In September 2026, OpenAI published a detailed analysis of how three companies—Basis, Clay, and Exa Labs—have moved beyond AI-assisted tasks to AI-driven operating capability. The findings are instructive for any business leader wondering why their AI adoption feels stuck in pilot mode.
The data gap is striking: OpenAI's Enterprise Signals shows that frontier firms (top 10% of AI usage) generate 8.3× as many output tokens per active user as typical firms, up from 2.6× in January 2026. The gap is widening, not closing. The reason is not access to better models—it is access to better workflow design.
At Vivid Visuals, we have been building AI-native systems for businesses since our original manifesto on AI-native software. The OpenAI analysis validates what we have seen in practice: the companies that win are not the ones with the most AI tools, but the ones that turn proven processes into repeatable, agent-driven workflows.
The Three Patterns: Teach, Persist, Execute
OpenAI identified three distinct patterns across Basis, Clay, and Exa. Each represents a different level of agent maturity.
Pattern 1: Teachable Workflows (Basis)
Basis, which builds AI agents for accounting firms, tackled employee onboarding. First-day onboarding previously took two hours of HR time. With a Codex-powered onboarding skill, it now takes 30 minutes.
The key insight: Basis demonstrated the onboarding process once, then turned it into a reusable skill with a clear trigger, known steps, access to the right tools, and a clear definition of "done." When recurring questions or exceptions appear, HR updates the skill before the next cohort.
This is the entry-level pattern: take a stable, repeatable process and make it teachable to an AI agent.
Pattern 2: Persistent Context (Clay)
Clay, a self-learning revenue engine for go-to-market teams, addresses a deeper problem: deal context scattered across CRM, email, Slack, calls, and presentations. A Clay GTM engineer built a persistent workspace and dedicated subagent for every account. Each subagent reviews primary sources and updates its deal folder overnight. Every morning, a coordinating agent turns those updates into a short list of priority moves.
The workflow saves roughly one hour of inbox triage per night. The supporting evidence stays close to each recommendation, so sellers can inspect sources before acting.
This is the intermediate pattern: give an agent persistent context that evolves as the work changes.
Pattern 3: Bounded Execution (Exa)
Exa Labs, which builds web search infrastructure for AI agents, pursued a goal it calls "Exa everywhere"—making its search API available wherever developers need it. Exa turned the sequence from opportunity discovery to integration into a defined workflow for Codex, with clear priorities, access to necessary sources, and human review before anything ships.
Codex now monitors for high-priority integration opportunities, gathers context, creates pull requests, runs tests, and prepares weekly updates. People still decide which opportunities matter and how external relationships are managed. Tests and review points make the agent's work visible before it ships.
This is the advanced pattern: let an agent carry a signal through to tested, reviewed execution.
The Six Steps OpenAI Recommends
OpenAI distilled the patterns into six implementation steps:
-
Choose one consequential value surface. Start with an end-to-end workflow where strategic priority, systems, handoffs, controls, and measurable stakes meet. It should repeat often enough to learn from and matter enough to justify redesign.
-
Define the outcome and how you will measure it. Name the accountable owner, KPI, baseline, and guardrails. Track depth through completed tasks, connected context, tool usage, and review load. Track value through cycle time, quality, cost, revenue, or risk.
-
Write the agent's job description. Define what triggers the work, the outcome, required context, tools, permissions, and how persistently the agent should work. Specify what evidence it must produce and where it must stop for human review.
-
Build the human system around the agent. Name who owns the business outcome, domain logic, access and controls, adoption, and daily use. Enterprises need explicit decision rights as workflows scale.
-
Make experimentation visible and reusable. Give employees room to test new use cases, then capture what works and package it as skills, plugins, or shared workspaces.
-
Carry the operating pattern forward. Preserve the context, permissions, evaluations, review points, owners, and measures that worked, then apply them to the next value surface.
What This Means for SMBs (Not Just Startups)
The OpenAI examples are startups, but the patterns translate directly to small and mid-market businesses:
| Startup Pattern | SMB Equivalent | Vivid Implementation |
|---|---|---|
| Basis onboarding skill | New hire training + setup agent | Define the process once, update quarterly |
| Clay account workspace | CRM enrichment + deal tracking | Agent reviews accounts nightly, flags priority actions |
| Exa integration discovery | Lead qualification + outreach | Agent monitors signals, drafts outreach, human approves |
The key is not the company size—it is the workflow maturity. A 10-person accounting firm can build a teachable onboarding skill just as effectively as Basis. A regional services company can give an agent persistent CRM context just like Clay. The barrier is not cost (these implementations are affordable on current platforms) but process clarity.
The Common Mistake: Starting With the Tool, Not the Workflow
The most common mistake we see at Vivid Visuals is businesses that start by selecting an AI tool, then searching for a use case. This approach produces pilots that never scale because the underlying workflow was never stable enough to teach an agent.
The correct sequence is:
- Identify a workflow that is stable, repeatable, and consequential
- Document it as a human would (steps, tools, decision points, definition of done)
- Then, and only then, encode it as an agent skill
This is why our AI agents vs. traditional automation framework starts with workflow stability before recommending agent vs. automation approaches. A stable workflow can become either a deterministic automation or an agent skill; an unstable workflow is a poor fit for either.
The Frontier Gap Will Keep Widening
OpenAI's data shows the frontier gap widening from 2.6× to 8.3× in nine months. This suggests a compounding effect: frontier firms learn faster because their agents improve workflows, which improve outcomes, which justify further agent investment. Typical firms stay in pilot mode, testing AI on isolated tasks without ever building the workflow infrastructure that would let them scale.
The implication for business leaders is urgent: the gap between AI-native companies and everyone else is not closing. It is accelerating. Companies that wait for "the technology to mature" before building workflows are not waiting for an easier entry—they are falling further behind with each passing month.
Read next from the blog : Know Your Agent: How Visa, Mastercard, and Ant International Are Building Trust Infrastructure for Agentic Commerce · How to Track Your Brand's Visibility in AI Search: The AEO Measurement Playbook · AI Agent ROI Tripled: What Salesforce's Agentic Enterprise Index Reveals · Why every business becomes an AI-native business