The man who automated other people's work says the automation work is next

Nick Saraev runs an AI automation agency and says he has earned close to $10 million from the work. On 5 October he published a video titled "Here's What I'd Learn Instead of AI Automation in 2027", and by this weekend it had 278,000 views. His argument: the service he sells, building automations for companies, is itself being automated, and anyone in it has about a year to move up a level.

He has an incentive to say the opposite. He opens by naming it: automation specialists, software sellers and content creators all profit from the idea that the work lasts. He also sells a course, so he has an incentive to say this too. Read the argument on its evidence.

Two reasons the building work is losing its value

The first is the benchmarks. Saraev points to AutomationBench, Zapier's test of end-to-end execution across core business functions, and to OpenAI's GDPval. Google reported on 30 September that its Gemini 4 Argon model ranks first on AutomationBench with 51.3%. By Saraev's account the benchmark went from zero to that score in under six months, and he expects it to be saturated by early 2027. His own description of the change: a year ago an agent had to be walked through API documentation and connectors; now it takes a one-line request and does, in his words, "a pretty good job".

The second is the knowledge gap. Saraev's business was arbitrage: he knew what AI could do and his customers did not, and he was paid for the difference. He says that gap is closing, and that it will close fast once business owners who have had their heads down see what agents now do. The analogy he reaches for is website building. The skills that once justified large fees became a hosted service and a prompt, and the designers who still earn well are the ones who are extraordinary at selling, with ordinary HTML underneath.

AI does the routine work. You do the thinking. Saraev's version of that sentence is that building the system is becoming the routine work. What stays is deciding what to build, getting a team to use it, and selling the result.

What he says to do instead

His prediction for the first half of 2027 is a surge of business owners asking for help, and for a human rather than a machine. The help they want changes shape, from implementation to direction. He lists the moves he is making:

• From building to teaching. He now runs workshops that train a team to use Claude, set up its shared skills, and give each member a weekly schedule to keep their level even. He says his cohort's own market tracking shows demand for drag-and-drop automation tools falling against agent coding tools, consulting and training.
• Sell the outcome. His sharpest line: time spent improving technical skill should be matched by a hundred times as much on sales, marketing and delivery, because everyone will soon be able to simulate technical skill with an agent. The questions that pay are about the funnel, the pitch and the follow-up.
• Higher-bandwidth outputs. Text and simple images are now cheap. He expects the market to pay for video, three-dimensional and simulation outputs, the things a prompt does not yet produce well.
• Short-lived opportunities. A faster market has more openings and each one closes sooner, which rewards people who can spot and act early.

He is explicit about the limit of his forecast. After 2027 he says he does not know, and he distrusts anyone who claims to.

Two of his numbers are his own estimates: the share of people who have never used a paid AI model, and the pace at which the benchmark will saturate. Treat them as a practitioner's reading, with the Google figure as the one hard number.

Your Next Move

• Move one level up in your own work. List what you produce, then ask which items an agent now does from a one-line request. Those are the implementation layer. The judgement above it, which problem to solve, what counts as done and who signs off, is where to spend your next quarter. The afternoon method for sorting a week of your tasks does this in detail.
• Teach what you know. If people ask you how you use AI, that is a service. Turn one answer into a one-hour session for your team, and keep a record of who could do what before and after. That record is the evidence an employer now grades.
• Count how you spend improvement time. For a month, log hours on tools against hours on how you explain, sell and deliver your work. Saraev's ratio is extreme, but the direction holds.