Sidekick Orchestration

How to scale with AI: what the top 1% do differently

Most businesses now use AI. Very few scale with it. What the companies at the very top do differently, why their advantage grows every time a new AI model comes out, and how a small or mid-sized business can start building the same way.

// In this article

This is the follow-on to our guide on how to adopt AI at your company. That guide gets your team using AI well. This one is for owners whose team already uses AI and who want to know why the gains have levelled off, and what the companies pulling ahead are doing instead.

The 1% puzzle

Ramp, the US company that runs business cards and bill payments for tens of thousands of firms, can see what its customers spend on AI. In September 2026, Ramp's lead economist said about 80% of its customers' spending on OpenAI and Anthropic, the companies behind ChatGPT and Claude, came from just 1% of firms, a concentration he said Ramp sees in no other kind of software. Ramp repeated the figure in its October report. In September, that top 1% spent about US$9,600 per employee per month on AI. For a 20-person firm, that would be about US$190,000 a month.

Part of the explanation is simple. Many of the top spenders are venture-backed tech companies, often building products on top of these models. Apollo's chief economist found that about 80% of the top 1% are backed by venture capital, against about 13% of typical spenders. A home builder in Delta is never going to look like them on a spending chart. And top spenders are not proven to be top performers: no study yet tests decision records directly. The evidence for the habits below is one company's own detailed account, survey patterns from large companies, and our own practice.

The more useful part is why they spend that much. Nobody spends that much by typing more questions into a chat window. Many of them spend it because they have set up their businesses so AI can keep working around the clock without waiting for a person to decide what happens next. In our view, that setup explains more of the spending than the tools do.

One of those companies published how it works. Spawn is a 10-person company building a game-creation platform. Its founder says it spent US$3.1 million on AI in one month, which he says puts it among Anthropic's biggest customers. Since its founding in 2024, Spawn has organised itself around one question: how do we build a business that gets more out of every improvement in AI? Most of what follows comes from the founder's own account. His figures are self-reported and unaudited, and we flag where it matters. We first heard the story on The AI Daily Brief, which covered it on October 30, 2026.

Sources: Ramp AI Index, October 2026 (spending by Ramp's own US customers; the labs' total revenue may be spread differently). Apollo's Torsten Slok on who the top spenders are, as reported by FourWeekMBA. Jacob Sansbury, “How I built the most AI-pilled company in the world” (October 28, 2026). The AI Daily Brief, October 30, 2026.

Dabbling, adopting, scaling

Most companies we meet are somewhere on a three-rung ladder.

Figure 01Three ways companies use AI
  1. A few curious peopleDabbling

    Individuals try ChatGPT or Claude on their own. Nobody owns it, and results depend on who happens to be interested.

    Stops when the curious people get busy
  2. The whole team, with trainingAdopting

    The company picks tools, trains people and uses AI on everyday tasks. Each person gets faster, often by a lot.

    Every decision still waits for a person
  3. People plus AI that can actScaling

    The company keeps a record of how its decisions get made. AI handles the routine calls and brings people only the new ones.

    Grows with each new AI model
Most businesses sit on the first or second rung. The jump to the third is a change in how decisions are recorded and handed off. A better tool alone does not make it.

Dabbling is a few curious people trying ChatGPT or Claude on their own. Adopting is the work our adoption guide covers: an owner, leaders first, the right tools, training on everyday tasks. It is worth doing, and it pays. At the scaling stage, AI makes the routine decisions and people make the new ones.

McKinsey's 2026 survey of 1,719 respondents points the same way. Only about 6% are what it calls high performers, getting at least 5% of their operating profit from AI. Nearly three-quarters of them have fundamentally redesigned how their work gets done, against about a quarter of everyone else. They are more than three times as likely to be scaling AI agents across most of the business, and twice as likely to have a defined way of measuring what AI is doing for them. The answers are self-reported. Its message is still clear: the leaders changed how work and decisions flow, and the rest added a tool.

Source: McKinsey, The state of AI in 2026 (August 2026).

Why adopting hits a ceiling

Watch someone at the adopting stage for a day. They ask AI to draft a proposal, read it, fix two things, decide which version of the price to use, and send it. Then they ask for the follow-up email, read it, decide whether the client needs a call instead, and so on. AI does more and more of the labour. Every judgment call still comes back to the person.

That sets a ceiling. In what we see with clients, people who direct AI task by task often get about 50% more done, and two to four times as much when they are really good at it. Those are rough figures from our own practice. Past that point, output is capped by how many decisions a person has time to make. Long chains of work make it worse: if each of 15 steps comes out right 95% of the time, the whole chain comes out right less than half the time, so a person ends up checking every step.

There is a second problem. In most setups, nothing the AI learns sticks. You correct it on Tuesday and make the same correction on Thursday. MIT's much-quoted 2025 study of company AI projects found most produced no measurable return. The main reason it gave was a learning gap: most AI systems could not keep feedback, adapt to the business or get better over time. The study's methods have been criticised, but that explanation matches what we see.

Ramp's own research hints at the size of the difference. Tracking more than 21,000 US firms, it found that companies adopting AI grew their headcount about 10% over two years, a gain that came entirely from the third spending the most on AI per employee (around US$30 a month, often across several models and coding tools). Light spenders showed no significant change. That is a correlation, and the heavy spenders were already growing faster, but it fits the idea that one group has broken through the ceiling and the other has not.

The top 1% have built their way past both problems. Four habits stand out.

Sources: MIT NANDA, The GenAI Divide (2025), as reported by PPC Land. Ramp and Revelio, Heavy AI adopters hire more (June 2026).

Habit 1: they record decisions, and the reasons behind them

Spawn recorded every call from the day it was founded and filed the text next to its work, long before AI was good enough to use it. The founder says they did it because they expected those records to become the best possible context for AI one day.

That record paid off this October. By then the AI was doing so much of the work that the only thing left waiting was the founder's own judgment. He describes waking up to 700 questions queued for him, most of which he had answered many times before. So the team turned two years of recorded decisions into a set of principles for each person and built AI “judges” that apply them. A judge can answer a question that has been decided before and say where, send a badly framed question back, or hold a genuinely new question for a person.

In the judges' first eight days they returned more than 11,000 rulings: about 7,300 answered from past decisions, about 2,200 sent back and about 1,900 held for a person. The founder says he had to give his own answer 27 times. He says the company shipped about 1,000 changes to its product a day the week before the judges, and 7,250 on October 10, in their first week: about seven times as much in a day.

A record like that is ordinary to build. For a service firm, one entry might look like this:

FieldExample
The decisionA repeat client asked for a rush job with a 48-hour turnaround
What we didSaid yes, no rush fee, moved one internal deadline
WhyThird project this year, and they paid early on the last two
What would change itA first-time client, or a week already over 90% booked
Who decidedOperations lead

The last two rows matter most. “Why” lets the AI apply the call to a situation that is similar but not identical. “What would change it” tells the AI when it should stop and ask. Nobody has to write these from scratch. Your AI reads the email thread or call notes and drafts the entry, and the person who made the call checks it and saves it.

Habit 2: they let AI earn its freedom

The top adopters do not hand AI the keys on day one. They widen what it is allowed to do one kind of decision at a time, as it shows it gets that kind right.

Anthropic's research on how people use its Claude Code tool shows this pattern in the data. New users let the AI carry out every step without asking in about 20% of their sessions. People who have used it hundreds of times do so in over 40%. The experienced users also step in more often, about 9% of the time against 5% for new users. They hand over more and watch more closely, both at once. Anthropic's own advice is that good oversight does not mean approving every action. It means being in a position to step in when it matters. The tool is used mostly for software work, but the pattern carries over: freedom grows with evidence, and watching never stops.

Figure 02How AI earns more freedom
  1. AI suggests the call

    It reads the record and proposes what your people would usually decide, with the reason.

  2. A person approves or corrects

    Corrections are saved with the reason, so the record gets sharper every week.

  3. Check the match rate

    The owner watches how often the AI's suggestion matches the call your people actually made.

  4. AI handles it and reports back

    Once it matches week after week, that kind of decision moves to the AI. People see a summary and the exceptions.

Fixed gatePayments, legal terms, client promises outside set limits

Always a person, every time.

One kind of decision at a time. A miss sends that kind back to step 02 of this loop. Anything that cannot be undone, like a payment, legal terms or a promise to a client beyond limits you set, stays with a person no matter how good the match rate gets.

Even at Spawn, where AI agents change the live product every hour, anything that touches the outside world and cannot be undone, such as sending a payment, goes through a small set of fixed checks. AI can do a great deal there, and payments still need a person.

The opposite approach has a well-known cost. In 2024 Klarna said its AI assistant was doing the work of 700 customer service agents. By May 2025 its CEO said cost had weighed too heavily in the decision and quality had suffered, and the company started hiring people again. It still uses AI for most chats. Our reading: autonomy given ahead of the evidence tends to get taken back.

Sources: Anthropic, Measuring AI agent autonomy in practice (February 2026). Maginative, Klarna dials back its AI customer service strategy (May 2025).

Habit 3: they feed every correction back in

A record only helps if it keeps learning. When Spawn's judges get a call wrong, the team says so, and the judges read that correction the next time. The founder says about 60% of Spawn's work starts from a signal rather than someone assigning it: an error in the product, a bug report, an odd entry in a log or a feature request.

The service-firm version is a habit more than a system. When the AI gets something wrong, the fix goes into the record with the reason, so the same mistake does not come back next week. When a client complains, the complaint and what was done about it go in too. Over a few months, the record becomes the most accurate description of how your firm actually works.

Habit 4: they build so every new model makes them better

This habit is why the leaders keep pulling further ahead. AI is improving very fast at paid, practical projects, the kind clients actually buy.

Figure 03How much paid project work the best AI can finish

Share of 240 real, paid freelance projects where the AI's finished work was judged at least as good as the professional's.

0%10%20%30%40%50%OCT 2025FEBJUNOCT 20262.5%15.83%21.25%41.3%NOV 2026
The Remote Labor Index, from the Center for AI Safety and Scale AI, gives AI 240 real projects that freelancers were paid for, in design, data, video, architecture and more, worth about US$600 each on average. Each point is the best score at the time, from the Scale AI leaderboard. The November point is Anthropic's newest model, Fable 5.2, at 41.3%, now first on the same leaderboard (reported by FourWeekMBA, November 11, 2026). A score of 41% means the AI's work matched or beat the professional's on 41% of the projects. It does not mean AI can do 41% of a freelancer's job.

Think about what a jump like that does for two different firms. One uses AI as a better search box and a faster way to draft emails. A better model gives it slightly better answers. The other has a decision record, an AI that already suggests most routine calls, and a habit of moving work to the AI as it proves itself. For that firm, a better model can take on more of the decisions the day it comes out. Cost moves the same way: Epoch AI, a research group, estimates that the same level of AI performance has been getting about 13 times cheaper each year since 2023, for firms that switch to the cheapest model that gets there.

Spawn's founder says each new model has made his team faster and more consistent, and that with the right structure every new model makes the whole company better. He also makes a point that is easy to miss: some of the instructions you wrote to make up for an older model's weak spots stop being needed, and they start holding the new one back. Scaling firms review and trim their instructions when a major new model comes out.

Sources: Remote Labor Index leaderboard, Scale AI. Epoch AI, The plunging price of thought (September 2026). Jacob Sansbury (October 2026).

What this looks like in a small or mid-sized firm

Spawn is a software company, so borrow the method. The setup will look different in your firm. You will not need AI changing a live product every hour. You will need your senior people's repeated calls written down, and a safe way to hand those calls over. A sensible order:

  1. Finish adopting first. If leaders are not yet using AI every week on their proposals, email and planning, start with our adoption guide. Scaling sits on top of it.
  2. Pick one decision that comes up every week. Something your senior people answer the same way almost every time, where a wrong call is cheap to fix. For a builder that might be scheduling a trade, booking inspections, or which warranty callbacks get a same-week visit.
  3. Record it for a month. The call, the reason, what would change it and who made it. The AI drafts each note and the person approves it.
  4. Let the AI suggest the call. A person approves or corrects, and corrections go into the record with the reason.
  5. Hand it over when the suggestions keep matching. The AI handles that kind of decision and reports back. People see a summary and the exceptions.
  6. Keep fixed gates on anything you cannot undo. Payments, legal terms and anything promised to a client outside limits you set in advance always go to a person.
  7. Add the next decision. Each one you hand over frees senior time for the next.

The record can live somewhere simple: a shared document or a project folder inside the AI tool your team already uses, such as a ChatGPT or Claude project, or a SharePoint folder Copilot can read. Keep it in your business account, and leave out client personal details the AI does not need.

Someone has to own the record. Usually it is the accountable owner from the adoption guide, the person who answers for whether AI is working. They decide what gets recorded, what stays out (sensitive details the AI does not need), and when a kind of decision is ready to move.

More from us: AI audit logs: the minimum system a small company needs and What happens to your staff when you bring in AI.

How we work: learning from the companies at the front

We don't work at an AI lab, and we're not inside a giant company on the bleeding edge. We are heavy, everyday users of AI. Our obsession is working out what the companies at the very front do differently, running our own business that way first, and then setting it up in our clients' companies. The labs and the leading adopters run the expensive experiments in public. We bring what works to firms that will never have an AI team of their own.

Sidekick runs on the habits in this article, starting with our own decisions. When Chase corrects the AI, it saves the correction, and those saved corrections become rules it reads before its next job. When a long piece of work is finished, the AI records what was done and why, along with the decisions and their reasons. We also use AI judges, like Spawn's, that rule on routine questions from that record before they reach Chase. We test them by replaying past decisions about running Sidekick, such as offers, priorities and which follow-ups to send, and comparing the judge's answer with the call Chase actually made. In two internal tests, on the decisions they ruled on, the judges matched his actual call 171 times out of 180 (95%) and 155 out of 172 (90%), and they passed the decisions they were unsure about to him. These are our own tests on our own decisions. They show the method working for us; your firm would run its own tests before handing anything over. Anything new, sensitive or uncertain still goes to him.

Common questions

Do we need to spend like the top 1% to scale with AI?

No. Their spending comes from the system they built. A 10-person firm can build the same habits (a decision record, earned autonomy, a feedback loop) on the business AI plan it already pays for. Your spending should rise only as AI takes on work that clearly pays for itself, and you decide when that happens.

Is it safe to let AI make business decisions?

Start with AI suggesting and a person approving, and only move a kind of decision to the AI after its suggestions match your people's calls week after week. Keep a fixed rule that anything you cannot undo, such as payments, legal terms and promises to clients, always goes to a person.

Which decisions should we hand to AI first?

Ones that come up every week, that your senior people answer the same way almost every time, and where a wrong call is cheap to fix. Scheduling, which template a reply uses, who gets copied and which internal deadline moves are common first picks.

How long does it take to go from adopting AI to scaling with it?

Expect a few months per area of the business, one kind of decision at a time. Most of the early weeks go into recording decisions and checking the AI's suggestions. That timing comes from our practice; we know of no published research on it.

What happens when the AI gets a decision wrong?

The correction goes into the record with the reason, and that kind of decision goes back to needing approval until the AI's suggestions match again. A wrong call teaches the system something, as long as someone writes down why it was wrong.

Does this only work for tech companies?

The best-documented examples are software companies, because they measure everything. The method suits any firm whose senior people make the same judgment calls over and over, which describes most service businesses: builders, accountants, agencies, HR and consulting firms.

Where Sidekick fits

Plenty of businesses are dabbling with AI, a smaller number have adopted it, and very few are scaling with it. The gap is rarely the tool. It is whether the business keeps a record of how its decisions get made, and lets AI take on more of them as it earns it.

That record is what we build. With Solo, we set this up for one leader: their AI connected to their email, calendar and files, their recurring calls recorded, and the AI taking repeat decisions off their plate. With Fractional, we act as your part-time head of AI and build the company version: an owner for the record, the first decisions handed over safely, fixed gates on what matters, and before and after numbers.

To see where this could pay off in your business first, try the free AI Opportunity Map. It takes about 3 minutes.

For companies

Get your team proficient with AI in 90 days.

Sidekick runs your AI rollout, and Chase leads it. We coach your leaders first, write your AI plan and staff policy with you, train your teams, and measure results against goals you set. Start with your free AI Opportunity Map.

Have a question first? Ask it here, text Chase at 778-999-0985, email, or connect on LinkedIn.