Lessons From Scaling Enterprise AI

In 2025, the adoption of enterprise AI reached a critical inflection point. On the one hand, artificial intelligence reached a sort of critical mass, with nearly 88% of organizations reporting that they regularly use AI, according to a November 2025 McKinsey report. On the other hand, most of those organizations are still mired in pilot-stage experiments, with only one-third stating that their companies have begun rolling AI programs out at scale, McKinsey said.

Therein lies a major gulf between AI’s potential and its reality, suggests the MIT Center for Information Systems Research. In an August 2025 research brief, it argued that organizations realize the greatest financial impact from AI when they move from early-stage AI experimentation to developing repeatable, scaled ways of working across the enterprise.

Exasperating AI’s unrealized value is a widespread skills and literacy gap, according to Larradin, whose platform measures and monitors enterprise AI adoption. In its “The State of Enterprise AI in 2025” report, it found that 73% of knowledge workers report using AI tools weekly. And yet, only 29% of the same workers rate their AI literacy as “advanced.”

Clearly, organizations pursuing AI excellence still have a long journey ahead. But as they navigated the widespread challenges of scaling AI in 2025, IT leaders gained critical knowledge that will help them accelerate AI adoption and maximize its impact in 2026, 2027 and beyond. AI analysts, advisors and practitioners shared with The Forecast the most important lessons that they learned so far.

AI Adoption Dies Without Executive Sponsorship

Because legal and IT departments are created to manage risk, they’ll never push AI adoption of their own accord, says Meryll Dindin, vice president of product and engineering at Parallel Learning, a K-12 teletherapy platform that uses AI to enhance its service offerings.

Indeed, Dindin’s company has learned firsthand that a business sponsor — someone with strategic oversight and budget authority — must own AI and drive its adoption across the enterprise. Even more important, sponsors must be willing to have uncomfortable conversations with department heads who confuse inertia with real security concerns, Dindin told The Forecast.

Many executives issue mandates to use AI. But that doesn’t provide the support needed for an actual operations model, argues Arvita Tripati, founder and managing director of the AI consulting company Vahana Labs.

“That dead zone between ‘we need to use AI’ and ‘here’s how we actually run AI inside this organization’ is where most companies stall,” Tripati told The Forecast in an interview. 

“They buy tools, run a pilot, get a win and then can’t replicate it because they never built the process layer: who reviews outputs, how exceptions get handled, what the escalation path looks like when something goes sideways. I call this ‘pilot purgatory.’ The companies that scale past it are the ones that treat AI adoption as an operations problem, not a technology problem.”

Governance Requires More Than an AI Policy

Tripati said almost every organization she works with has a written AI policy, but very few answer the harder question of who owns the decision when the AI output is wrong. Case in point: She recently worked with a mid-market healthcare company whose executive team had approved an AI strategy, but had no roadmap for actually executing it. 

“No one could tell me who was accountable for model performance once it was in production,” Tripati said. “They had a policy doc. They didn’t have an operating model. The fix wasn’t more documentation; it was assigning clear ownership at the workflow level, not the org-chart level.”

Dindin’s organization learned about governance the hard way when team members across several departments began using an unauthorized meeting notes tool, which was especially problematic given the organization's highly regulated industry. In response, company leaders formalized clear boundaries by approving tools with defined data-flow rules, security scoping and escalation paths.

“We built role-based access control directly into our internal AI systems, so a sales rep querying our data warehouse from Slack sees different tables than someone in clinical or finance. Every AI query gets logged with the user, the model version and the sources it touched,” Dindin told The Forecast. 

“When someone asks for restricted data, the system redirects them to a safe equivalent instead of refusing outright. That keeps trust intact.”

Organizations Should Focus on Literacy Instead of Training

Many organizations give top-down mandates to use AI, which breeds resentment when they come without the proper context and support, Dindin noted. Instead of training, Parallel Learning uses shadowing, where team members sit with each other during a real workflow and learn how to use a tool to cut off manual work in their actual workflow. Learning tied to a real pain point converts skeptics faster than any training deck, the company has found.

“One of our operations leads used to wake up at 4 o’clock in the morning for manual documentation. After a single session demonstrating voice-to-text for SOP creation, she became our most vocal AI advocate and now trains others on her team,” Dindin said.

One of Tripati’s clients used a similar approach by replacing their quarterly AI awareness sessions with a 30-day sprint. Instead of check-the-box training, an internal champion who is available for questions guides a single department through using AI tools on actual work throughout the sprint. Her client saw adoption move from “performative” to “functional” in a matter of weeks.

Embedding AI Directly Into Workflows Increases Adoption

AI pilots and proofs of concept succeed because the early adopters involved in them tend to be motivated and enthusiastic about AI, according to Barbara Roos, founder of Trailhead Communications, a consulting firm that focuses on the human side of AI. But when organizations begin scaling AI pilots across the enterprise, employees often are overwhelmed with other work, anxious about new technology or even skeptical.

“Companies of all sizes and across all industries are seeing programs fail between proof of concept and scaled adoption specifically because the human dimension wasn’t factored in,” Roos told The Forecast in an interview. 

“Building cross-functional ‘change champion’ networks, which include both enthusiasts and skeptics, is one of the most effective structural interventions I’ve seen for bridging that gap.”