The 3.5× Gap: Why Frontier Firms Use More AI Per Employee, With the Same Models
OpenAI's recent B2B Signals research surfaced a number that has been hard to shake for anyone working inside enterprise AI. Frontier companies use 3.5× more AI intelligence per employee than the typical firm. The instinct on first reading is to assume frontier firms simply spend more on AI, but the data does not support it. Frontier and typical firms have access to the same models, often the same vendors, often comparable annual spend. The intelligence-per-employee gap opens up before the procurement line item is reached. It opens up at the workflow design layer.
We see the same shape inside our deployments. Two companies, comparable size, comparable budget, same model vendor, very different intelligence-per-employee output six months in. The difference is never the model. The difference is what the company asked the model to do, and how the surrounding workflow was redesigned to let the model do it.
This piece walks through what the frontier firms appear to be doing differently, why the gap compounds rather than flattens, and what the catch-up move looks like for Greek enterprises in 2026.
What "intelligence per employee" actually measures
The phrase sounds abstract. Operationally it is concrete. Take the volume of cognitive work an organisation produces in a quarter (contracts reviewed, tickets resolved, drafts generated, analyses run, code shipped, decisions made). Divide by headcount. That is intelligence per employee at the analogue baseline. Now layer in the AI-assisted volume produced by the same headcount in the same quarter, against the same quality bar. The ratio of those two numbers, before and after AI, is what OpenAI's research is tracking when it talks about an intelligence multiplier.
Frontier firms are running at 3.5× the typical-firm multiplier. The same employees, with the same week, are producing 3.5× more reviewed work product. That is the gap, and it shows up on the P&L as higher revenue per employee, faster time to market, lower cost to serve, or all three.
Why the gap opens
The starting question matters more than the spend.
Typical firms tend to start from a position of "we already do X, where can we add AI to make X faster?". The natural output of that question is AI features bolted onto an existing process. The process boundaries do not move, the tooling does not change, the headcount allocation stays the same. The AI shows up as a productivity assist inside a workflow that was designed for humans only. The lift is real but capped, typically in the 15 to 30% range, because the underlying process is the constraint.
Frontier firms start from a different question. "What does this model do natively better than a human, and what work can we route to it as the primary owner?". The natural output of that question is a new workflow design, with the AI as a first-class actor rather than an assistant. Process boundaries move. Tooling changes. Headcount is reallocated upward into oversight, design, and exception handling. The lift is uncapped, because the workflow was designed around the AI's capabilities rather than retrofitted onto an existing one.
Both groups are using the same models. Both are paying roughly the same per-token. The intelligence-per-employee divergence is entirely on the workflow design side. This is the part the model evaluations cannot capture, and it is the part that matters.
The compounding mechanism
The 3.5× gap is not a snapshot, it is the result of a compounding loop. Frontier firms get more leverage per dollar of AI spend, so they reinvest more, so they get more leverage, so they reinvest more. The loop tightens. Typical firms get a smaller initial lift, see a smaller ROI line in the next budget cycle, and reinvest less. Their loop loosens.
Three years of this divergence is the difference between a company that has rebuilt its operating model around AI and a company that has automated a handful of internal tasks. The frontier firms will look like genuinely different operating beasts by 2028. The typical firms will be running improved versions of their 2024 process maps.
There is a window in which the gap is closeable. We think that window is roughly 18 months from today. After that, the compounding will have moved the frontier far enough ahead that catch-up requires a structural reset rather than a workflow redesign.
The catch-up move for Greek enterprises
We work with companies that sit on the typical side of the gap and want to move toward the frontier inside the next 12 months. The playbook is consistent.
Start from capability, not from the existing process
The first conversation is not "what do we want to automate". It is "what does this class of model do natively well", which we can answer for any given workflow because we have seen the failure modes and the success modes across thirty-plus deployments. The output of that first conversation is a candidate workflow where the model's native strength matches a process bottleneck the business cares about. The absorption gap article goes deeper into why this matters for Greek market specifically.
Rebuild one workflow end to end, do not retrofit five
Picking five workflows and bolting AI onto each is the classic typical-firm pattern. Picking one workflow and rebuilding it around the AI is the frontier pattern. The first approach produces 1.2× on five workflows. The second produces 3× on one. The math is decisive. The hard part is the discipline of saying no to the other four until the first one is real. Most enterprises lose this argument internally, which is one of the structural reasons the gap exists in the first place.
Reallocate headcount upward, do not just save hours
The frontier firm move on labour is not "AI freed up two FTEs, we redeploy them on lower-value tasks". It is "AI freed up two FTEs, we redeploy them on the design layer of the next workflow we are about to rebuild". The intelligence-per-employee number compounds because the freed capacity goes into building more leverage, not into running more tickets. This is the move that requires a CEO conversation, not a vendor conversation.
Measure intelligence per employee, not cost per task
Cost-per-task is a typical-firm metric. It pushes optimisation toward cheaper execution of the same work. Intelligence-per-employee is a frontier-firm metric. It pushes optimisation toward more cognitive output from the same team. The numbers move in roughly opposite directions on the dashboards, and the strategic implications are nearly opposite. Pick the metric that matches the kind of company you want to be in 2028.
The Greek angle on this gap
Greek enterprises have two structural advantages on this specific problem. First, the smaller average headcount means workflow redesigns can be executed by a single owner in a single quarter, instead of negotiated across multiple departments over a year. Second, the relationship-based decision culture allows the CEO conversation about reallocating freed-up capacity to happen in one meeting rather than five. Both advantages compress the time-to-frontier from years to quarters, if the leadership team chooses to spend that compressed time on workflow redesign rather than on incremental automation.
We have watched several mid-sized Greek companies move from typical to near-frontier inside a single year by following this playbook deliberately. The pattern is replicable. The companies that move now compound their advantage through 2027. The companies that delay accumulate retrofit debt that gets harder to pay down each quarter.
What to do next
Two questions worth taking into the next executive meeting. First, what is one workflow inside our business that we could plausibly let an agent own end to end, if we redesigned the surrounding process to make that possible. Second, if AI freed up two FTEs from that workflow next quarter, where would we redeploy them to compound the leverage rather than redistribute the workload.
If both questions have clean answers, you have a frontier-firm move on your hands. If either one is fuzzy, that fuzziness is exactly where the 3.5× gap is hiding inside your organisation.
We help enterprises identify which workflow to start with, redesign it around the agent rather than around the existing process, and put the right oversight layer on top so the deployment survives its first quarter. The agents we deploy (AI IR Assistant, AI Disclosure Co-Pilot, Enterprise AI Search, AI-Powered CRM, AI Customer Support, AI Contract-to-Cash) each replaced an existing typical-firm workflow rather than augmenting one. If 2026 is the year you stop bolting AI onto existing processes and start rebuilding the processes around AI, get in touch at inbusiness.gr.