AIで成功するのはツールでなくサービス企業

Sequoia partner Julien Bek recently published the thesis I have been betting on since 2022. He calls it "Services: The New Software." The argument in one line: the next trillion-dollar company will sell the work, not the tool, because for every dollar of software spend there are six or more dollars of services spend sitting next to it, and AI just turned that services budget into something a startup can actually attack. I laughed when I read it, because we have been building that company since 2022. ColdIQ. Seven million in ARR. Four hundred B2B clients. Thirty-plus people across ten countries and four continents. Bootstrapped. I left an eighty-thousand-dollar operations job at Worldcoin, the iris-scanning company Sam Altman co-founded. My first client paid three thousand a month. We (@MichLieben and I) ran the campaigns ourselves for a year and wrote down everything we learned. This article is the operating manual I wish someone handed us when we started. Two readers who can take the most away: The operator sitting on an idea they have never acted on. The founder plateaued between a small consultancy and a real business. If that is you, keep reading. Even if you don't fit in the above two buckets, you can still take away lots of scaling and business advice from here. The math, in one picture Bek uses a simple example that makes the whole thesis obvious. A company spends around ten thousand a year on QuickBooks and a hundred and twenty thousand on an accountant who actually closes the books. That is roughly one to twelve in that specific vertical. Most services categories run closer to one to six. Every software founder is chasing the QuickBooks budget, and the real money is sitting inside the accountant budget. Cold outbound has the exact same shape. A B2B company pays a few seats of sales tooling, a single SDR salary, and a meaningful multiple of both to an agency when they hire one. Software is a small line on that budget. The work is the expensive one. I did not need a Sequoia spreadsheet in 2022 to see this. I needed my first client to say yes. Copilot versus autopilot, and why only one survives the next model release Bek draws a line between two models that look similar on the surface and behave completely differently underneath. A copilot puts AI in the hand of a professional. The professional takes the output, carries the risk, and owns the client relationship. Harvey sells to law firms that way. Rogo sells to investment banks. An autopilot skips the professional entirely and sells the outcome directly to the buyer who actually wants it. Crosby ships the NDA to the company that needs one. WithCoverage closes the insurance policy for the CFO. ColdIQ books the meeting. The customer is buying the thing they came for, and the tool is invisible underneath. Every AI-tool founder I know is asking the same question right now: What happens when the next model update turns my product into a feature? It is the right question to ask. A company selling the tool is in a permanent race against the model and has to win by staying upstream of it forever. A company selling the work improves every time the model does, because the cost of delivery drops while the price to the customer stays put. The margin widens. The data moat deepens. That is the whole difference. One model compresses your business. The other compounds it. Worldcoin to $7M ARR: My Journey In 2022 I was doing operations at Worldcoin, Sam Altman's iris-scanning project. The job paid eighty thousand a year and came with the kind of resume line that opens other doors. Then Mich (@MichLieben) told me one day that I was his biggest potential competitor if I stayed where I was. That was the nicest insult I had ever received. A month later I quit. The first client at ColdIQ paid three thousand a month. The early campaigns were run by hand, and every campaign that worked or failed got written up the same day it happened. That discipline of documentation is what eventually turned a services company into a software company. You run the work manually until you have seen every edge case the market can throw at you, and then you encode the pattern, automate it, and price what comes out of the pipeline like a product instead of labor. By month thirty-one we had crossed six million in ARR. The company today looks like four hundred clients, thirty-plus employees across ten countries, and $7M+ in ARR. Somewhere in the last stretch I stepped out of daily delivery entirely, and the business has been better since. I was the ceiling. None of that required being smart. It required being early, being wrong in public a hundred times, and being stubborn enough to refuse the lazy version of "agency" that just means a body shop charging for activity. The playbook Here is how I would do it today if I had to start again from zero. Six steps. The order matters more than anything else on the page. 1. Pick one outsourced line item inside one industry. Go narrow. The narrower the slice, the faster you accumulate proprietary data, and the data is the moat. Three questions help you evaluate a niche: Is the work already outsourced to someone today, so that you are swapping into an existing budget line rather than inventing one? Is the work mostly intelligence work, meaning pattern recognition and rule application, rather than genuine strategic judgement that only a human can hold? Is the services spend in that category meaningfully larger than the software spend? Cold outbound cleared all three. That is why it worked. 2. Land your first clients yourself. You do not need a website, a deck, or a funnel to start. You need a short list of paying customers who will tell you what they hated about the last vendor they hired, so you learn which part of the work is worth productizing later. Charge a retainer floor you want to be anchored to in three years. ColdIQ opened at three thousand a month because anything below that left us unable to actually deliver. Sell the outcome your client actually wants, not the activity you happen to generate. Record and transcribe every sales call. The objections you hear in the first ten calls become the copy on the sales page you publish after the first ten clients. If a prospect ghosts you twice, move on. You have a targeting problem, and no amount of follow-up will fix it. 3. Do the work by hand for as long as it takes, and document every step. I know you want to build the tool first. Please resist that. The data you accumulate while running the work manually is worth far more than any MVP you could ship in the same window. At ColdIQ the early campaigns were all run manually, and that period is the training set that let us run more than two thousand two hundred campaigns across hundreds of clients in the years that followed. Four artifacts to produce from day one: One markdown file per repeatable task, written as if you are handing it to someone who starts on Monday. A Loom any time the work involves a cursor. One decision log per client with dated notes on why you made the calls you made. One file of failed campaigns and the reasons they failed. That last file becomes the most valuable artifact you produce in your first year. 4. Price the work like a service, and report on it like a product. Your clients do not care how the work gets done. They care about the outcome, and they care about being able to watch what is happening. Sell the outcome, and build dashboards, telemetry, and reports the way a SaaS company would build them. One pricing shape that works: an upfront setup fee to cover onboarding, a monthly retainer tied to an outcome metric such as meetings booked or closes delivered, and a performance bonus on top when you exceed the target. One reporting cadence that works: a live dashboard for every client from day one, a weekly update from the account owner covering wins and misses and next week's plan, and a quarterly conversation with the decision-maker rather than only the operator. Retention in a services business starts looking like SaaS retention when the client experience looks like SaaS too. 5. Replace yourself on delivery before you scale anything else. Every agency founder falls into the same trap. You keep doing the work personally because nobody else does it quite as well. You are probably right that they cannot. You are also the ceiling. Hire in this order: First a delivery operator who can run multiple accounts alone after a short shadowing period. That role is pure execution and does not require a founder's judgement. Second a technical automator, fluent in whatever stack you run, whose job is to turn the markdown workflow files into running agents. Third a head of delivery who manages the first two and owns quality across the book. Avoid hiring a marketer, a salesperson, or a COO until the delivery layer runs without you. Every early hire outside of delivery is a rounding error at best and a burn-rate accelerator at worst. When I stepped out of the last piece of delivery work that still needed me, the business moved forward the next month. 6. Compound the data moat before you compound the software. Bek's convergence point is the whole long game. Judgement turns into intelligence as your dataset grows, and three years of running the work gives you an asset no pure-SaaS competitor can buy, license, or ship around. Save every third-party input you pull, in both raw and cleaned form. Save every output you delivered to a client, tagged with the outcome it produced. Save the reasoning your team used when they made judgement calls over the top of the system, because the why matters more than the decision once you start training on it. Save every client objection and the response that closed it. Give this process enough time and you end up with a data asset your competitors cannot bootstrap into existence. Eventually the agents run delivery while the humans run judgement. That transition happens quietly in the background for years before anyone outside the company notices. The accelerator Two years into ColdIQ I started getting a steady stream of messages from operators who wanted to build something like this. Some were still at real jobs sitting on real ideas. Others were running small agencies plateaued somewhere in the middle. A few were asking whether the math still worked in 2026. I turned the answer into a program. The ColdIQ Accelerator now has more than two hundred and eighty-seven members, and the same model has helped one hundred and seventy-three agencies add somewhere between five and fifty thousand in monthly revenue. Members get the agents we use internally, the list builder we built for our own delivery, the template library behind more than two thousand two hundred campaigns, and the same playbook that took ColdIQ from zero to three hundred fifty thousand a month. The program is the operating stack handed over to someone who wants to do what we did without the years of false starts underneath it. The pattern I see in the members who win is always the same: They pick one outsourced, intelligence-heavy niche, and they run the work by hand until it teaches them what to automate. If you recognized yourself in the description at the top of this piece, check out our Accelerator program. Sequoia can name a category. Founders can raise on a category. The head start still belongs to whoever is willing to do the work by hand until the work tells them what to automate. The companies that look like services firms and earn like software firms are never accidents. They are what happens when one person decides the agency model is lazy, and sets out to prove it. Start before the category closes.