What Should a Small Business Hand to AI, and What Should Stay With the Owner?

A ten-person startup runs most of its engineering through AI systems. Here is what a small Northeast Indiana business can borrow from that model, and what the owner should keep.

What Should a Small Business Hand to AI, and What Should Stay With the Owner?

A small business should hand AI repeatable work that follows clear, written rules, while the owner keeps direction, exceptions, and the decisions that shape the company. That division is the practical lesson from a recent episode of The AI Daily Brief, in which the host examined how Spawn, a ten-person game-creation startup, runs most of its engineering through a fleet of persistent AI systems. Although the scale is unusual, the underlying structure translates well beyond software and into the day-to-day operations of ordinary local businesses.

The most useful point in the episode is not the size of the fleet itself. The host argued that most companies describing themselves as AI-native still run the old company, only faster, so their outputs remain similar and their org charts would look familiar to a manager from twenty years ago.

Spawn took a different path, treating AI as a reason to dismantle the habits originally built for human workers. That choice changes which tasks you should hand off and how carefully you should watch them once they leave your hands.

Key Takeaways

  • Most AI transformation speeds up the existing company. Real change removes routines that were designed for human workers and no longer serve a purpose.
  • Spawn reported about 108 code changes an hour and a token bill roughly 30 times its payroll last month.
  • Spawn removed pull requests because its AI systems understand the whole codebase and do not need to narrate every individual change.
  • Two years of founder decisions became written principles that AI judges now apply, and the founder personally made 27 of more than 11,000 decisions in eight days.
  • People keep three jobs: taste, deletion of redundant work, and asking for the ambitious leap that AI systems rarely propose on their own.
  • A small business can copy the structure by writing its rules, keeping an approval step, logging its decisions, and keeping the owner focused on direction.

What does an AI-native company look like in practice?

An AI-native company organizes its work around systems that keep running over time, rather than around individual tasks that start, finish, and disappear without leaving much behind. The episode describes Spawn's fleet as persistent, and each AI system carries a name, its own memory, and ownership of one function, such as the game engine, bug hunting, or finance, which means the system is accountable for an outcome rather than a single request.

A persistent system accumulates context that a disposable helper, spun up for one job and then discarded, never gathers in the first place. It remembers what changed last week, which change caused a failure, and which decision a person approved. That accumulated memory allows work to carry forward without a human re-explaining the situation every morning. A shared and current record performs the same function for a human team, which is why the host treats that record as the backbone of the entire model.

Spawn's own numbers show how far the idea has been pushed in everyday practice, well beyond what most owners would ever attempt. The company spent $3.1 million on tokens last month, roughly 30 times its payroll, and the host says that spending makes Spawn one of Anthropic's largest customers.

During the same month, the fleet shipped 77,000 code changes, which works out to about 108 an hour. These figures come from the founder's account as reported on the show, so they should be read as one company's own report rather than as an independent industry benchmark.

Why should a small business care about a game studio's token bill at all? The answer is that the underlying pattern scales down considerably, even when the budget and the infrastructure do not. A landscaping crew, a dental office, or a contractor does not need thousands of helper systems running in parallel. It needs a small number of clearly defined responsibilities, such as who drafts estimates, who answers routine questions, who follows up on quotes that went quiet, and who approves the final price before it reaches a customer.

Why does removing old routines matter more than speeding them up?

Removing old routines matters more than speeding them up because accelerating a broken process only produces broken output sooner and in greater volume. The host's clearest example involves internal communication, the constant messages that keep people aligned on shared work. Spawn's AI systems originally spent much of their time announcing what they were about to change, and each announcement touched other pieces of work and therefore triggered further replies. The founder reported that doubling the number of systems roughly quadrupled the total volume of messages.

Human teams depend on that kind of chatter because a database specialist and a front-end designer each see a different part of the same product. The AI systems, by contrast, understood the whole system, so the constant announcements served very little purpose. Removing them did not represent a loss of coordination. It represented an honest recognition that the coordination had been engineered for a different kind of worker with different limitations.

A small business has its own version of this problem, and most owners will recognize it immediately. The owner who texts the entire crew about every schedule change, or who copies three employees on every customer email, is quietly paying a coordination tax on each message.

The remedy is not more messages sent more carefully. It is a shared and current record that shows what changed, who owns the next step, and what remains unresolved. The episode makes the same point in its principle about letting the work carry its own context.

Which human habits did Spawn remove, and what did it gain?

Spawn removed pull requests, the review step that human developers rely on to confirm that everyone understands a proposed code change before it is merged. Without that step, the AI systems work directly inside the codebase. When a change breaks something, the company spawns a fixer system to diagnose the problem and repair it. Automated test suites run every fifteen minutes, and the entire fleet must reach a consistent state once an hour.

The founder reported about 1,000 reverts out of 123,000 changes over the past year, which he presented as evidence that the approach is resilient. He also said the arrangement is durable enough that no human gets paged at two in the morning to fix a fatal bug, because the fleet keeps iterating on the code as problems surface. Spawn treats falling behind on shipping as an outage, a deliberately strict standard that it chose on purpose.

The price of this approach is compute, and it is substantial, since no small business will ever run that infrastructure. The episode cites 5,000 CPUs, 17 terabytes of RAM, and 150 GPUs running on AWS, plus another 60 percent of capacity held in reserve for surges. The lesson for an owner is narrower and more practical: removing a routine saves real time only when you can clearly see what replaces it. Spawn replaced its review step with scheduled tests, a fixer on call, and a visible shipping rhythm.

The bug loop offers a second example worth studying. Instead of humans relaying production problems to the AI systems by hand, players' in-game AI detects bugs and files reports automatically. A small model then triages each report and routes it to the right owner. A small business rarely operates a product of that kind, but it can still adopt the principle, which is to send each problem to the person or system able to fix it, with the relevant details already attached.

How did founder decisions become precedent that AI judges apply?

Spawn converted two years of founder decisions into written principles, then built AI judges to apply those principles, so that most routine questions no longer wait for the founder to answer them personally. The turning point arrived when the founder woke to 700 requests overnight. The systems were surfacing questions faster than he could possibly answer them, and many of those questions were ones he had already settled hundreds of times before.

The judges have three possible outcomes for any given request, depending on what the existing record shows. If a prior human decision covers the question, the judge applies the same outcome and records which human decision it followed. If the request does not make sense, the judge returns it to the originating system for revision.

If the situation is genuinely new, the judge escalates it to a person. In the first eight days of operation, the judges made more than 11,000 decisions, with about 7,300 resolved from human precedent, roughly 2,200 sent back for revision, and about 1,800 handed to a human.

The founder personally made 27 of those calls during that window. The host reports that daily shipped changes rose from about 1,000 to more than 7,000 once the judges scaled up. The precedent idea is the most portable piece of the entire model. Every owner already carries rules inside their head, such as when to discount, which jobs require a site visit, and when a complaint should reach the owner directly. Writing those rules down is the essential first step toward letting AI apply them consistently.

Keep a log of each decision, the reason behind it, and the case it covered, because that log keeps a teammate consistent when you are not present in the room. For guidance on what a small business should record, read what a small business should log about an AI teammate.

What work still belongs to people?

In this model, people keep three jobs, namely taste, deletion, and ambition, and none of them is easy to automate. Taste means judging whether a product or service feels right, including telling an AI system no even when its work technically functions. The host notes that the founder describes his engineers as "paratroopers" who drop into whatever situation needs them, become familiar with it quickly, and then unblock the systems working on it.

Deletion means removing work that a newer model has made redundant, which is harder than it sounds. In June, the founder scrapped six months of work covering 70 percent of the game engine, because a new model had rendered that effort unnecessary. Software teams rarely discard their own output on purpose, yet the episode presents this habit as one of the more valuable disciplines a team can practice.

Ambition means asking for the large leap that a system will not propose on its own initiative. The host reports that Spawn's AI systems build the next small increment every week, but they are poor at noticing when they have been heading in the wrong direction for days. The team's most significant recent build grew out of a human prompt asking how to make a world a thousand times the size of GTA load on a phone within one second.

A business owner can read this as a job description, and it is a demanding one. You keep the decisions about what to offer, what to stop doing, and what to attempt next. The AI handles routine execution within rules you wrote, and it flags the cases that genuinely require your judgment.

What else was in the October 30 episode?

Beyond the main thread, the episode covered several headlines that matter to owners who follow the broader economy. The New York Times reported that Weil, Gotshal & Manges, a 95-year-old law firm, filed for Chapter 11 protection. The host says departing partners, a failed succession plan, and a scuttled merger caused the collapse, not artificial intelligence. A legal AI startup named Crosby has submitted a letter of intent to buy some of the firm's assets, including its brand and internal data.

The host's interest in that bid centers on the archive. A firm's precedent documents, contract templates, and workflows can be worth far more to an AI buyer than the firm's equity, which often sits near zero once its senior partners depart. One lawyer cited in the episode noted that Google had bid $10 million for airline operational data, which is considerably less specialized than the knowledge embedded in a bankruptcy practice's files.

Earnings reports told a split story across the largest technology companies. Amazon Web Services grew 39 percent, its fastest pace in almost five years, yet Amazon's overall revenue and profit both missed forecasts. Intel's data center business grew 78 percent, but a softer forecast sent its shares down 23 percent after hours. The host argues that legacy businesses are subsidizing the AI build-out, and that any sustained slowdown in those core businesses would place greater pressure on AI financing.

The episode also reported on a New York Times story describing a Chinese intelligence concern involving Nvidia's chief executive. Nvidia disputes the claims, and the host characterized the reporting as thin in places. It is included here because it came from a major outlet and because the host was explicit about its limits.

What does an AI teammate do with this idea, and what stays with you?

An AI teammate handles the repeatable work you define, and it remains inside the access level you choose. Hey Button sets up each teammate around real workflows, beginning with an intake conversation about how work arrives, which tools you already use, and where busywork tends to pile up. Access comes in three levels: read-only, draft-and-approve, or full access, and you select the level that fits each workflow.

Our default is draft-and-approve, which means the teammate drafts a reply, a follow-up message, or a quote, and nothing is sent without your explicit approval. An illustrative example shows why that matters: a customer asks for a spring cleanup quote. The teammate organizes the request, identifies the missing details such as the address and photos, and drafts a reply for your review, after which you approve, edit, or reject it. Customer messages, spending, and sensitive changes remain behind the approval rules you set.

The host's principle about human judgment maps directly onto that setup, because direction remains a human responsibility that no rulebook can fully replace. Ownership of direction stays with you, while the teammate handles routine execution inside rules you wrote down ahead of time. An AI teammate will occasionally be a little wrong, so review and clear limits form part of the setup itself rather than an afterthought added later. For a closer look at how that review works in practice, see how review works before customers see AI work.

Where should a Northeast Indiana owner start?

Start with one workflow, write its rules down, and keep an approval step in place until the results are clear and consistent. Choose a task that repeats every week and has a clear correct answer most of the time.

For an HVAC company in Fort Wayne, that might be follow-up on estimates that have gone quiet. For a clinic in Auburn, it might be appointment reminders and intake forms, with any clinical question routed directly to staff. For a retail shop, it might be order status questions, and for an agency, it might be turning client requests into a shared task list.

Each workflow needs the same three components: the rules, the exceptions that always return to you, and a running record of decisions. Keep those three components visible to everyone who touches the workflow. Expect the first month to feel slower than you would like, because the Spawn team reworked its habits across many iterations before its system ran smoothly. A small business will face the same learning curve.

Measure what matters, such as response time and the share of drafts you approve without edits, rather than counting the number of messages sent.

This approach does not suit every business. If your work depends on in-person judgment with little written process behind it, begin with the decisions you already make repeatedly rather than the entire operation. Writing the rules is the difficult part, and it is worth completing before any tool touches a customer. Take the first step by starting the Hey Button questionnaire, which walks through how your work arrives and where it piles up.

Sources

  • The AI Daily Brief, "How the Most AI-Pilled Company in the World Works," October 30, 2026: https://aidailybrief.ai/e/2026-10-30. Figures used: Spawn's $3.1 million token spend and 30 times payroll; 77,000 changes last month, about 108 an hour; about 1,000 reverts out of 123,000 changes; 5,000 CPUs, 17 TB RAM, 150 GPUs, 60 percent surge capacity; 700 overnight requests; more than 11,000 judge decisions in eight days, split about 7,300, 2,200, and 1,800; 27 personal founder decisions; daily changes from about 1,000 to over 7,000; 70 percent of the engine scrapped in June; the Weil filing and Crosby letter of intent; AWS growth of 39 percent; Intel data center growth of 78 percent and a 23 percent share drop; a $10 million Google bid for Spirit Airlines data, as cited by a lawyer on the show.
  • Spawn founder Jacob Sanbury, "How I Built the Most AI-Pilled Company in the World," the post cited on the episode. Company-reported figures in this article come from that account as relayed on the show.

Frequently Asked Questions

An AI-native company organizes its work around AI systems that keep running, with defined ownership and shared memory, instead of adding AI to routines built for people. The episode's example, Spawn, removed pull requests and reworked its cadence. Speeding up an old process rarely changes the outcome much, which is the host's central point.
No, but software is where the pattern is easiest to see. The same logic applies to any business with repeatable work that follows written rules, such as quoting, scheduling, follow-up, and order questions. A plumber, a clinic, or a retail shop can apply the principle without building anything technical.
A decision log is a running record of each rule you applied, the reason behind it, and the case it covered. Keep entries short, one or two sentences each, and review the log monthly. Remove rules that no longer match how you work, so the teammate does not follow outdated instructions.
Usage charges can rise with volume, so ask for them in writing before you commit. Hey Button quotes setup and ongoing support around your workflows and integrations, and identifies any software subscriptions or usage charges up front, with no surprise line items.
Not at first. The default is draft-and-approve, where the teammate prepares replies and nothing goes out without your approval. Read-only and full access are the other levels. Move to more autonomy only after you have watched the work on real examples and the rules have held.
What is an AI-native company?
An AI-native company organizes its work around AI systems that keep running, with defined ownership and shared memory, instead of adding AI to routines built for people. The episode's example, Spawn, removed pull requests and reworked its cadence. Speeding up an old process rarely changes the outcome much, which is the host's central point.
Does this model only apply to software companies?
No, but software is where the pattern is easiest to see. The same logic applies to any business with repeatable work that follows written rules, such as quoting, scheduling, follow-up, and order questions. A plumber, a clinic, or a retail shop can apply the principle without building anything technical.
What is a decision log, and how long should it be?
A decision log is a running record of each rule you applied, the reason behind it, and the case it covered. Keep entries short, one or two sentences each, and review the log monthly. Remove rules that no longer match how you work, so the teammate does not follow outdated instructions.
Should I expect AI costs to climb as usage grows?
Usage charges can rise with volume, so ask for them in writing before you commit. Hey Button quotes setup and ongoing support around your workflows and integrations, and identifies any software subscriptions or usage charges up front, with no surprise line items.
Should an AI teammate send messages on its own?
Not at first. The default is draft-and-approve, where the teammate prepares replies and nothing goes out without your approval. Read-only and full access are the other levels. Move to more autonomy only after you have watched the work on real examples and the rules have held.

Hey Button

Want a teammate that handles this for you?

Answer a few questions about your business and Hey Button will show you what an AI teammate would take off your plate, with you approving anything that goes out.

Start the questionnaireSee what teammates do
Lucas M. Button

Written by Lucas M. Button

Founder, Hey Button

Lucas builds AI teammates for small businesses across Northeast Indiana and writes about what works, what doesn't, and what to hand off first. More about Lucas