Frontier Firms Use 3.5x More AI Per Worker. The Gap Isn't About Volume — It's About Execution Complexity.
OpenAI's B2B Signals data shows the AI gap has stopped being about who uses it and started being about what they use it for. Typical firms ask AI questions. Frontier firms build AI into multi-step execution — and the difference is widening.
By Forge Team
If you use AI every day at work, you're probably not the problem anymore. The firms pulling away from the middle category are not using AI more often than you. They're using it for different work — not questions and answers, but multi-step execution. OpenAI's B2B Signals research (May 7) found frontier firms now use 3.5x more AI intelligence per worker than typical firms, up from 2x in April 2025. The gap is not about prompting volume. It is about what kind of work AI is actually handling.
What the data actually says
OpenAI defines the distinction plainly: typical firms use AI for questions. Frontier firms use AI to execute — research pipelines, coding, workflow automation — with humans reviewing outputs at checkpoints rather than handling every step themselves. The same week this data published, Anthropic closed a $1.5B joint venture with Blackstone and Goldman Sachs (Fortune, May 4) to embed AI engineers directly inside mid-market companies and redesign workflows around agents. At a New York City event (Fortune, May 5), JPMorgan's Jamie Dimon built a working Treasury dashboard using Claude in 20 minutes while a room of executives watched.
This is not a large-enterprise story. Anthropic's joint venture targets mid-market companies specifically. The practical implication: agent-augmented workflows are arriving in 60-person operations whether the people running them are ready or not.
What to do differently on Monday
The practical shift is from AI as a question-answering tool to AI as a step in a designed process. That requires two things most professionals have not built yet: the ability to decompose a repeatable task into steps where AI handles some and you handle others, and the discipline to decide where you review before you start — not after the output arrives.
Notion's Head of Product described it on Lenny Rachitsky's podcast (May 3) as: "the first 10% of projects is now free." The constraint isn't generation. It's agency — directing, evaluating, and deciding whether the output is good enough to act on. The free 10% only compounds if someone is choosing what to do with it.
What it looks like in practice
Priya manages demand generation at a 40-person B2B SaaS company. Before quarterly planning, she researches five competitors. She opens a chat, asks questions, reads the answers, asks follow-ups. It takes about an hour and produces a set of notes she then writes up herself.
A workflow approach looks like this: she builds a research template once — five competitor names, six consistent fields (positioning, recent messaging, ICP signals, pricing, notable content, relevant news). She gives AI the template, the competitor names, and a brief on what to look for. The research comes back structured and comparable across all five. She reviews it, adjusts where her judgment changes the read, and shares it with the leadership team. It now takes 15 minutes of her time per cycle, and the output is more useful than her freeform notes were.
The difference is not a better prompt. It is a designed process where AI handles retrieval and formatting while she handles judgment.
Build a repeatable AI workflow for a task you currently do as a single back-and-forth conversation.
The counterpoint worth taking seriously
You might object: designing a workflow sounds like more work upfront, not less. That's true the first time. The question is whether the task recurs. A one-off research request doesn't need a workflow. A task you do every month — competitor analysis, vendor summaries, client status reports, budget variance notes — earns back the setup cost within two or three cycles.
Marcus runs operations at a 65-person logistics company. He produces a weekly vendor performance summary — five vendors, six metrics each, sourced from emails, invoices, and a shared spreadsheet. He currently pulls numbers manually and writes the summary himself. He has tried asking AI to "write the summary" and the results are generic because the inputs are unstructured.
The step he is missing: treating the summary as a three-part process rather than a one-shot request. Step 1: AI extracts and categorises the data from the documents he uploads. Step 2: AI formats it against a consistent template. Step 3: Marcus reviews the draft, adjusts any line where his judgment changes the read, and sends. The workflow handles the mechanical work. He handles the interpretation.
Practice breaking a recurring task into three steps where AI handles the mechanical work and you handle the review.
The ladder
The progression the OpenAI data implies runs in four levels: Level 1 is asking AI questions. Level 2 is using AI to draft and iterate. Level 3 is building repeatable multi-step workflows. Level 4 is delegating to agents with defined guardrails and human checkpoints. OpenAI's data places typical firms — the majority — between Level 1 and Level 2. Frontier firms are operating at Level 3 and, with the Anthropic JV building agent infrastructure into mid-market companies, starting to reach Level 4.
The frontier firms are not smarter. They have moved up this ladder faster — and the distance between levels keeps growing.
Practise choosing the right level of AI involvement for the task in front of you — before you open the chat.
Like this post?
Get the next one in your inbox. Practical AI skills, no filler.