Inside the Quiet Revolution of Enterprise Artificial Intelligence

For the majority of the past three years, the story of AI has been told through keynote addresses, such as those given by Sam Altman on a stage in San Francisco, demonstrations of chatbots composing sonnets, and rounds of funding from a small number of frontier labs. The narrative unfolding within corporate IT departments has been more subdued, slower, and, in some respects, more significant. It’s the kind of shift you only become aware of when you start looking at the spreadsheets instead of the keynotes.

The back office of a mid-sized pharmaceutical company, where someone is matching invoices to purchase orders, is a helpful starting point. That was a person a year ago. These days, a software agent typically reads the invoice, verifies the contract, highlights the disparity, and only sends the exception to a human when something truly unusual occurs. According to Deloitte’s 2026 enterprise survey, 78% of organizations use AI in at least one function, up from 55% the previous year. In business technology, such a leap is uncommon. It typically indicates that a tool has transitioned from curiosity to infrastructure.

The shift is most noticeable and unsettling in software development. In addition to autocompleting, tools like GitHub Copilot Workspace, Cursor 2.0, and OpenAI’s Aardvark plan, act, and verify. Large companies are hesitant to release precise numbers, in part because their own engineers disagree, but estimates of the percentage of production code that is now AI-assisted range from 30 to 50 percent. Senior developers believe that the work has become denser rather than easier. The human makes more decisions every hour, while the agent deals with the boilerplate. According to Brent Orrell of the AEI, this is “leverage, not magic,” which seems more accurate than most statements made about it.

Of all of this, the pharmaceutical sector may be the cleanest case study. With 36 ongoing Phase 3 trials, the 150-year-old Indianapolis-based Eli Lilly reported $65.2 billion in revenue last year. It has invested about $55 billion in manufacturing since 2020 and established a co-innovation lab with Nvidia. That is not a story about a tech startup. It tells the tale of a legacy company that, rather than viewing AI as a product, quietly saw it as plumbing. 2026 will be the first year with enough AI-assisted trials in the final stages to even begin answering the question of whether the technology can finally overcome pharma’s stubborn 90 percent clinical-trial failure rate.

Then there’s SAP, the massive German enterprise software company that most customers couldn’t choose from. RPT-1, a “relational pre-trained transformer,” pronounced “rapid one,” was introduced in late 2025. It was trained on the structured tables that actually run businesses, such as ledgers, payroll, and supply-chain logs, rather than on language. If you’re accustomed to chatbots, it looks weird. Additionally, it solves the most tedious and costly issue in corporate AI, which is that businesses have spent decades attempting to train limited models to make poor predictions one at a time.

As this develops, it’s difficult to avoid feeling that the public discourse has been slightly misguided. The attention is drawn to the eye-catching demos, while the long-lasting changes take place in poorly filmed areas. There are still a lot of things that could go wrong, such as the lack of strong governance, the fact that corporate use cases don’t make hallucinations any less problematic, and the fact that only roughly one-third of leaders surveyed claim to be genuinely reimagining their business rather than merely adding AI to their current workflows. As it is, the revolution resembles a gradual shift in the weather rather than a thunderclap. The majority of its occupants are still undecided about whether to be anxious or hopeful. Most likely both.