There is a place where corporate AI projects tend to stop: pilot purgatory. The test is approved, the demo impresses, the board applauds — and none of it reaches the operation. Months later, nobody can say what actually changed in the numbers.
The most quoted figure on this came out of MIT. The GenAI Divide: State of AI in Business 2025, published by Project NANDA in July 2025, reviewed more than 300 publicly disclosed AI initiatives, interviewed representatives from 52 organizations and surveyed 153 senior leaders. Its conclusion: despite US$ 30–40 billion in enterprise investment in generative AI, 95% of organizations are getting zero return. Only 5% of integrated pilots extract meaningful value.
The report is blunt about what matters to decision makers: the divide is not driven by model quality or by regulation. It is determined by approach.
The finding holds across the wider market. A survey by S&P Global Market Intelligence covering more than a thousand companies in North America and Europe, reported in March 2025, found that 42% of them scrapped most of their AI initiatives that year, up from 17% the year before. The average organization killed 46% of its proofs of concept before they reached production.
For a company in Brazil or elsewhere in Latin America, the message is direct: the risk is not being left out of AI. It is spending time, budget and internal credibility on tests that never become company capability.
The pilot does not die because of the model
The MIT report shows exactly where the math breaks. General-purpose chat tools are widely adopted: more than 80% of organizations have explored or piloted something like ChatGPT or Copilot, and roughly 40% report some form of deployment. But those uses raise the productivity of individuals — not the performance of the business.
When a company attempts the next step, the funnel narrows sharply: 60% of organizations evaluated enterprise-grade systems, only 20% reached the pilot stage, and just 5% reached production. The reasons cited are always the same three: brittle workflows, missing context, and misalignment with day-to-day operations.
Translated: the pilot dies because the AI has no idea where it is. It does not know how the company works, which documents count, where the work actually happens, or who is asking for what. A brilliant, uninformed assistant is still an uninformed assistant.
Three disconnections that kill a pilot
1. The right knowledge is not there
In the demo, someone attaches two files and the answer comes out perfect. In the operation, the question depends on a policy revised last month, a procedure that changed, and a table only one team understands. The content exists, but scattered — and mixed with old versions nobody retired.
Without a rule for what may become official knowledge, AI answers with whatever it finds. Knowledge intake has to be treated as a process, not an upload: a critical document goes through review and approval, with a named owner, before it informs any answer.
2. Processes and systems are left out
A pilot that only chats produces text. A project that moves the numbers has to reach the place where the task ends: the ticket that gets opened, the order that gets updated, the spreadsheet that feeds the monthly close. As long as AI lives apart from the operation, every gain depends on someone copying and pasting — and that someone gives up in week three.
3. Nobody knows who is asking
The question "was that person allowed to see this?" has no answer in a tool that does not know the company's structure. Without an answer, the project is halted on security grounds — correctly. Many pilots ended that way: not because they failed, but because there was no way to say who saw what.
The pilot that worked also needs somewhere to go
There is a second, quieter and equally expensive way to fail. Someone on the team finds a genuinely good use — drafting the proposal, reviewing a contract, summarizing the weekly report. It works. And it stays in that person's private history. When they move to another team, the learning leaves with them.
Individual productivity does not accumulate on its own. It becomes company capability only when there is a path for what worked to be reused by the people entitled to use it: the agent, the prompt or the flow leaves individual use and gets a reach defined by people, groups and roles. That is one of the central ideas behind Skyller's design.
Reuse is not opening everything to everyone: personal memory stays personal, and documents and integrations keep their own scopes and credentials.
What has to be in place
The difference between a test and an operation is not the model chosen. It is an environment that brings knowledge, processes and systems together with clear rules about who accesses what. That calls for mechanisms a loose pilot does not have:
- Knowledge with an owner and a shelf life. A critical document goes through review and approval before guiding answers and agents, and the answer shows what it drew on.
- Genuine corporate identity. People sign in with their network account, and an offboarding in the company directory carries through to access.
- Function-by-function access. In an integration with dozens of functions, releasing the two a team needs without handing over the rest.
- Approval matched to risk. Sensitive actions stop and ask for human confirmation before they happen.
- An audit trail. Creating an agent, approving a document, changing a permission: what happened stays reconstructable.
- Reuse with a defined reach. What worked for one person becomes team capability, without opening what should stay closed.
That is how Skyller was designed: approved knowledge, corporate identity, permission by function, approval matched to risk, and a record of what happened, in one environment.
Governance is not the brake — it is what lets you accelerate
Market data points the same way. Deloitte, in its State of AI in the Enterprise research with 3,235 business and IT leaders across 24 countries, found only 21% of organizations with mature governance models for AI agents — while adoption of those agents runs well ahead of the controls.
Returns follow the same curve. BCG, in The Widening AI Value Gap, published on 30 September 2025 with 1,250 executives across nine industries, classified 60% of companies as laggards reporting minimal revenue and cost gains, and only 5% as capable of generating value at scale. Another 35% are scaling up and starting to see results.
Read together, the picture is uncomfortable and useful: companies without control do not scale, and companies that do not scale capture no value.
A route out of pilot purgatory
Crossing that line does not require a transformation programme. It requires changing the question. Instead of "which AI tool should we test?", ask "which of our work will now be done differently?".
- Pick a real process, not a demo. One with volume, an owner and a number attached — response time, rework, closing deadline. If nobody can name the number that should move, it is not a project yet.
- Decide who accesses what before you switch anything on. Which documents count, who may consult them, which actions require human approval, and who answers for them.
- Measure where the pain lives. Compare the same indicator before and after, on the same terms. User praise is a signal; it is not a result.
- Turn what worked into team capability. A good individual use should become a reusable agent, prompt or flow with defined permissions — otherwise you will run the same pilot again in the next department.
- Only then repeat. One process at a time, with the same standard of knowledge, permission, approval and record.
Three questions to bring to your next AI meeting:
- Which process in our operation will be measurably different — and which number proves it?
- If someone asks today who saw which information and who approved which action, do we have the answer?
- Is what our best person discovered with AI last quarter already available to everyone entitled to use it?
