies, your time costs something too, and "acceptable" isn't always good enough. But weak, and getting weaker as the tools improve.
Does it require knowing your business specifically? Not knowing your industry. Your business. Your voice, your customers, the decisions you've made and why, the constraints that aren't written down anywhere. If the answer is yes, that relationship has a durability that a subscription doesn't. Institutional knowledge is the moat, and it's one of the few moats that AI makes stronger rather than weaker, because the gap between someone who knows you and a tool that doesn't keeps widening as your business gets more specific.
If both answers point the same way, the decision is clear. If they diverge, the task is well scoped but the person knows you deeply, that's where judgment comes in, and there's no formula for it.
The Thing Nobody Says
Most of the founders who have cut commodity freelance work in the last two years have found something unexpected on the other side: the people they kept matter more now, not less.
When everything average is automated, genuine talent becomes visible in a way it wasn't before. The writer who makes you sound like yourself. The designer who sees something you couldn't articulate. The developer who talks you out of a bad idea before you build it. Those relationships have gotten more valuable because the baseline around them has been replaced. It used to be that good work existed in a field of okay work, and the difference was easy to miss. Now the okay work is gone, and the good work stands alone.
This isn't consolation framing. It's what actually happens when you make these decisions honestly. Cutting the work that AI can do is what creates the budget and the attention to properly value the work it can't.
There's still an email in the inbox. The honest question isn't whether you like the person or whether they've been reliable. It's whether what you're buying from them is something a tool can now do, or something only they can do.
Most founders already know the answer for at least one of their current arrangements. The ones who are honest about it — in both directions, cutting what should be cut and valuing what should be valued, end up building something better than what they had before.
