Most owners have used AI. Almost none have watched it do a week of work.

On October 11, Andrej Karpathy, a founding member of OpenAI and former Director of AI at Tesla, reposted something he wrote in April and said the gap it described is getting wider. In April he described two groups of people talking past each other about AI. This time he drew a funnel, and a steep one.

His numbers come with his own label, napkin math. About 6 billion people, roughly three quarters of everyone alive, have barely touched an AI chatbot like ChatGPT. Somewhere between 1 and 2 billion use a free one now and then and, in his words, treat it "a bit like a better Google search, a writing aid". About 20 million, around 0.2%, use the most capable versions professionally in math and code, and have seen "large, complex projects that used to take them weeks/months" get "completed by agents with a prompt". An agent here means AI software that works through a multi-step task on its own: it reads the email, opens the file, drafts the reply. And roughly 5,000 people have internal access to frontier systems inside the AI labs themselves.

Nobody can audit those figures, him included. What I take from them is that each layer is having a different experience of the same technology, and each has fair reasons for believing what it believes.

He also explains why the layers exist. The first is ambition: "You need a reservoir of big, difficult problems that you really care about." People who ask small questions see small results. The second is that the capability is jagged: it "peaks in domains that are digital, verifiable and economically valuable", which is why coders and mathematicians saw it first. And the tier of access matters, because the free version, the paid version and what runs inside a lab are not the same product.

Where small firms actually sit

What follows is my read, not Karpathy's. He was describing programmers, researchers and the general public. He said nothing about law firms, insurance agencies or freight brokers.

Still, when I talk with people who own or run firms of 10 to 300 people, the pattern is consistent. Almost all of them have tried ChatGPT. Most use it the way you would use a smarter search box: draft an email, summarize a PDF, settle an argument about a policy wording. Very few have watched it take a week of their team's real work and do it.

The survey data points the same way, with numbers that swing depending on what counts as "using AI":

The Census Bureau's Business Trends and Outlook Survey puts AI use at about 17% to 20% of US businesses, 19.8% as of early May 2026. Use rose among firms with 20 or more employees and did not change much for firms under 20. Finance and insurance sits at 33.9%.
Goldman Sachs' 10,000 Small Businesses Voices survey of 1,256 owners, fielded early this year and reported by Fortune, found more than three quarters use AI to some degree. Only 14% have it embedded in core operations, and more than 70% said they would benefit from more training and implementation help. The owners surveyed come from Goldman's own small-business program, so the sample probably skews toward more engaged owners than average.
Intuit's QuickBooks 2026 AI Impact Report, published in May, found regular use among US owners climbed to 77%, yet only about 1 in 10 businesses pay for dedicated AI tools. The barriers owners named were privacy and security, fear of errors, and uncertainty about what AI can do. Cost was not at the top.

In the Goldman survey, the 76% and the 14% come from the same 1,256 owners, and those are owners engaged enough to be in a business program. Even in that group, about five in six who use AI have not put it into the work that runs the business. The reasons they give are sensible ones. They are not sure it can be trusted with work that matters, and nobody has tested it on theirs.

Nobody in that 76% doubts AI exists. The open question for an owner is whether it can take the weeks.

What I've watched agents do with a prompt

Before you weigh anything I say here, you should know where I'm coming from. I'm not an engineer. I direct AI systems. My judgment comes from years of running operations. As interim COO of Nussbaum APC, a law firm, I handled billing, AR, AP, accounting and payroll for a staff of 10 and a book of more than 1,000 clients. That is the seat most readers of this page sit in, and it is why I care more about the Tuesday afternoon backlog than about benchmarks.

I founded Agor AI in late 2025. Here is what I have shipped working with agents:

Four iOS apps are live on the App Store, all with in-app purchase, shipped solo.
Dialogues of the Machine, a book of 25 dialogues: 824 pages, about 313,000 words, directed and shipped.

Across all my projects, that adds up to about 3,500 commits of my own work directing AI systems.

For a business owner, the part to pay attention to is how I run my own agents. I run my own fleet of agents in production, with governance gates. In plain terms, they start switched off, draft first, and only act once a person arms them. There is a kill switch that stops them all, and each one can be set back to OFF. Each one sits on a ladder. OFF means it does nothing. DRAFT means it does the work and a person approves it before anything leaves the building. LIVE comes only after its drafts have held up. And when an agent hits something it isn't sure about, the rule is to hand it to a person instead of acting.

I run these on my own company first. That is where I learned which jobs agents handle well and which ones need a person watching every step.

By Karpathy's description, that puts me in his third group: someone who has seen weeks of work done from a prompt. I'm not in the fourth. I don't work inside a lab, and I don't have that view. What I have is the view from a small company that runs its own agents this way, which is closer to yours.

Where it breaks

Here are the limits, as plainly as I can put them.

I have no clients yet. Everything I've described is my own systems. I'm not going to dress it up as anything else, and you should weigh it accordingly.
Agents are strongest where the result can be checked. Matching invoices against a ledger is checkable. Replying to an upset customer with the right mix of firmness and apology is much less so. Karpathy's point about jagged capability is the most practical thing in his post for an owner: pick work where a person can tell at a glance whether the output is right.
Review falls behind output. Karpathy puts it bluntly: "human review and comprehension are starting to fall behind." Inside a firm, that is the real risk: an agent sending 200 emails that nobody read. It is why everything I build starts switched off and spends time in draft before it goes live.
Loosely specified work resists automation. In the April thread, Megan McArdle pushed back (my paraphrase): iteration works on well-defined problems and struggles with loosely specified human systems. She is right, and most small firms run on loosely specified human systems. "Dana knows how we handle that client" is a process, but nothing an agent can read. Finding out whether a workflow is specified enough, or what it would take to write it down, is the first thing I check.

Which of your team's weeks are ready

Pick one workflow, the one that eats the most hours in your week, and answer six questions about it.

Is it repetitive? Does it take the same basic shape most times it comes up?
Is it already written down? A checklist, a template, an email you keep reusing. Badly written down still counts.
Can someone check the output quickly? If a person can tell in a minute whether a draft is right, it is a good sign: an agent can draft and a person can check.
Is it digital end to end? Email, PDFs, portals, spreadsheets. Paper and phone calls are harder, though not always impossible.
Is there one clear owner? Someone who would approve the drafts and be accountable for them.
Is it costing you a hire? Overtime, a backlog, or a job posting you keep putting off.

Four or more yeses and it is a real candidate. Two or fewer, and I would leave it alone for now, or start by writing the process down.

Where I would expect a lot of yeses, by kind of firm:

Law: new-matter intake forms and follow-up, and billing follow-up on overdue invoices.
Insurance: new-client intake, and chasing clients for renewal documents.
Accounting: chasing clients for documents and organizers.
Property management: vendor invoices coming in and getting coded for payment.
Agencies: invoicing clients and following up on what is overdue.
Logistics: carrier invoices coming in, and billing follow-up.

These are the hours that fill a team's week at most firms this size. They mostly live in email and PDFs, they are done the same way each time, and a person can check them quickly.

The smallest safe step

If your workflow came out as a candidate, this is what I sell. It is called the AI Agent Architecture Review.

$2,000, five business days, one workflow you pick.
You get a map of that workflow: which steps an agent can take, which stay with a person, and the safeguards around each handoff (what runs in draft, who approves, what shuts it off).
You get a build plan with the costs and risks spelled out, and a fixed quote for the build. For the four builds I quote, intake, billing and AR, AP intake, and document chasing, it falls between $9,000 and $18,000.
If you sign the build within 30 days, the $2,000 is credited in full.

The Review does not include a demo. What you get is the map and the fixed price. If the answer for your workflow is "not yet", you will know why and what would have to change first, and you keep the map either way.

Karpathy's funnel describes who has seen the thing work. Most owners haven't, and the reason usually has little to do with intelligence or appetite. Nobody has pointed it at their work. The Review points it at one workflow of yours, with every step starting in draft and a person approving it.

Which workflow eats the most hours in your week? Start there. Book the AI Agent Architecture Review: $2,000, five business days, one workflow, credited in full to a build signed within 30 days.