Copilot or autopilot: where the money is and where the margin is

In March 2026 Julien Bek, a partner at Sequoia, published “Services: The New Software”, and here is his thesis in my own words.

Companies spend far more on services than on software, and the next large AI companies will not sell tools but the work that professionals do. The best place to start is work that is already outsourced and follows rules. And each time a better AI model comes out, that service gets cheaper to deliver.

Foundation Capital had already named the idea in 2024: Service-as-Software. And Sequoia later gave it a sharper line: “A copilot sells the tool. An autopilot sells the work.”

Deep down, the buyer does not want a tool. What do you think a sales director really wants, a lead scoring app? What they want is qualified leads on Monday morning.

A tool also has to be used, because someone has to learn it, feed it and check what it says. In most companies that someone does not exist, so the tool sits there, and an autopilot removes exactly that step.

And the budget for the work is already there. If a company pays an agency or a consultancy for a task, it has already decided the task is worth money, so there's no new budget line to invent.

The result has to be defined before you start. If you charge per qualified lead, both sides must agree on what a qualified lead is before the first one, because otherwise every invoice turns into an argument.

Someone has to answer for the result. A tool that fails is the user's problem, but an autopilot that fails is the seller's. And that means a named person accountable for it, insurance where it makes sense, and a signature where the law asks for one.

Running it also costs money, because each task pays for the AI models it uses and the tools it relies on. And those costs grow with every task, unlike most of the cost of software.

So the margin sits in between: an autopilot earns more than an agency and less than a software company. Better models lower that cost, but they do not remove it, which is why I would not sell anyone software margins on day one.

Here are five questions, and if a task gets five yeses, it is a good first autopilot.

Do you already pay someone else to do it? If so, the budget and the idea of “done” already exist.

Does it follow rules? Sequoia calls this intelligence, as opposed to judgement, and a grant shows the difference, because filling in the form is intelligence and deciding which grant to go for is judgement.

Can anyone check the result? A lead, a report or a checked case file can be checked, but “better positioning” cannot.

Is there a clear unit to charge for? It could be per task, per case or per qualified lead.

Is the liability low? If a mistake can end in court, you're better off starting with another task.

Before you buy another AI tool, look at what you already outsource, because the first autopilot is usually hiding in an invoice you pay every month.

Ask any provider who answers when something fails and what each task costs to run. And be careful, because if they cannot answer, chances are they are selling you a copilot with a new name.

That is how I work: first a diagnosis to pick the task, then the autopilot and, if we agree on it, a monthly report with the cost of each task.