The AI Adoption Arbitrage: Why Most Companies Are Stuck

"Some of them are 50-to-100-million-dollar companies, and that's how they're using it. We're spending an extra five dollars a month per license so people can chat with the Copilot. It's a shitty version of AI."

That was a strategy advisor at this week's AI Roundtable, describing the AI adoption pattern among the mid-market companies he works with. He calls it stage one — typing into ChatGPT, not yet knowing there's anywhere else to go. Operators who understand how to integrate AI more deeply into their businesses have started to pull away from the pack.

What follows are notes from this week's Executive AI Roundtable discussion, shared under the Chatham House Rule.

Stage One Everywhere

The pattern shows up at every scale. The strategy advisor described it across his panel attendees:

"Go outside your bubble, and you're not gonna find a lot of people doing more than just typing shit into ChatGPT. They are stage one everywhere, and they don't even know how to get out of that, and they don't know that there's something more than that."

A founder who offers AI workshops for CEOs echoed the same gap from a different vantage. Attendees show up describing themselves as "completely AI native" — "use it all the time" — and it turns out they have a free ChatGPT account. Others are "absolutely shit scared." There is an enormous mix of hype, fear, and ignorance among leaders.

This is the hidden tech misalignment many leaders don't quite comprehend. The CEO assumes the team has adopted AI because the company bought Copilot. The team assumes it's adopted AI because there's a ChatGPT tab open in a browser. Neither perspective matches what adoption looks like in 2026, which includes agents that enhance or replace business workflows, systems that improve over time, and continuous learning at the organization level.

The No-Code Brownfield

The same gap looks different from the legacy side. A founder who'd spent years working in the no-code ecosystem described what happened when ChatGPT arrived. The pitch for no-code had been to democratize software creation. AI takes that concept to the next level.

Net revenue retention at his no-code clients is healthy. Old customers are sticky. However, new logos have nearly plateaued. The engineering hours are going into brownfield maintenance — patching platforms whose architectural assumptions predate the LLM foundation models. Cash flow from the legacy product makes the maintenance work rational; however, the market is moving without them. New customers are scarce. The reasoning new customers go through is short: "Why wouldn't I just do this in Claude Code?"

"I was investing in buggies when the car came out."

This is the innovator's dilemma playing out in real time, at the platform level. Three questions follow from it.

First: if your business supports a stack designed before 2023, look at where your engineering hours actually go this quarter. Are you balancing maintenance spend with innovation to maintain long-term product-market fit?

Second: which of your software vendors is the buggy-maker now? If you're happy with them, you're good. However, if you're counting on their platform to evolve, double-check their roadmap.

Third: is your own product something AI has already made obsolete? If a customer could get a usable substitute by typing into Claude for ten minutes, your roadmap dollars need to be aimed somewhere less exposed to AI risk.

The Wait-It-Out Bet

There's a third way to miss the arbitrage that has nothing to do with ignorance or legacy maintenance: deliberate paralysis. Some leaders see the field shifting fast and decide to wait it out before committing.

"They're just buying time, waiting for things to settle out."

Waiting is a strategy with its own cost. Every quarter the operators pull further ahead. The bet pays off only if the field actually settles into something stable. It hasn't, and it isn't about to.

What Operator-Class AI Actually Looks Like

The conversation got specific about what "past stage one" actually means in practice. A founder who runs onboarding programs for other operators walked through a system he's built on top of Claude and Dropbox that he runs his entire business through:

• His sales conversations get transcribed, then a Claude project compresses each one into 800-word notes pulled out as jobs-to-be-done moments — what the prospect actually wanted, in their words.
• A separate agent watches for conflict-of-interest situations against his published policies and flags potential breaches before he commits.
• A publication agent turns the transcribed material into LinkedIn content, marketing pages, and talking points.

How that system bootstraps is itself instructive:

"It creates what it thinks is a skeleton of your business, an outline based on public information. And then you start to define that, and then you start to use it."

The agent does first-draft scaffolding from whatever's public; the operator refines it iteratively until it matches the real thing.

The operating-system framing matters. Each agent has a defined scope, a small explicit set of instructions, and a clear handoff back to a human for judgment. The work compounds because the inputs (transcripts) feed multiple outputs (sales follow-up, marketing, governance) through a single pipe.

A different operator described the same pattern from another angle. He'd built a constellation of his own mental models out of years of his writing and talks. The agent he trained on top of it sifts through meeting transcripts and links each surfacing idea back to a mental model: "Here's an idea that actually links to a mental model." Each meeting becomes another star in the constellation.

What separates these two examples from the stage-one pattern is that AI powers an explicit operating system for their business. For the foundations beneath that — mapping how your business actually runs before pointing a model at it — see Map Your Operating System Before You Apply AI.

A data-and-ops expert in the room came at the operator-class arbitrage from a different angle still. For decades, a class of work resisted automation: spreadsheets someone filled in by judgment, SOPs with subjective decisions baked into them, free-form text fields no rules engine could parse.

"There's all these messy workflows that previously couldn't be automated — some subjective SOP, somebody decides something, the data is messy, free-form text fields that can't go into a standard workflow. But now all that's possible."

The operator-class move there is plainer than agents: name the messy work that finally became automatable when language models arrived, and pipe it through Claude with a few clean rules.

The arbitrage in 2026 lives in this gap. One attendee named the same point directly:

"There's such an arbitrage opportunity, isn't there, between companies that can and companies that can't."

Another, surveying his own client work, was blunter:

"It is incredible how much low-hanging fruit there is."

The five-dollar Copilot company has a fundamentally different ceiling from the company running an agentic business operating system. And the gap is widening.

One area where the AI adoption gap hasn't shaken out is regulated industries. There is a lot of talk and hype around locally-run AI models. "I have not seen it operationalized beyond what I would call relatively trivial forms of AI," as one attendee put it. They pointed to transcription tools like Whisper and text-to-speech models as examples of the local or on-device models used in business today. There may be opportunities for businesses that can supply enterprise-grade hardware and AI models that can run in a corporate server room. Or perhaps these businesses will overcome their reluctance and sign enterprise agreements with the major AI providers.

The Approval Theater That Rewards Leakage

From the same operator: a story relayed from a data science conference last fall.

A large company published an "approved AI" list. The list was conservative — locked-down internal tools, mediocre output. A new hire, a 25-year-old fresh out of school, joined believing he'd landed at an AI-first company. He took one look at the approved list, decided it wouldn't do what he needed, went home, used personal ChatGPT against company data, and brought back a working prototype.

Leadership celebrated him at the all-hands. He got a massive bonus. The rest of the engineering organization was asked why they weren't being equally innovative.

The rest of the staff drew the obvious inference: the way to succeed at this company is to bypass the approved-tool list. Internal data started moving out to personal accounts at scale. The "approved" governance posture was now an incentive structure for shadow AI.

"Their staff knows they should leak because they're being rewarded for it."

This is the hidden tech misalignment most CEOs can't talk about in public. The official posture is we have AI governance. The actual incentive is bypass it to get promoted. Most enterprise AI rollouts contain some version of this — a controlled-AI policy paired with a results-rewarded culture pulling the other way. The policy says no. The bonus says yes. Staff either seek less dysfunction by leaving the company, or they start ignoring the rules.

You can avoid all of this by offering better internally approved tools — ones that outperform what a 25-year-old can build at home on a Saturday. When the approved stack loses to the alternatives, leadership is failing, and the user trying to work around it is just being rational.

For the broader pattern of vendor AI quietly adding features that train on your data, and the credential-firewall approach to scoping AI access, see Shadow AI: The Tools Your Team Uses Without You.

Where to Start This Quarter

If you suspect your company is closer to stage one than your strategy deck implies, the first two items below earn the right to do the rest. Audit before you build; build one pipe before you scale:

1. Audit what "AI adoption" actually means at your company. If the answer is "we pay for Copilot" or "engineers can use ChatGPT," you're still in stage one. The next-stage answer names specific workflows — sales transcript compression, governance auditing, marketing publication pipelines — that run against specific company artifacts.
2. Build one operating-system pipe before ten chat workflows. Pick a single high-volume input (sales calls, support tickets, hiring conversations). Pipe it through Claude with project-level instructions. Pull two or three outputs (jobs-to-be-done notes, follow-up drafts, governance flags) from the same pipe, and use those new insights to build better business processes. Ideally, you're building a stronger system with each pass.
3. Reconsider your "approved tool" list against what's actually performing. If your approved stack falls short of what an enthusiastic new hire can build at home, your technical team has fallen far behind the curve. You need to update your tools and take a serious look at the leadership behind your AI choices.
4. Take inventory of brownfield drag. Where are your engineering hours going this quarter? If most are maintaining infrastructure whose architectural assumptions predate ChatGPT, the company is paying compound interest on a three-year-old bet. Re-evaluate your engineering priorities.
5. Find someone who's past stage one. As one attendee put it, "most people are being criminally badly advised about what they should do" — the signal-to-noise in the AI advisory market is "absolutely bonkers." The operators running these systems can build them with you in weeks. For the full deployment playbook, see How to Roll Out AI Across Your Company.

The gap between the stage-one company and the operator-class company will be a major business divider in 2026. The operator-class playbook is short: a handful of agents on top of a written-down operating system, fed by the work your team already does. Writing the operating system down — the part most teams skip — is where the cost lives.

Resources From the Roundtable

• Claude Code and Claude Team — the operator-class default AI toolset for most Roundtable attendees.
• Microsoft Copilot — the stage-one default at mid-market scale; the five-dollar-per-seat add-on most companies mistake for an AI strategy.
• Whisper and WhisperKit — open-source local transcription, preferred for sovereignty-sensitive work; runs on-device with much smaller models than language LLMs.
• AI Stages of Adoption (AISA) — Mario Thomas's June 2024 framework; the "Experimenting" Stage 1 maps closely to the in-room shorthand used throughout this essay.
• The Innovator's Dilemma — Clay Christensen's framework, cited directly when describing the no-code companies' brownfield trap.
• Jobs to be Done — the framework underneath the transcript-to-marketing pipeline; the practitioner most often cited in the room is Bob Moesta.
• Open Data Science Conference (ODSC) — the source of the "approved AI / rewarded leakage" story.