| Highlights: |
|
AI has spread through the workplace quickly. Employees use it to summarize documents, draft emails, research topics, analyze spreadsheets, and prepare presentations. According to McKinsey’s 2025 State of AI survey, 88% of organizations use AI in at least one business function.
That figure suggests that AI has already become part of everyday business. Another number tells a different story: only 7% of organizations say they have fully integrated and scaled AI across the company.
The distance between those two figures is where many businesses now find themselves. They have purchased tools, tested use cases, and run small pilots. Some employees may already use AI every day. Even so, the organization has not found a reliable way to turn those activities into wider business results.
This is the AI adoption paradox. Access is becoming common. Consistent value is not.
A Successful Pilot Proves Less Than People Think
Imagine a company testing AI in its customer service department.
A small group of experienced employees use an AI assistant to draft responses to common customer questions. The team chooses a narrow set of inquiries, provides approved reference materials, and checks every answer before it is sent. During the pilot, response times fall and employees report that the tool is helpful. The test appears successful.
Problems may emerge when the company tries to extend the same system to hundreds of employees. Some departments use different customer records. Product information may be outdated or stored in several locations. New employees may accept incorrect answers because they don't have enough experience to notice them. Security rules may prevent the system from accessing information it needs.
None of these issues means the original pilot failed. It means the pilot answered a limited question: Can AI help a small group complete a specific task under controlled conditions?
Scaling raises harder questions. Can the system work with ordinary data? Can employees use it consistently? Does it fit existing security rules? Who is responsible when the answer is wrong? Can the company prove that the tool is improving service rather than simply producing more text?
Small pilots often avoid these complications. Company-wide use cannot.
Scaling AI Means Changing the Work Around It
Many AI projects begin with a tool. A leader sees a demonstration, a department receives access, and employees are encouraged to find useful applications.
That approach can generate ideas, but it rarely changes how the organization operates.
Consider a monthly reporting process. Employees may spend days collecting figures from different systems, correcting formatting, asking departments to explain missing information, and waiting for approvals. An AI tool could help write the final report, but writing may represent only a small part of the delay.
The larger opportunity involves improving how information is collected, deciding which approvals are necessary, and creating a shared source for the data. AI can support that work, but the company must first examine the full process.
This helps explain why workflow redesign appears so often in research on successful AI programs. McKinsey found that organizations reporting the strongest results were much more likely to have changed how work was completed, rather than adding AI to the existing process.
The difference can be easy to miss. An employee may save an hour by using AI to prepare a document. The organization benefits only when it decides how that hour will be used. The employee might handle more customer requests, spend additional time checking quality, or focus on work that previously had been delayed. Without that decision, the time saved may never appear in a business metric.
The same problem applies to goals. “Increase AI use” is difficult to measure and gives employees little direction. A clearer goal might be to reduce the time required to prepare a proposal from five days to three, improve the accuracy of invoice coding, or shorten the wait for a customer response.
These measures give a team something specific to test and make it easier to stop projects that are interesting but not especially useful.
People Usually Determine Whether the System Lasts
AI implementation is often assigned to technology teams, but many of the obstacles appear elsewhere.
Employees may not know which tools are approved. Managers may have different expectations about when AI should be used. One department may allow employees to enter internal documents, while another prohibits it. Some workers may avoid the technology because they don't trust it while others may rely on it too heavily.
Training can help, although a general demonstration is rarely enough. Employees need examples connected to their own work.
A finance team may need to understand how AI can assist with variance explanations while recognizing that the figures still require verification. A human resources team may need guidance on handling employee information. A sales team may need rules for reviewing claims before they appear in customer communications.
Employees also need a clear way to raise concerns. If a tool repeatedly provides outdated information, produces biased results, or adds unnecessary work, people should know where to report the problem. Otherwise, they may quietly stop using it or develop their own workarounds.
Fear of job loss can also affect participation. Employees are less likely to support an AI project when they believe their knowledge is being used to remove their positions. Leaders should explain what is changing, what remains the responsibility of employees, and how roles may develop.
That conversation should happen before a broad rollout. Employees who perform the work often understand its exceptions better than anyone else. They know which customers require special handling, where data is unreliable, and which steps exist for reasons that may not be visible in a process diagram.
Leaving those employees out may produce a system that works during a demonstration and fails during a normal workday.
What the Organizations Making Progress Tend to Do
The small group of organizations reporting substantial financial value from AI doesn't appear to rely on a single model, platform, or technical strategy.
Their advantage seems to come from how they manage the change.
Senior leaders remain involved after approving the budget. They use the tools, ask for evidence, and assign responsibility for the results. Business teams help choose the problems being addressed. Technology, security, legal, and data teams are included early enough to prevent avoidable delays.
These organizations also tend to invest beyond software licenses. They spend money on data preparation, system connections, training, support, and process changes. Those expenses may be less visible than purchasing a new platform, but they often determine whether the platform becomes useful.
They are also more willing to pursue outcomes beyond cost reduction. AI may help a company answer customers more quickly, develop products faster, improve forecasting, or identify sales opportunities. Efficiency still matters, but it's only one possible result.
Most importantly, these organizations are selective. They do not attempt to scale every experiment.
A useful pilot should have a clear owner, a measurable result, and a reasonable path into everyday operations. Before expanding it, the company should know what information the system needs, how its output will be checked, and what employees are expected to do differently.
When those answers remain unclear, wider deployment usually adds confusion rather than value.
Moving From Experiments to Routine Work
Companies that are stuck in pilot mode don't need to begin with a company-wide transformation. They need one useful process that can be improved and measured.
A team might start with customer inquiries, contract reviews, proposal preparation, employee support requests, or internal research. The right choice depends on the business, although the process should occur often enough to study and matter enough to justify the effort.
The team should first document how the work happens today. That includes delays, informal approvals, repeated corrections, and the exceptions employees handle manually. A process that looks simple in a policy document may be very different in practice.
Next, the company should decide where AI will be used and where human judgment remains necessary. Someone must still own the final result – that responsibility cannot be assigned to the system.
The company also needs a baseline:
- How long does the process currently take?
- How often do errors occur?
- How many requests can the team handle?
- What do customers or employees think of the experience?
After the pilot begins, the same measures should be tracked again. A successful result should be visible in the work, not only in employee enthusiasm or the quality of a demonstration.
The final test is whether the process can continue without unusual support. If technical specialists must attend every meeting, repair the data manually, or rewrite most of the output, the system is probably not ready to expand.
Scaling becomes more realistic when the new process works during busy periods, with different employees, and with the imperfect information that exists in ordinary business settings.
The 88% figure shows how easy it has become to start using AI. The 7% figure shows how much work remains after that first step.
Companies that close the gap will likely be the ones that pay attention to the less exciting parts of implementation: reliable data, clear responsibilities, employee involvement, realistic testing, and changes to the way work is organized.
