In 2023, the National Bureau of Economic Research (NBER) published a study tracking 5,179 customer support agents. Half had access to a generative AI assistant integrated into their workflow; the other half did not. The result was clear: those who used AI became 14% more productive, measured in issues resolved per hour. The study also recorded improvements in customer sentiment and in agent retention.
It was one of the first concrete numbers on AI productivity at real work scale. And it raised an obvious question: if 14% is the average per-person gain, why do so many companies that adopt AI at scale not see it reflected in quarterly results?
The same study already offered the first clue. The 14% average concealed a wide gap between profiles: novice and lower-performing workers gained 34%, while the most experienced ones gained almost nothing. AI was not distributing its gain evenly — it was spreading the best practices of the already-strong performers to those who weren't there yet.
A second study, run by Harvard Business School with Boston Consulting Group, made the same point from another angle. Researchers gathered 758 consultants, gave some of them access to generative AI tools, and measured performance across 18 realistic consulting tasks. Within what AI does well, the gain was large: 25% faster, 12% more tasks completed, and quality rated 40% higher. And again, the gain was uneven: the bottom half of performers improved 43%, against 17% for the top half.
But the most uncomfortable finding came when the researchers deliberately designed a task outside AI's frontier of competence. There, those using AI were 19 percentage points less likely to get it right than those working without it. The authors named the phenomenon the "jagged frontier": AI is excellent on one side of the line and harmful on the other — and the line is written down nowhere.
The real gain versus what the company sees
In May 2024, Microsoft and LinkedIn published the Work Trend Index, a survey of 31,000 knowledge workers across 31 markets. The headline finding that circulated: 75% already used AI at work. The secondary finding, less publicized, was more interesting: there is a group of intensive users — people who use AI several times a week and have reorganized their own routine around it — who save more than 30 minutes a day compared to skeptics. That's 2.5 hours per week. 12.5 hours per month. Roughly 150 hours per year.
If a team of 100 people reached that level, it would mean 15,000 hours of work freed up per year. In cost terms, that's the difference of several full-time people. The problem is that the group who gets there is a minority — and got there on its own, each person their own way.
That's where the math stops working. McKinsey surveyed 1,719 executives in 2026 and found both sides of the same coin: nearly 9 out of 10 organizations already use AI regularly in at least one function, and 80% of respondents say AI has improved individual productivity. But when the question shifts to company results, only 37% report meaningful impact on earnings before interest and taxes (EBIT) — a number unchanged from the year before. The group attributing 5% or more of EBIT to AI is about 6% of respondents. And what sets that group apart isn't the tool: it's having redesigned how the work happens.
Gartner reached a similar conclusion by another route. In a survey of 724 respondents, only 34% of teams using generative AI reported high productivity gains — slightly worse, in fact, than teams using traditional AI, at 37%. It's not that AI doesn't work. It's that working locally is not the same as working globally.
And there are cases where the apparent gain is the opposite of the real one. METR, a nonprofit research organization, followed experienced developers in 2025 working on open-source repositories they maintained themselves. Before starting, they expected to be 24% faster with AI. The measured result was 19% slower. And the detail that should trigger a red light for any executive: even after being slower, they still believed AI had sped them up by 20%. Among the contributing factors is precisely the time spent steering the assistant and reviewing what it produces — work that exists, consumes the day, and shows up nowhere.
Why the obvious solution doesn't work
When the company president asks IT to "put AI everywhere", what they usually do is: buy a per-person license, publish training, and hope. Each person gets their own personal account in the AI tool. Some discover good ways to use it. Most follow protocol.
The problem is that knowledge stays locked in the person. The analyst who spent weeks building the perfect spreadsheet analysis routine — the right columns, the ideal prompt, how to handle exceptions — is the only one who has it. When they leave the company, the routine leaves with them.
And more: when 100 people use AI in 100 different accounts, nobody knows what works at scale. There are 100 ways to do the same thing, and each person has to work out alone where that jagged frontier sits — which tasks AI helps with and which ones it hurts. Nobody is reusing anything: not the wins, and not the lessons from the failures.
What has to be in place
An environment where individual gain becomes company gain needs one specific mechanism: shared agents, prompts, and workflows with the right access.
Reusable agents, with permission by role. When someone designs an agent that handles tax collection analysis, that agent can be made available to others in finance — not to the entire company, only to those with a collection function. And not with "access to everything", but only the access they need based on their role. The person running the agent doesn't build a tool from scratch; they reuse the one that was approved and tested — one that already carries the knowledge of where AI helps and where it gets things wrong.
Documented and versioned prompts and workflows. The prompt that became the standard for report summaries stays on record, it doesn't die in someone's head. When a law changes, the prompt changes once, for everyone. Nobody is reinventing the wheel.
Knowledge that stays with the company. The conversations, the agents, the workflows — all knowledge created stays with the company, not trapped in a personal account. If the analyst leaves, the company doesn't lose their work.
Shared credits, not per-person licenses. When everyone has their own license sitting idle, the company pays for unused capacity. When they share an AI budget by area or department, whoever needs more uses more. Consumption stays transparent — it's the lever that lets you measure real gain.
This is how Skyller was designed: the gain one person discovers is immediately usable by their team, because agents, prompts, and workflows are shared within the access controls of each role.
From individual gain to team gain
The numbers are clear: using AI well delivers real gain. 14% in call centers, 25% more speed in consulting. The problem is not the technology. It's how the company organizes it — and that is exactly what separates the 6% who capture results from the other 94%.
When each person has their personal account, the gain stays scattered and invisible. The professional gets measurable savings at their desk. But the company can't see whether it saved 5% or 25% for real — nor whether, in some area, it is actually losing time. And when someone leaves, they take the knowledge with them.
The right architecture doesn't have to be complex. It needs three things: reuse with permission, knowledge that stays, and real measurement by area. When these three exist, the gain stops being a silo phenomenon and becomes a team result.
What changes: a department of 10 people that saves 1 hour a day, but nobody knows how, is worthless. Ten people that save 30 minutes a day using a reusable, documented agent — one that can be expanded to the branch or to the parallel team — that's a gain the company can replicate. And measure.
Three questions for the next meeting
Before buying more AI licenses, it's worth answering these three questions with IT and the CFO:
-
If an analyst discovers an excellent way to use AI for price analysis, how does the company ensure the next analyst doesn't start from scratch? If the answer is "it stays documented and shared", great. If it's "everyone discovers their own way", the problem is identified.
-
How do you measure today how many hours AI actually saved by department? If there's no metric, there's no visibility — and without visibility, there's no way to know if AI is worth the investment, or where it is costing time instead of saving it. Shared credits by area create that transparency immediately.
-
If someone leaves, what agents and workflows they created stay with the company? If the answer is "none", it's clear that knowledge leaves with the people.
