What does AI actually replace in a creator deal?

Written by Priyanshu Mishra, CEO of Lumienzo. Sources are named inline. Last reviewed 15 September 2026.

In June 1998, in the city of Leon in northern Spain, Garry Kasparov sat down across the board from Veselin Topalov for the second time in five weeks. The first match had not been close. Kasparov won it four games to nothing. This time each man had a computer at his elbow. Kasparov used a program called Fritz 5, Topalov used ChessBase 7.0, and the format was Kasparov's own invention, dreamed up a year after a machine beat him in public. He called it Advanced Chess. The question he wanted to ask was not who wins, man or machine. It was what the two of them can build together.

The six main games ended level, three to three, before Kasparov won a short tie break. Read that again. The player he had beaten four to nothing a month earlier had just held him to a draw. He explained why in the New York Review of Books in February 2010:

"A month earlier I had defeated the Bulgarian in a match of 'regular' rapid chess 4-0. Our advanced chess match ended in a 3-3 draw. My advantage in calculating tactics had been nullified by the machine."

The machine had taken the one thing Kasparov was the best in the world at, and handed the contest back to preparation, judgement and nerve. That is the shape of what is happening to your work right now, and it is a friendlier shape than the headlines suggest. A tool does not take a job. It takes the steps whose correct answer could have been written down before anyone started. In an ordinary brand deal, about half the steps are that kind, and you can name which half before you begin.

We build AI for creative work, so read this knowing that. Every figure below names who published it and when, the real losses come before the good news rather than after it, and the last section lists the popular numbers we refused to use because they do not survive checking.

The rule, and it is older than the chatbots

Three economists wrote it down more than twenty years ago. David Autor, Frank Levy and Richard Murnane published the finding in the Quarterly Journal of Economics in November 2003: computers take over tasks that can be written down as explicit rules, and they increase the amount of work in the tasks that cannot.

Almost everybody gets this backwards, because we assume a machine starts with the easy work and climbs. It does not. Playing chess is hard and a machine does it. Knowing which of two photographs makes someone cry is easy for you, and no machine can do it. The line is not difficulty. The line is whether the right answer can be fully described in advance.

Hold that rule up to your own week. Some of what you did could have been specified on a form before you opened your laptop. Some of it could only be decided by a person who knows the client, the audience, or the room. The first kind is getting cheap very quickly. The second kind is about to become the whole job, and the whole job is the part you probably liked.

The losses are real, and they go first

A page that tells you only the happy version has not earned the right to reassure you about anything. So here is what has already gone wrong for people who make things for a living.

The Society of Authors in the United Kingdom asked its members in January 2024 and got 787 replies. A quarter of the illustrators, 26 percent, had already lost work to generative AI. More than a third of the translators, 36 percent, said the same. Those are not forecasts. That work is gone.

It gets less comfortable. Xiang Hui, Oren Reshef and Luofeng Zhou studied a large freelance marketplace and published in Organization Science in November 2024, and the detail that should stop you is who got hurt: the best rated freelancers were hit hardest, not protected. Being good was not a shield. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab reported in August 2026 that employment for 22 to 25 year olds in the most exposed occupations is running about 19 percent below where it would be if it had tracked less exposed work. The youngest people are carrying this one.

History is blunter still. The US Bureau of Labor Statistics reported in September 2025 that one hour photofinishing went from 17,320 jobs in 2000 to 437 in 2024. That is a fall of 97.5 percent, and it is not a transition, it is a disappearance. When AT&T replaced human switchboard operators across more than half the American network between 1920 and 1940, James Feigenbaum and Daniel Gross found in the Quarterly Journal of Economics in August 2024 that younger women moved into other work while older women largely did not. A total that recovers can still be made of individual people who never did.

Keep all of that in mind while you read the rest, because the rest is genuinely hopeful and it is only worth something if you believe this part first.

A deal, one step at a time

Picture a normal campaign. A brand wants four creators for a product launch in six weeks. Here is every step, marked.

Writing the first plan. This goes. You can describe what a good plan contains before it exists: the product, the buyer, what to show, roughly what to say. Shakked Noy and Whitney Zhang put 453 college educated professionals through real tasks from their own jobs and published the result in Science in 2023. Time fell about 40 percent and quality rose about 18 percent, but the sentence worth keeping is about shape rather than speed. The tool, they wrote, "restructures tasks towards idea-generation and editing and away from rough-drafting". You stop starting from a blank page and start from an argument.

Deciding whether the plan is any good. This stays. No rule returns the answer "this is boring". And there is evidence that the human half of the pair is still the more original one. Leonard Boussioux, Jacqueline Lane, Miaomiao Zhang, Vladimir Jacimovic and Karim Lakhani ran an innovation challenge with 125 solvers and then had 300 independent evaluators judge the results, published in Organization Science in 2024. Their finding splits honestly: "while human crowd solutions exhibited higher novelty, both on average and for highly novel outcomes, human-AI solutions demonstrated superior strategic viability, financial and environmental value, and overall quality." People brought the surprising ideas. People steering a machine brought the workable ones. Neither side won outright, which is why the study is worth trusting.

Finding candidates. This goes. Matching stated requirements against measurable audience data is a rule shaped task, and it is the kind of counting that used to eat a junior person's fortnight.

Choosing which of them is right for this brand. This stays. A shortlist is a ranking and anybody can produce one. The choice is taste, and taste is what the client is actually paying for.

Agreeing the price, writing the contract, holding the money, chasing approvals, posting the product. All of this goes. Every one of those steps has a right answer you could specify on a Tuesday for work that happens in a month. The money side is the clearest case: the fee sits with a regulated escrow partner and releases when the work is approved, and a rule does that perfectly at three in the morning. This is where most of the hours in this industry currently disappear, and it is where they will stop disappearing.

Filming it. This stays, in the way that matters. A machine can generate a video. It cannot generate the fact that four hundred thousand people already choose to watch one particular person, and it cannot inherit the trust they built by turning up every week for three years.

Reading every comment afterwards. This goes, and honestly it was never done properly by hand. Nobody reads eight thousand comments. Automation is not taking this step away from a person. It is doing a step that was quietly skipped for a decade.

Deciding what to do about those comments. This stays. A tool can tell you that several hundred people asked where to buy the thing. Deciding that the next video opens with the link, and standing behind that decision in public when it does not work, is a person's job and will stay one.

This is not a page about agencies losing

It would be easy to read that list and conclude that the middle of the industry is finished. That is the wrong conclusion, and the sharpest correction comes from David Autor again, writing in the Journal of Economic Perspectives in 2015 about bank tellers rather than agencies: "a bank teller who can tally currency but cannot provide 'relationship banking' is unlikely to fare well at a modern bank."

Counting money could be specified, so it went to the machine. Knowing the customers could not, so it stayed, and the teller who knew the customers became worth more than the teller who counted well. The same test cuts through every role in a deal. If your product is processing plans and assembling reports, you are doing the counting. If your product is knowing which idea suits which client, holding the relationship and carrying the blame when it goes wrong, you are doing the banking. That applies to a talent manager, a producer, a video editor and a creator exactly as it applies to an agency.

There is a chess story about this, and it is the one everybody tells wrong. In June 2005 a website called Playchess.com ran a tournament where anything was allowed: bring a grandmaster, bring a supercomputer, bring both. It was won by two men from New Hampshire with no titles at all. Steven Cramton coached soccer at a prep school. Zackary Stephen was a database administrator with a masters in statistics. ChessBase reported in June 2005 that they played on three ordinary desktop computers and won the final two and a half to one and a half, against opponents rated around a thousand points above them. They were not better players and they did not have better hardware. They knew which program to believe in which kind of position, they split the analysis between them, and they argued about it. Kasparov, in that same 2010 article, described what they had found: "Weak human + machine + better process was superior to a strong computer alone and, more remarkably, superior to a strong human + machine + inferior process."

Better process beat both more talent and more computing power. Process is what a good agency sells. That is not a threat to the business. It is a description of it.

One more finding for anyone running a small team. Fabrizio Dell'Acqua and colleagues from Harvard Business School and Wharton ran a pre-registered experiment with 776 professionals at Procter and Gamble, published as a National Bureau of Economic Research working paper in April 2025. They reported that "individuals with AI matched the performance of teams without AI". Read that as a statement about what a small team can now attempt, not as a plan for having fewer people, because the second finding is the interesting one: the usual split, where the technical people propose technical answers and the commercial people propose commercial ones, disappeared. Everybody produced balanced solutions. People also said they felt better while working. The tool did some of the work a good colleague does.

What machines free people up to do

Here is the part to hold on to when the news is loud.

For more than ten years, scientists could not work out the shape of one enzyme from a monkey virus related to HIV. They had the crystals and they had the computers. Every automated method failed. So in 2011 a group at the University of Washington around David Baker put the problem into Foldit, a puzzle game where ordinary people fold proteins on screen for fun, and gave the players three weeks. The players solved it. The paper in Nature Structural and Molecular Biology in 2011 says it plainly: "Following the failure of a wide range of attempts to solve the crystal structure of M-PMV retroviral protease by molecular replacement, we challenged players of the protein folding game Foldit to produce accurate models of the protein." The author list carries the scientists and also two teams of gamers, credited by name. The software kept getting stuck on answers that looked locally fine. Humans are good at glancing at a shape and feeling that it is nearly right. The computers did the arithmetic. The people did the seeing.

Then the machines got very good at the arithmetic indeed. On 9 October 2024 the Royal Swedish Academy of Sciences gave half the Nobel Prize in Chemistry to Demis Hassabis and John Jumper for predicting protein structures, and the other half to David Baker for designing proteins that never existed in nature. Heiner Linke, who chaired the committee, described the pair of achievements in the announcement: "One of the discoveries being recognised this year concerns the construction of spectacular proteins. The other is about fulfilling a 50-year-old dream: predicting protein structures from their amino acid sequences. Both of these discoveries open up vast possibilities."

Look at what that prize did. It split the money between a machine that lifted an enormous piece of drudgery off humanity's hands and a human being using computers to invent things. And the prediction machine did not empty the field. The same announcement says more than two million people in 190 countries have used it, working on antibiotic resistance and on enzymes that break down plastic, on problems they could not previously have attempted at all.

That is the pattern, and it is not romantic, it is mechanical. Take away the counting and people do not do less. They reach further.

The honest ending to the chess story

You should know the part that gets left out. The advantage those teams had did not last. As the engines got stronger the human contribution in a chess game shrank towards nothing, and the one job the person used to do, choosing which program to trust, is now done by another program. The economist Tyler Cowen wrote on Marginal Revolution in February 2024 that in this kind of chess "the entity making those choices is now a program, not a human being". No formal tournament result settles it, so treat that as the accepted view among people who follow the game rather than as a measured fact.

It does not undo the argument. It sharpens it. Chess is a closed world with one goal and a perfect scoreboard, and that is exactly the kind of world a machine ends up owning. Your work is not that world. There is no score for whether this was the right campaign, no rule that settles which creator suits this brand, and no formula for what to make next. Where the goal itself is up for debate, the human is not a temporary advantage. The human is the point.

Where the tool will let you down

Being encouraging is not the same as being soft, so here are three warnings, each from people who measured them.

Fabrizio Dell'Acqua and colleagues also ran a field experiment with 758 consultants, published in Organization Science in 2025. On tasks the tool was good at, the people using it did more work, faster, at higher quality. On one task chosen deliberately because it sits outside what the tool handles well, the people using AI were about 19 percentage points less likely to reach the right answer than people working with no AI at all. They named the problem the "jagged technological frontier". The tool is brilliant at some things and confidently wrong at others, and it gives you no signal about which one you are looking at.

Medicine found the same edge. Philipp Tschandl, Harald Kittler and colleagues published a large study of skin cancer diagnosis in Nature Medicine in 2020, and their good news is excellent: support from a well built system made doctors better than either the doctors or the system managed alone, and the least experienced clinicians gained the most. Their warning sits in the same paper: "we find that faulty AI can mislead the entire spectrum of clinicians, including experts." Seniority is not protection. Attention is.

There is also a cost to everyone using the same helper. Anil Doshi at University College London and Oliver Hauser at the University of Exeter had 300 people write short stories, some starting from an AI idea and some not, and published in Science Advances in July 2024. The writers who had scored lowest on creativity improved the most, which is lovely. But the stories written with AI help were about 10.7 percent more similar to each other than the ones written without it. Individually better, collectively narrower. Treat a first draft as something to argue with, never as something to accept.

What to do with this on Monday

Take your own deal and mark it up. Write out every step from brief to payment and put one of two words next to each: specifiable, or judgement. Be strict. If somebody could have written the right answer on a form before the work began, it is specifiable, however long it currently takes you.

Then hand the specifiable list to a tool and watch what happens to your week. The plan, the shortlist, the contract, the money, the chasing, the comment reading: all of that is the counting. What is left is the part you are paid for and, if you are honest, the part you actually wanted to do.

That is the line Lumienzo is built along: it runs the specifiable half of a deal, from the first plan through to the escrow payment and the comment analysis, and leaves every judgement call with the person whose name is on the work. If that is how you want to work, get started.

Kasparov did not lose to the machine at Leon. He got a harder opponent and a better game, one that was suddenly about the things only he could bring. That is on offer here too, and it is closer than the doomsday version of this story would like you to believe.

The figures we refused to publish

This subject is full of numbers that fall apart when you check them. If you see these quoted elsewhere, now you know.

• "Two amateurs with laptops beat a supercomputer." Wrong twice. ChessBase reported in June 2005 that the winners used three ordinary desktop computers, and that both Hydra supercomputers were knocked out before the quarter finals, so the final was against a team of grandmasters with machines. The true version is remarkable enough without the embroidery.
• "Kasparov's Law", presented as his words. The widely shared wording is somebody else's rewrite. The sentence he actually published, quoted above, is from the New York Review of Books on 11 February 2010, and we cite that rather than the paraphrase or the TED talk we could not verify.
• "90 percent of online content will be AI-generated by 2026." Nobody measured this. It began as a personal forecast in a 2020 book about deepfakes, a Europol report cited the author in 2022, and the internet turned a citation of one person's guess into an apparent finding by a police agency.
• "Goldman Sachs says AI will destroy 300 million jobs." The March 2023 report says exposed, which means some tasks overlap with what a tool can do. The same report put full substitution at about 7 percent of US employment and complementarity at about 63 percent.
• "AI destroyed 204,000 entertainment jobs." CVL Economics asked 300 executives in late 2023 what they expected by 2026, and published in January 2024. A survey of expectations is not a count of losses.
• Any layoff-tracker count of creative jobs lost to AI. Those tally the reason an employer chose to state in its own announcement. That is self-attribution, not measured cause.
• Any pilot employment number used to prove that automation grows jobs. We checked, and the two US Bureau of Labor Statistics occupation codes do not agree with each other, so we dropped the claim rather than pick the flattering one.

Common questions

Which steps of a brand deal can AI actually do?

The ones with a right answer somebody could have described before the work started. Writing the first plan, building a shortlist, drafting the contract, holding the money safely, chasing an approval and reading every comment afterwards are all that kind of step. Deciding which idea suits this brand, judging whether a cut is good, and being the person a client calls at nine in the evening are not.

Is there a rule for which work gets automated?

Yes, and it is older than generative AI. David Autor, Frank Levy and Richard Murnane published it in the Quarterly Journal of Economics in November 2003: computers take over tasks that can be written down as explicit rules, and they increase the amount of work in tasks that cannot. It is not hard against easy. Chess is hard and a machine plays it. Knowing which of two photographs makes someone cry is easy for you and a machine cannot do it.

Does this mean AI replaces agencies?

It takes the paperwork, not the judgement, and the difference decides who does well. David Autor wrote the same distinction about bank tellers in the Journal of Economic Perspectives in 2015: "a bank teller who can tally currency but cannot provide relationship banking is unlikely to fare well at a modern bank." An agency whose product is processing plans and assembling reports is doing the counting. An agency whose product is knowing which idea is right for a particular client is doing the relationship.

Has automation ever wiped out a creative job completely?

Yes, and pretending otherwise would be dishonest. The US Bureau of Labor Statistics reported in September 2025 that one hour photofinishing fell from 17,320 jobs in 2000 to 437 in 2024, a drop of 97.5 percent. Developing the picture was destroyed. Taking the picture was not. The rule held: the step with a specified correct output went, and the step needing judgement stayed.

So what is left for people?

Three things, and they are the same three every time. Taste, which means knowing which option is the good one. Relationship, which means being the person somebody trusts or watches. Accountability, which means carrying the responsibility when it goes wrong. None of the three can be written down in advance, which is exactly why no tool has taken them.

Is there real proof that people and machines do better together?

There is. In 2011 a group at the University of Washington could not work out the shape of one enzyme, and neither could any automated method, after more than a decade of trying. They handed the problem to players of a puzzle game called Foldit and gave them three weeks. The players cracked it, and the paper in Nature Structural and Molecular Biology in 2011 credits two teams of gamers as authors. The computers did the arithmetic. The people did the seeing.