Where you are, and where you'll be
You're at A: you bought AI the way you bought software, by the seat, meaning a per-person license. Your people still do the job, and they still check every output. So you pay twice, once for the seat and again in staff hours. By the end, you'll be at B: you'll be able to write one recurring revenue task as a finished piece of work with a finish line and a check, which is the spec you'd hand to a vendor or an AI agent (AI that works through a multi-step job on its own).
The one idea
When you buy or build AI for a revenue team, scope it as finished work with a pass/fail check, not as a tool seat. Do it this way because the value sits in the work, not the tool. A seat measures access. A finished deliverable measures output.
Why the second bill is the big one
Issue 1 did the math. A sales manager costs about $106 an hour, fully loaded, based on Bureau of Labor Statistics figures for sales managers. Each hour a week that manager spends pasting, reading, and correcting AI output costs about $5,512 a year. The seat shows up on an invoice. The hours don't, and that's why the hours rarely get cut.
Angle 1: the principle
Sequoia's Julien Bek put it in one line in March: "A copilot sells the tool. An autopilot sells the work." A copilot is AI that helps a person do a task. An autopilot does the task and hands back the result. He adds that "for every dollar spent on software, six are spent on services," a Sequoia estimate with no stated source.
The idea isn't new. In 2024, Sequoia's Sonya Huang and Pat Grady wrote: "Cloud companies sold software ($ / seat). AI companies sell work ($ / outcome)." Most of that conversation is about how AI companies should price. Our angle is the buyer's side. You can scope your own AI purchases and builds this way today, one task at a time, whatever your vendor charges.
Angle 2: an example
Bek's example: "A company might spend $10K a year for QuickBooks and $120K on an accountant to close the books." That's Sequoia's illustration, not survey data. The company doesn't want QuickBooks. It wants closed books.
The revenue-team version: you don't want a CRM seat, you want a clean pipeline every Monday. The seat is the tool. The cleaning is the work you value.
Angle 3: a number
We checked Sequoia's ratio against official data. The US Census Bureau's Quarterly Services Survey shows that in Q2 2026, professional, scientific, and technical services firms brought in about $816 billion. Software publishers brought in about $177 billion. By our math, that's about 4.6 times as much, and about 6.4 times if you add administrative and support services. Call it roughly 4 to 6 times, depending on what you count. Two caveats: these are sellers' revenues, not buyers' spending, and the ratio is ours, not the Census Bureau's. And software revenue is still growing faster, 14.7% versus 6.4% a year, so the money hasn't visibly moved yet. The point here is where the value sits today, not where it's flowing.
Three angles, one point. If what you value is the finished result, pay for it and measure it as a result, because a seat tells you who has access, not what got done.
Granted, autopilot isn't here yet
Granted, most companies aren't buying finished work from AI, and AI can't yet deliver most of it alone. A Census Bureau working paper from April 2026 found that "most users (66%) rely on AI solely to augment tasks," and that AI-related job cuts happen "in only 2% of firms." (It's a working paper, not yet through the Bureau's full review.) In TheAgentCompany, a peer-reviewed benchmark presented at NeurIPS 2025, the best model fully finished 30.3% of 175 simulated office tasks on its own. Newer models may do better, but that's the latest peer-reviewed figure we have.
However, that's the reason to scope the work with a check, not a reason to wait. You don't need full autopilot. You need a finish line, a checker, and one person who approves, which is the loop from Issue 1. If the AI finishes, the check passes and you approve. If it doesn't, the check fails and shows you where to step in. Either way, you judge against a standard instead of proofreading. Start with one task, not the whole job.
The system we ran: this issue
On October 2, I set up a team of AI agents to run this newsletter, with research, curation, writing, and editing each owned by one agent. None of them gets "access to a tool." Each owes a finished deliverable that has to pass a check.
Research delivers a source file where every figure was read on the primary page. For this issue, it pulled and read 9 sources and flagged 3 Sequoia figures that came with no stated source. That flag is why the six-to-one ratio above is labeled an estimate and checked against Census data.
The editor delivers a brief with the claim, the evidence, and the objection the issue has to answer, because the writer can't hit a finish line nobody wrote down.
The writer delivers a draft that passes a 14-item checklist, including that every number names its source. A draft that fails goes back.
I approve. I don't write or fact-check. I read finished work and say yes or no.
Every handoff is finished work with pass/fail checks, and the only person in the loop approves. That's selling the work, inside our own company.
Copy this: write one task as finished work
Pick one weekly revenue task, like the Monday pipeline summary. Write three things.
The finished output. For example: "A one-page pipeline summary in leadership's inbox by 9 AM Monday." Be this specific because a check can only test a concrete result, and "help with pipeline" can never fail.
Three yes/no checks. For example: "Does every deal over your deal-size cutoff have a dated next step?" Use yes/no because a pass/fail answer is what turns a draft into finished work.
Who approves it. Name one person, because if nobody owns the yes, everyone goes back to checking.
That's the spec. Hand it to a vendor and ask them to price the output, or give it to an agent running Issue 1's loop. Either way, you're buying the work.
Recap
A seat buys access, and you still pay your people to do and check the work.
Sequoia estimates six dollars in services for every dollar of software, and Census data puts it at roughly 4 to 6 times.
Most AI still helps people rather than finishing work alone, which is exactly why you scope the work with a check.
You're at B if you can now write one revenue task as a finished deliverable with three pass/fail checks.
Scoreboard
AI agents on the content team: 5
Sources the research agent pulled and read for this issue: 9
Sequoia figures it flagged as having no stated source: 3
Items on the pre-publish checklist: 14
Saturday: why your AI starts ignoring your instructions halfway through a long job, and the one check we run to catch it.
P.S. If you'd like a task like this scoped and running for your revenue team, just reply to this email and tell me which one eats your Fridays.