In 2023, researchers at Harvard Business School brought together 758 consultants from Boston Consulting Group and split the group in two: half with access to GPT-4, half without. Everyone received the same 18 realistic consulting tasks. The result, published in the study "Navigating the Jagged Technological Frontier": those who used AI on tasks within the tool's reach worked 25% faster, completed 12% more tasks, and delivered results rated 40% better.
A second study, from the National Bureau of Economic Research, measured the same kind of gain in a different setting: 5,179 customer support agents, with staggered access to an AI assistant. The average was 14% more issues resolved per hour. The most telling detail isn't the average — it's the distribution: novice or lower-performing agents gained 34%; the most experienced ones, almost nothing.
Two studies, two industries, the same message: AI multiplies the performance of people who already know how to use it well. The question that remains for decision-makers is a different one. If the individual gain is so clear and so well documented, why does it almost never show up in the company's overall results?
The uneven gain nobody sees
The Harvard and BCG researchers themselves gave the problem a name: the jagged frontier. AI is excellent on one side of the line and harmful on the other — and that line is written down nowhere. When researchers deliberately tested a task outside the tool's reach, consultants using AI performed 19 percentage points worse than those working without it.
The customer support study shows the same phenomenon from another angle. The researchers describe the mechanism behind the 34% jump among novice agents: AI was spreading the working style of the most experienced agents to those who hadn't gotten there yet. In other words, the gain didn't come from the tool alone — it came from one person's knowledge being transferred, through the tool, to another person.
That changes the question every company should be asking. It isn't "does AI work?" It's "is the knowledge of whoever already found the right way being transferred to the rest of the team, or is it stuck in one person's account?" Most companies still don't have an answer, because they've never put the question into practice.
The pattern repeats in companies across Latin America, and it tends to show up more sharply in groups with several units or branches operating independently. One branch discovers a good way to summarize contracts or triage support tickets; the branch next door, doing the same kind of work, never finds out. Without a formal path to carry that forward, the good discovery stays with whoever made it — and the company keeps paying the same learning cost, unit by unit.
Why the gain never reaches the company's results
Boston Consulting Group tried to answer that question at scale, looking not at individuals but at entire companies. Research published in 2025 found that only 5% of companies are what the consultancy calls "future-built" — organizations that already turn AI use into measurable financial results: 5 times more revenue growth and 3 times more cost reduction than average.
On the other side, 60% of companies invest in AI and get little or no material value in return. It isn't a lack of people using the tool — many of those companies have high individual adoption. What's missing, according to the analysis, isn't the technology. It's turning isolated initiatives into organizational capability: taking what one person discovered alone and making it work for everyone else, systematically.
That's exactly the pattern from the two earlier studies, just at company scale. When 100 people use AI in 100 separate accounts, each one reinvents their own script on their own — the right way to ask, the right sequence of questions, how to handle exceptions. One or two find something very good. And that something dies with them, because there's no path to publish it for the rest of the group.
What has to be in place
Closing the gap between individual gain and company results requires a specific mechanism: reusable agents, scripts, spaces, and workflows, with reach defined by person, group, and role.
A reusable agent, with permission by role. When someone builds an agent that handles collections analysis well, for instance, that agent can be made available to the rest of the finance team — not the whole company, only those with that function. Whoever receives the agent isn't starting from zero: they're reusing something already tested, one that carries within it the knowledge of where AI helps and where it gets things wrong.
A documented script, not lost in someone's head. The way of asking that became the standard for summarizing a type of document stays on record, with an owner and a version — it doesn't die in the personal history of whoever discovered it. When a rule changes, the script changes once, for everyone.
The right reach — neither the whole company nor just one person. Not everything needs to become the whole company's asset, and not everything should stay locked to one person. The middle ground — my group, my area, my role — is where most real reuse actually happens.
A ready-made process so nobody starts from zero. The company shouldn't have to wait for someone to discover their own way of handling a common routine. A catalog with more than 170 ready-made policy and process templates shortens that path before the first attempt even begins.
This is how Skyller was designed: what one person discovers becomes available to the right group, with the right permission, instead of dying in a conversation only that person can reopen.
From individual experience to shared capability
Go back to the numbers from the start. One consultant, working 25% faster, is good news for that consultant. One novice agent jumping 34% in productivity is good news for that agent. Neither one, by itself, changes the company's quarterly results — because the gain is contained in a single person, in a single work session.
What changes the results is the same gain, replicated. Not through the effort of repeating the discovery at every desk, but because the path that worked was published once and reused many times. A team of ten people with a shared, tested agent that has a clear owner doesn't depend on ten people finding the same solution on their own. It depends on one person having found it, and on there being an official place to make it available to the other nine.
That's the difference between the 5% BCG calls future-built and the 60% that still see no return. It isn't access to the technology — that's already widely distributed. It's whether what works at one desk can actually leave it.
Three questions for the next meeting
Before buying more licenses or announcing more training, it's worth answering these questions with operations leadership:
- Has anyone on the team already found a way to use AI that saves real time? If the answer is yes and nobody knows who that person is, the problem isn't technology — it's visibility.
- Is that script somewhere another person can open and use tomorrow, without asking the one who created it? If the answer is no, it exists, but it isn't available to the company.
- Who else, with the same role, would benefit from the same agent or the same script? Defining that reach — person, group, role — is the step that turns an isolated discovery into shared capability.
- Is there a ready-made process the company could use today, instead of waiting for someone to reinvent it? Most common routines already have a tested template; the real work is usually adapting it, not building it from scratch.
- When that person goes on vacation — or leaves the company — does the knowledge go with them? If the answer is yes, the gain was never the company's. It was borrowed.
