How Frontier Firms Lead in AI Adoption

Most organizations now use AI. A much smaller group reports deeper use, broader deployment, and stronger outcomes. The gap is not about licenses.

The default belief is that adoption equals access: buy seats, run a pilot, wait for the returns to arrive. The evidence points somewhere else. The firms pulling ahead are not the ones with the most tools switched on. They are the ones that went deep — into more functions, more advanced capabilities, and more governance — and then turned individual wins into repeatable systems.

That is the thesis of this article. It is also a thesis with a weak evidence base, and I want to be honest about that up front. The headline numbers come largely from vendor-commissioned research and vendor telemetry. They are directional signals, not audited outcomes. Read them as a map of where the leaders say they are, not as a controlled experiment proving what caused the lead.

The Adoption Gap Is a Depth Gap

Start with definitions, because two different numbers keep getting conflated.

The first is headline adoption: how many organizations use AI at all. A Microsoft-commissioned IDC study of more than 4,000 business leaders, published in late 2025, found 68% of surveyed companies using AI. The Canadian cut of the same study reported 65%. A separate U.S. Census Bureau survey cited in spending coverage put business AI use at 22%. The spread tells you something important: "adoption" depends entirely on who you ask and what counts.

The second number is the one that matters here. Within that same research, a small cohort — the "frontier firms" — is defined not by size or industry but by breadth and depth of AI use. Microsoft's Canadian release put frontier firms at 14% of Canadian respondents versus 22% globally. That is the real shape of the market: a wide base of users, a narrow tip of deep users.

So what separates the tip from the base? The research uses a few proxies for depth:

• Output tokens per active user. OpenAI's enterprise research defines frontier firms as the top 10% of AI usage each month and reports they generate 8.3× as many output tokens per active user as typical firms. A token is the unit of text a model reads or writes; token volume is a rough proxy for how much work is actually flowing through the system.
• Advanced capability usage. Among weekly active users, OpenAI reports 21% at frontier firms use plugins and 19% use skills, versus 9% and 3% at typical firms. Plugins and skills are ways of connecting a model to external tools and repeatable instructions — the difference between asking a question and wiring a capability into a workflow.
• Number of business functions served. Frontier firms report using AI across roughly seven business functions on average.

One caveat before we go further: the 3x and 4x return figures come from vendor-commissioned IDC and OpenAI studies. Treat them as vendor-reported signals, not independent audits. The direction is plausible. The precision is not established.

What Frontier Firms Actually Do Differently

If the gap is depth, the next question is what depth looks like in practice. The sources point toward four recurring practices, though they measure different things — survey-reported behavior, telemetry proxies, and playbook recommendations — so read the convergence as directional rather than proven.

Breadth across functions. Frontier firms report using AI across roughly seven business functions, with over 70% deploying it in customer service, marketing, IT, product development, and cybersecurity. These are functions where AI can automate a workflow, generate content, or detect anomalies in real time — not where it produces a clever demo.

Depth in advanced capabilities. The plugin and skill usage numbers above are the clearest signal. Frontier firms are not just prompting more; they are connecting models to company context, tools, and repeatable steps. That is the difference between an assistant and a system.

Customization. Microsoft's research reports 58% of frontier firms using custom AI solutions, with 77% planning to within 24 months. Custom solutions embed proprietary knowledge, tone, and compliance requirements — fine-tuned on internal data or industry-specific knowledge. The strategic point is not the technology. It is that a competitor cannot copy a workflow built on your data and your review process.

Governance and observability. Microsoft's playbook describes frontier firms embedding governance, security, and compliance into AI systems rather than bolting them on after scale. Observability here means you can see what the system did, why it did it, and where it failed. That is not a compliance checkbox. It is what makes a workflow safe to promote from one team to the whole company.

A note on evidence quality: all of this is self-reported survey data and vendor telemetry. There is no controlled comparison here. The patterns are consistent and plausible, but they describe what leaders say they do, not what an independent study measured.

Why the Gap Widens Instead of Closing

Here is the mechanism I find most defensible, stated as a model rather than a law.

Deeper usage produces artifacts: workflow data, evaluation traces, and reusable patterns. Evaluation traces are records of what the system did on real inputs and where it failed. Those artifacts make the next deployment cheaper and faster, because you are not starting from zero. The first workflow costs you discovery. The fifth costs you configuration. That is a compounding loop, and it explains why an early lead can widen rather than close.

But the loop only turns if the complementary investments exist. OpenAI's research is explicit that access alone may not be enough to scale AI, pointing to continuous employee learning, shared workflows, data infrastructure, and governance as the supporting conditions. Model access is now cheap and widely available. Those four things are not.

The bottleneck is usually organizational, not technical. In my experience building automation systems, the failure mode is rarely the model. It is that a working individual workflow has no owner, no review path, and no way to become a shared one. It stays a personal trick. The person who built it gets promoted or leaves, and the capability leaves with them.

There is also a counter-signal worth taking seriously. Spending data from Ramp, which tracks roughly 70,000 companies, showed 56% of its customers paying for AI products in August, up just 0.4% from the month before. The same coverage notes price competition pushing customers toward cheaper or open-weight models. If the frontier lead were purely about capability access, cheaper models would erode it. The fact that the lead persists suggests the advantage lives in the workflow and governance layer, not the model layer — but that is interpretation, not measurement.

Where the Frontier Narrative Is Weakest

I would not treat the frontier playbook as a guaranteed formula. Four weaknesses matter.

Vendor incentive. The firms publishing the frontier research also sell the platforms. Microsoft commissioned the IDC study. OpenAI published the enterprise research. The definition of "frontier" and the success metrics are not neutral. That does not make the data wrong. It means you should read the framing as marketing-adjacent.

Survivorship and selection. OpenAI's companion working paper found that enterprise adopters among U.S. public companies held more assets, employed more workers, and had higher R&D investment than non-adopters. Firms with more resources were better positioned before they adopted anything. Some of the "frontier advantage" may be pre-existing advantage wearing an AI costume.

Measurement gaps. Token volume and capability usage measure activity, not business value. A firm can look deep on every usage metric and still ship nothing that changes a P&L. The research measures the inputs to value, not the value itself.

Governance lag. Stanford's 2026 AI Index, cited in recent research, reports organizational AI adoption at 88% while the average Foundation Model Transparency Index score dropped from 58 to 40 in 2025. Deployment is outrunning the ability to evaluate and understand what is deployed. That is a real risk for any firm scaling fast.

The open question: does the frontier cohort stay ahead once cheaper and open-weight models close the capability gap on ordinary tasks? Reuters coverage suggests open-weight models are gaining ground for basic operational tasks while frontier labs retain an edge on demanding work like coding. If that holds, the frontier advantage narrows to the hardest problems — which may be exactly where the durable value was anyway.

A Practical Path for Firms Behind the Frontier

If you are running a mid-size organization and you are not on the frontier, the research suggests a sequence. It is not a platform program.

Pick one high-friction workflow with a measurable output. Not a company-wide rollout. One workflow where the friction is visible and the output is countable — tickets resolved, invoices processed, reports generated. Prioritize recurring, high-volume work with a visible baseline, bounded risk, accessible data, and an available reviewer. Defer workflows where errors are hard to detect or where authority is ambiguous; those are the ones that stall after the pilot.

Instrument it before scaling. Define the output metric, the review point, and the failure path. The review point is where a human checks the output before it goes anywhere that matters. The failure path is what happens when the system gets it wrong. Without both, you cannot tell whether it worked.

Promote what works into a shared pattern. The single biggest leak in most organizations is that a working workflow stays personal. Write it down, template it, hand it to the next team.

Assign ownership. Microsoft's playbook makes this concrete: treat each deployed agent or workflow as a product with an owner responsible for uptime and accuracy, like a product manager for a software application. An agent here means a system that can take actions — calling tools, retrieving data, completing steps — not just answer questions.

Fund learning alongside tooling. The constraint is usually skill and workflow design, not licenses. Budget for it the way you budget for the software. The capabilities that matter map directly to the steps above: workflow and agent design for the first step, evaluation and observability for the second, data and integration fundamentals for the third, and governance literacy for the fourth. The fastest way to build them is to build and inspect a small working system — one workflow, one evaluation, one review point — rather than read about one.

The decision rule I would apply: if you cannot name the output metric and the human review point, you are not ready to scale that workflow. Naming them is the test. Everything else is enthusiasm.

What to Watch Next

Four signals would confirm or falsify the frontier-lead thesis over the next few quarters. I am framing these as scenarios, not predictions.

• Spending and usage growth. Does it re-accelerate, or continue to flatten across the broader market? The Ramp data is one of the few direct spending datasets available and worth tracking as a leading indicator.
• Open-weight erosion. Do cheaper and open-weight models close the frontier advantage on ordinary tasks while frontier labs retain an edge on demanding ones like coding? Reuters reporting suggests this split is already forming.
• Governance catch-up. Do evaluation, transparency, and review practices close the gap with deployment speed, or keep falling behind?
• Independent measurement. Do the frontier cohort's reported returns survive scrutiny from someone who is not selling the platform?

None of these are settled. The honest position is that we have a plausible model, consistent vendor-reported signals, and no independent audit. Pick your highest-friction workflow this quarter. Instrument it. Let the output decide whether the frontier thesis applies to you.