In 2025, the Boston Consulting Group put a number on a proportion that had already been circulating informally among people running AI programs inside large companies: the 10-20-70 rule. According to the firm's report "The Leader's Guide to Transforming with AI," only 10% of the effort behind a successful AI transformation should go into algorithms and models, 20% into technology and data, and the remaining 70% — the decisive share — into people and processes.
The math is uncomfortable because it inverts what most companies actually do. AI budgets tend to run the other way: most of the spend goes to tool licenses and technical integration, and whatever is left over becomes a few hours of training for the team. Organizations that get this ratio backwards, BCG notes, rarely manage to measure any return on the investment.
That does not mean training the team is pointless. It means training without giving people an official place to apply what they learned solves only a small slice of the problem — and it is precisely the slice that determines the least of the outcome.
In Brazilian and Latin American companies, this distortion tends to show up even more directly: AI training becomes an HR line item or a one-off event, while the technology team keeps treating tool access as a separate matter, solved through individual licenses. The two sides rarely talk to each other, and the result matches what BCG describes — except without even a record of how much was spent trying to fix the problem from the wrong end.
The paradox of training with nowhere to apply it
Deloitte's "State of AI in the Enterprise" report, based on interviews with 3,235 business and technology leaders across 24 countries between August and September 2025, lays out this paradox in numbers. Workforce access to AI tools grew 50% in one year. Yet only 30% of organizations are redesigning key processes around AI, and 37% use the technology only at a surface level, with little or no change to the processes underneath the work.
In other words: far more people have access to AI, and almost nothing has changed in how the work actually gets done. Access grew; process did not.
The same gap shows up on the training side. LinkedIn Learning's "2025 Workplace Learning Report" found that companies with strong career-development programs are 42% more likely to be frontrunners in generative AI adoption, and 32% more likely to be rolling out AI training programs this year. The pattern is consistent: training on its own does not separate the companies pulling ahead from the ones falling behind — what separates them is training paired with a work environment able to absorb what was learned.
When that environment is missing, training turns into scattered knowledge. The person learns to write a better prompt, understands the model's limits, leaves the course motivated — and goes back to the same personal account as before, with no oversight, no shared history, and nobody at the company aware of what she found.
Why training alone does not change the outcome
Training without an official environment has a specific side effect: the company teaches the person to get more out of her own personal AI account. It is a real investment, but it stays with the person, not the company — and it evaporates the moment she changes teams, leaves the company, or simply forgets what she learned months after the course ended.
There is also a measurement problem. When there is no record of what was actually done with AI, the company cannot tell who genuinely applies what they learned from who took the course and never touched the topic again. Without that data, the training program itself cannot be adjusted — the company keeps investing in the same formula without knowing whether it works.
The third problem is the quietest one: what one person discovers stays with that person. A well-built response script, an efficient way to summarize a contract, an agent configured for one specific routine — none of it spreads to the rest of the team unless there is a shared place where it can be stored and reused. The company pays for the whole team's training and collects the knowledge one person at a time.
What has to be in place
An environment that turns training into real capability rests on verifiable mechanisms, not on good intentions.
Corporate identity, not a personal account. People sign in with the same login as the company network, sourced from the corporate directory — and when someone is let go, AI access drops with it, without depending on anyone remembering to revoke a standalone account.
Access by role. What each person can see and do is defined by the role they hold, including inside a connected tool — not all-or-nothing.
Human approval based on risk. Sensitive actions pause, ask a person to confirm, and only proceed afterward. Documents that will become official knowledge for the AI can go through a two-step review and approval before they are put to use.
Audit trail. Who created what, who approved it, when an agent was used: all of it recorded and searchable, without relying on anyone's memory.
Reuse with permission. What one person builds — an agent, a script, a workflow — can be made available to other people or groups, with reach defined by role and permission, instead of dying in a single account's history.
This is how Skyller was designed: identity coming from the company directory, permission by role, and reuse of agents and scripts as the platform default, not an advanced feature.
From individual knowledge to collective capability
When that environment exists, training changes function. It stops being the end of the process — "the team took the course" — and becomes the start of a cycle: a person learns, applies it inside an environment with identity and permission, what works gets recorded, and someone else on the team reuses it without learning from scratch.
It is a similar effect to the one LinkedIn Learning measures when it shows that companies with a strong career-development culture pull ahead in generative AI adoption: training does not act alone, it acts inside a system that absorbs and distributes what was learned. Without that system, every new training round starts from zero, because the previous one never left a trace anywhere.
For whoever owns the budget, this is the difference that actually matters: training without an environment is a recurring cost with no accumulation. Training with an environment is a cost that turns into capability, because what one person learns first makes learning cheaper for whoever comes next.
That logic also changes how training itself gets measured. Instead of counting how many people completed the course, the relevant question becomes how many agents, scripts, or workflows built by one person got reused by someone else on the team — a number that only exists when there is a shared environment recording what happens.
A roadmap to get started
Before approving another AI training budget, these questions are worth putting on the table:
- Where will the team apply what they learn in the course, as early as tomorrow morning? If the answer is "in each person's own account," the training leaves with each individual's memory.
- Does what one person discovers with AI today reach anyone else on the team, or does it stay only with the person who found it? Without a place to reuse it, every new training round starts from zero.
- If someone who was trained leaves the company tomorrow, what is left? If the answer is "nothing," the investment never stayed with the company — it stayed with the person.
- Can the company measure who applied the training and who merely attended it? Without that record, the next training budget repeats the same bet blind.
