Frontier Firms vs. Everyone Else: The 8.3x AI Agent Output Gap

Frontier firms now pull 8.3 times more output from the same AI models everyone else is using. Not a bigger model. Not a bigger budget. Same tools, eight times the result.

This is for any operations or technology leader running an agent pilot right now, past the initial excitement and into the harder question of why some teams get so much more out of identical software. It isn't a claim that one AI vendor beats another, and it isn't a claim that spending more wins. The data below is from mid-2026, and it's specific about what "more output" actually means and what it doesn't.

What is a "frontier firm," and why did the gap just triple?

A frontier firm, in OpenAI's enterprise research, is one of the top 10% of business accounts ranked by monthly AI usage per active user. It's a usage percentile, not a size category or an industry label. As of mid-2026, frontier firms generate 8.3 times as many output tokens per active user as typical firms, up sharply from 2.6 times in January the same year (OpenAI, 2026).

Sit with that for a second. The ratio didn't creep up. It more than tripled in five months, while the underlying models available to both groups stayed roughly the same. And that's the real story here: the gap is opening because of how the top decile uses the tools, not because they have access to something better.

A second "frontier firm" term is circulating from the same period, from Microsoft's Work Trend Index research, describing an organizational archetype built around agent-augmented teams (Microsoft WorkLab, 2026). Related idea. Different company, different data, different claim. Conflating the two muddies what's actually being measured, so this piece sticks to OpenAI's usage-percentile definition and its 8.3x figure specifically.

Why does usage pattern, not budget, explain the gap?

The clearest signal in OpenAI's data is what the heaviest users are doing differently: delegating full tasks to agents instead of asking for help with a piece of one. As of June 2026, Codex, OpenAI's coding agent, generated 64% of combined Codex-and-ChatGPT output tokens among enterprise customers, a sign that agentic, multi-step work is overtaking single-turn chat even inside the same accounts (OpenAI, 2026).

Plugin adoption tells a similar story. Each week, 21% of active users at frontier firms use plugins that connect agents to company systems and tools, compared with 9% at typical firms. Plugins are what let an agent actually reach a CRM record, a ticket queue, or a document store, instead of just answering a question about them.

A few patterns recur across the accounts pulling ahead:

• Agents get access to real company context and tools, not just a chat window.
• Usage concentrates on two or three workflows tied to revenue, cost, or cycle time, not spread thin across dozens of shallow use cases.
• Leaders use the tools themselves and talk openly about what worked and what didn't, a pattern Microsoft's own leadership research found among the fastest-moving adopters too (Microsoft Cloud Blog, 2026).

What do frontier firms actually do differently with agents?

This is where the AI Work Redesign Framework's first three steps, diagnose, redesign, enable, explain the behavior behind the numbers.

Diagnose means naming the actual workflow an agent will touch, end to end, before deciding what it should own. Frontier firms aren't asking "where can we add a chatbot." They're asking which workflows most directly move revenue, cost, risk, or response time, then working backward from that answer (Microsoft Cloud Blog, 2026).

Redesign means rebuilding the handoffs in that workflow so an agent can complete a full step, not just draft a suggestion a human still has to rewrite. Enable means giving the agent the access it needs to do that step for real, the plugin, the data connection, the tool permission. Skip enable, and the agent stays a glorified autocomplete no matter how good the underlying model is.

None of this requires a bigger AI budget.

It requires picking fewer workflows and actually finishing the redesign on each one.

How do you tell if your company is assisting or delegating?

A rough self-test, built from the patterns above:

1. Do outputs from your AI agents get used as-is, or does someone rewrite them before anything happens?
2. Can the agent reach the actual system of record (CRM, ticketing, documents), or only a chat window?
3. Is usage concentrated in one or two high-value workflows, or scattered thin across many?
4. Do your leaders use the tools themselves, or is usage pushed down to junior staff only?
5. Is there a named owner for each agent's scope, or is "AI" everyone's job and no one's?

Two or more "no" answers means you're closer to assistance than delegation, regardless of what's on the invoice.

What's the real risk of chasing frontier-firm status too fast?

The case against racing to close this gap is fair, and worth stating plainly: handing agents more autonomy before governance catches up creates real accountability problems. When an agent can plan and execute across many steps, unclear ownership of its outputs becomes the actual risk, not the model's capability (Fast Company, 2026).

Where I land: that argument supports sequencing, not stopping. But sequencing only works if it has an endpoint, not an indefinite pause. Treat each agent like a new hire, with a defined role, explicit guardrails on data and tools, and a clear point where it has to stop and check in with a person. That structure is what lets a company expand delegation without losing track of who answers for a mistake.

How should a leader measure whether this is actually working?

Output tokens are a usage metric, not a return metric. The measure step of the framework means tying agent activity back to something the business already tracks: cycle time on the redesigned workflow, cost per resolved case, or revenue tied to a faster response.

Pick the metric before the rollout starts. Set a baseline. Check it against the same measure 60 to 90 days after the redesigned workflow goes live. A frontier-level usage ratio with no change in the underlying business metric is a sign the redesign step got skipped, not a sign the data is wrong.

I'd be curious whether this 60-to-90-day window holds up outside a usage-heavy function like engineering or support. If you've measured a redesigned workflow on a longer or shorter clock and gotten a cleaner read, I'd like to hear what you found.

Frequently asked questions

What is a frontier firm in AI adoption?

A frontier firm is OpenAI's term for the top 10% of enterprise accounts ranked by monthly AI usage per active user, not the company that bought the most licenses. The ranking measures depth of use, including how much work gets delegated to agents versus kept as simple chat assistance.

Does 8.3x more agent output mean 8.3x more business value?

No. Output tokens measure usage depth, not revenue or cost impact. A company closing the gap still has to tie agent output to a specific workflow metric, such as cycle time or resolved cases, before it counts as ROI.

Do we need to buy more AI licenses to become a frontier firm?

Not necessarily. OpenAI's data shows frontier firms use the same underlying models as typical firms. The difference is which workflows get agent access and how much of the task gets delegated, not how many seats get purchased.

What's the difference between OpenAI's "frontier firm" and Microsoft's "Frontier Firm"?

They are two different definitions from two different companies. OpenAI's term is a statistical usage percentile from its enterprise signals data. Microsoft's term is an organizational archetype from its Work Trend Index research. Don't cite them as the same finding.

What's the first workflow to redesign if we want agents to do more?

Pick one workflow that already touches revenue, cost, risk, or customer response time, and that currently stalls on a human handoff. Redesign that single workflow for agent delegation before expanding, rather than deploying agents broadly across many shallow use cases.

Key takeaways

• Frontier firms are the top 10% of enterprise AI accounts by usage per active user, not the biggest spenders; the gap over typical firms grew from 2.6x to 8.3x in five months.
• The gap comes from delegation depth: giving agents real tool access and letting them complete full steps, not from a better model.
• Two or three redesigned workflows, chosen for revenue, cost, risk, or speed impact, beat a broad, shallow rollout across many use cases.
• Expanding agent autonomy needs a defined role, guardrails, and an escalation path per agent, or accountability becomes the actual risk.
• Usage volume isn't ROI. Tie agent output to a workflow metric you already track, and check it 60 to 90 days after launch.