Stop hiring for operations. Redesign the company.

There are entrepreneurs who use AI. There are companies that ship AI features. And there is a much smaller group asking a different question: what happens if you stop adding AI to a company — and redesign the company around it?

Valeriy Sterligov ran that experiment on his own operating business, a nationwide auto glass and insurance-claims company. By his account, as the operating model was rebuilt, the operations team shrank by roughly 97%, and most of the work that used to pass between people is now done by software and AI.

The number is striking. What he took from it is more interesting: the hard part of AI is no longer how smart the model is. It is the system you build around it.

Context · Memory · Tools · Permissions · Verification · Accountability · Several independent models · A human at the critical points

In autumn 2026 he took part in Anthropic's AI Interviewer study, a research conversation about the role AI plays in people's lives. The conversation had an unusual extra layer: while one AI asked the questions, another — ChatGPT, which has worked with him for a long time — helped him put his answers into English. One AI interviewing a man about his relationship with AI, while a second AI helped him explain it.

What did you last use AI for?

I was building an AI-based business automation system. I rarely use AI just to ask it something anymore. It works more like a technical co-founder or a systems architect: it helps me understand an existing system, find the weak points, design the changes and gradually automate work that used to take several people.

At some point it stopped feeling like using a tool. It started feeling like working together.

What moment stands out most?

The moment I saw that AI could understand a real production system. Not a demo. A live business: source code, processes, databases, documents, access rights, integrations, change history. Instead of generic advice, it traced dependencies, separated symptoms from root causes and proposed concrete changes.

I used to think the question was how smart the model will get. Now the question is what environment you have to build around it before you can actually rely on it.

Context. Memory. Tools. Clear limits on authority. Verification. Escalation to a human. The ability to roll back. And several specialized AIs instead of one universal brain. That's when I understood this isn't a better way to work. It's a different way to build companies.

The experiment

So I stopped theorizing and tested it on my own company. Not a small AI pilot. Not a chatbot. A gradual rebuild of the operating model itself. The question was: how far can you go if AI doesn't help an employee do the work — if the work is designed from the start to be done by a machine?

Over time, headcount fell by roughly 97%. But I wouldn't call it replacing people with AI. That's too crude. We started eliminating the need for many of the old operations altogether.

A single piece of information used to pass through several people. One receives it. Another classifies it. A third moves it between systems. A fourth checks the document. A fifth tracks the status. A sixth writes the reply. Then you ask: why does this process look like this at all? When you design around machine intelligence from scratch, a huge amount of human work simply stops existing in its old form.

"I stopped asking who we need to hire"

It changed me as an entrepreneur. When an operational problem came up, the natural question used to be: who do we hire to handle this? Now the first question is: can we change the process so that no person is needed for this step at all?

People stay where they are truly needed: relationships, accountability, complex judgment, exceptions. But a person should no longer work as a router between two pieces of software.

Convincingly wrong

Has AI ever made things worse?

Of course. Those cases taught me as much as the successes. One of the most dangerous traits of current models is that they can build a very convincing argument on an incomplete picture.

Once, an AI misidentified which layer of a running application was the source of truth and proposed editing an already-built production artifact instead of the source. Another time, a tiny patch escaped a string incorrectly. A microscopic error — and the production bundle stopped working.

What struck me wasn't the mistake. It was how logical all the reasoning around it looked. That's where the rule came from: confidence is not evidence.

Look first. Then think. Then change. My protocol now: read-only investigation of the real system first. Then establish what we know, what we don't, where the source of truth is and what depends on what. Then a backup. The smallest possible change. Build. Tests. Smoke check. A rollback plan. Only then the next step.

Intelligence without verified context can be operationally dangerous.

Where delegation ends

I'm comfortable letting AI analyze, prepare decisions, compare options, criticize my ideas and take many reversible actions. But delegating work and delegating responsibility are different things.

If a decision can irreversibly affect someone's rights, create a large financial obligation, have legal consequences or seriously affect human relationships, I want a person at the final point. Not because humans always reason better than machines — sometimes it's the opposite. The questions are: who is accountable if it's wrong, and can it be undone?

As systems become more reliable, that line will move. But automating intelligence shouldn't mean automating away accountability.

Two AIs that know the same person

Several AI systems have now worked with me for a long time. They know me not because I once told them who I am, but because they've seen thousands of my decisions: how I argue, when I change my mind, what I call a mistake, which ideas I reject, at which stage I start to doubt.

Sometimes I use one AI to help me put a thought into words for another. This interview is an example. The substance is my own experience; ChatGPT helped me translate and structure it, because over a long history it already understands much of my working context.

With a long enough history, a system starts to see patterns that are hard to notice from inside your own life: where you repeat the same scenario, where your words and actions diverge, where you're unusually effective, which decisions keep leading to the same mistake. That's not quite an assistant anymore. I'd call it cognitive infrastructure around a person. Part memory, part analyst, part critic, part partner.

If AI understands me better tomorrow than it does today, I want that understanding to give me more options, not fewer.

A system should never decide, "I've figured out who Valeriy is." People change. Good AI memory should be able to say: this is a fact; this is a recurring pattern; this is a hypothesis; here is contradicting evidence; this may be out of date. Knowing someone well and being overconfident about them are not the same thing.

Several minds instead of one

I don't think the future is one giant AI that does everything. In my own work I already see the value of a different architecture. One model proposes a solution. Another solves the same problem independently. They critique each other. A third loop compares the arguments. A human steps in where there's real uncertainty or a high cost of error. Different models have different blind spots, so competition between them is sometimes worth more than a few extra points of performance from a single model.

What AI should change

A huge part of human life goes not to real complexity, but to the consequences of badly designed systems. Copying information from one window to another. Finding a document. Explaining your context again. Filling in another form. Reminding someone of the next step. Checking a status. Approving something a machine could have approved on its own.

I'd like AI to give that time back — in business, government, education, medicine. Not by making people unnecessary, but by removing work that exists only because our organizations and software can't understand context.

At the same time, not all difficulty should be removed. Some friction is just waste. Some of it forms a person: judgment, character, craft, taste, responsibility, courage. I want AI to remove pointless friction, not meaningful struggle — to expand people's ability to act, not quietly replace it.

What should Anthropic build next?

The next big leap is around the model, not only inside it. Long-term context. Real memory. Tools. Clear permissions. Working for months toward one goal without losing the thread. Understanding its own uncertainty. Verifiable conclusions. Several independent intelligences working together.

I want AI to tell me not just "here's my answer," but: here are the facts I saw. Here's what I concluded. Here's how confident I am. Here's what contradicts it. Here's what I don't know yet. Here's what would change my mind. That kind of AI isn't just good for conversation. It's fit to run real systems.

The next conversation

For a business owner, "How do I use AI in my company?" is now too small a question. The better one is: "If I built this company from scratch today, knowing machine intelligence exists, would it look the same at all?"

How many departments would remain? How many approvals, interfaces, managers, manual checks? How many people are doing work today only because two parts of your company can't understand each other?

Sterligov is already running that experiment. Which is why a conversation with him rarely starts with "show us your AI assistant." It starts with: "Here's my company. If you were building it today, what would you keep?"