TL;DR
The short version
The distance between elite and average AI users tripled in six months, and it has nothing to do with who has the better model. It comes from delegating real work to agents. If you lead a team, the fix is operating patterns, not more tool budget.
On The AI Daily Brief, Nathaniel Whittemore reads OpenAI's enterprise data: everyone has the same models, but the frontier hands far more of its work to agents — and that lead compounds.
Built on The AI Daily Brief episode "What the Top AI Users Are Doing Differently" (2026-08-26), citing OpenAI's Enterprise Signals data.
The gap tripled, and the model isn't the reason
OpenAI ranks its enterprise customers each month by output tokens per active user. In January, the top 10% of firms used 2.6x more than a typical firm. By June, that gap was 8.3x.
OpenAI, Enterprise Signals (updated Aug 2026).
The part worth sitting with: every one of those firms has the same models. The frontier isn't buying better AI — it's putting the same AI to work differently.
This is not a model question. It's about how they put those models to work.
Nathaniel Whittemore, The AI Daily Brief
That means the gap is closable. It also means it keeps widening for anyone who treats AI as a smarter search box.
The shift is from asking to delegating
A year ago, enterprise AI was almost entirely chat: ask a question, get an answer, paste the output somewhere. As of June, agentic work (OpenAI's Codex) made up 64% of combined Codex and ChatGPT output tokens. In under a year, the balance flipped from asking to delegating.
The surprise is who's driving it. The fastest growth isn't engineering. Since February, weekly active Codex users grew 108x in legal, 41x in sales, 41x in recruiting, and 26x in marketing — against 5x for engineers (finance 20x and healthcare 24x per VKTR). These are functions that had almost no agent use six months ago. As one legal analogy put it: LLMs are for lawyers what spreadsheets were for accountants.
What frontier teams actually do differently
Two operating habits separate the frontier from everyone else.
First, they build reusable context. Frontier firms put skills (reusable instructions for common workflows) and plugins (bundles connecting an agent to company apps and data) in front of far more people: 19% of weekly users on skills vs. 3% at typical firms, and 21% on plugins vs. 9%. For scale, 95% of OpenAI's own employees use plugins weekly — the ceiling is high, and most teams are nowhere near it.
Second, they climb a use-case ladder: generation (draft an email or a formula), then synthesis (combine sources into something new), then execution (act inside real systems), then maintenance (keep a system running). Chat lives on the bottom rungs. Agentic work moves up into systems-level tasks. In legal, that looks like agents comparing contract terms, flagging deviations, and drafting redlines — while humans still set risk tolerance and approve final language.
The real bottleneck is habit, not capability
The most honest moment in the episode came from Sam Altman, admitting he still works the way he always has despite having a coding agent on his desk: "By revealed preference, I have a better way to do it now and I still do it the old way… I just think we all get stuck in the patterns of how we've always done things."
If the CEO of OpenAI defaults to the old workflow, your team will too. Access to a tool changes nothing. Rebuilding the muscle memory of how work gets done is the actual project — and it's uncomfortable enough that people avoid it while telling you they're all for AI.
What to have your team do differently
Stop measuring AI adoption by seats and licenses. Measure depth: how much real work is being delegated, not how many people logged in.
- Pick two or three recurring workflows per function and turn them into skills, so the good pattern is reusable instead of re-invented in every chat.
- Connect agents to the systems where work already lives — the CRM, the contract repo, the codebase — through plugins. An agent with no tools is just a chatbot.
- Push agents up the ladder deliberately: from drafting, to acting inside a system, to maintaining it, one rung at a time.
- Name the boundary out loud. Agents handle coverage and coordination; people keep risk judgment and accountability.
One caution before you chase the numbers: output tokens measure consumption, not value, and OpenAI says so itself. MIT's 2025 research found most enterprise AI pilots returned nothing to the P&L. The goal isn't to burn more tokens than the firm next door — it's to move real work off your people's plates and keep the judgment where it belongs.
