Deliberate AI Adoption Creates The Evidence Behind Smarter Investment Decisions

Enterprise AI spending has produced more announcements than measurable results. Eighty-eight percent of organizations now use AI in at least one business function, while just 39 percent report earnings impact at the enterprise level. Explanations for the gap usually name model quality, budget, or regulation. More of it comes from decisions made before a single user touches the system, and from whether those decisions produce evidence a company can use on its next deployment.

As Partner and Head of the AI Practice at Simpson Thacher & Bartlett, Matt Kelly advises public and private companies on AI governance, cybersecurity incidents, regulatory investigations, and technology-driven transactions across financial services, healthcare, private equity, and Fortune 500 technology organizations. He's also one of a small group of US attorneys ranked by Chambers in both artificial intelligence and cybersecurity. Throughout his career, Kelly has helped dozens of clients build digital risk controls and model governance frameworks, work that turns on the same question of what to build first and what to build after that.

"Often, you get started on the next thing before you're done with the first. But if you see that first choice through to completion and learn from the experience, every choice you make from that point forward can be a better one," says Kelly. In his view, what a company learns from a first deployment carries into every decision that follows, and that learning can be worth more than the deployment itself. With a new technology that is meant to transform the way we work, having too many projects running at once produces nothing to examine, because none of them are finished. Without allowing the first project to reach a result, and then taking a moment to reflect on what worked and what didn’t, a leader making a second decision has evidence to work from and is bound to repeat the same mistakes or forgo the same opportunities.

The cost of moving first

Pressure to deploy started with the public release of ChatGPT, and it came framed as a question of survival. Companies that had been running technology projects for years repositioned them as AI initiatives and talked about them loudly. Some of those claims went far enough that they  drew the attention of the SEC and other regulators, who became concerned that firms might be overstating what their systems could do. But executives and directors reading the headlines or sitting in the audience had no way to pressure test their competitors’ claims or to get a realistic benchmark about use of the technology. "Nobody was raising their hand saying, 'We are nowhere,'" says Kelly. "So everybody just assumed that everyone else was further ahead than they were, and they came to fear for their survival if they didn’t join the first wave of adoption." 

Kelly resisted this survival framing for a while, partly because of his observation that companies hadn’t actually deployed AI as widely as the headlines assumed. And those that had deployed the technology widely also were not reaching anything that looked like escape velocity. This trend of lackluster near-term outcomes continues, as more than 40 percent of agentic AI projects are now reported to be on track for cancellation by the end of 2027, with escalating costs, unclear business value, and inadequate risk controls named as the causes.  

But Kelly also rejects the idea that technology adoption is a zero-sum game in which companies either move first or get wiped out. "No one's really been able to point me towards a historical antecedent in which an established company was locked out of a market because they were not the first to move to adopt a new IT or technology advancement," Kelly notes. "Of course, if they never moved at all, that’s a different story. It’s very hard to survive if you don’t evolve. But the assumption that you needed to move early or risk extinction, I just haven’t seen anything to support it."

Companies that came to the internet years after their competitors still built businesses on it. Among law firms and financial services companies, the ones now recognized for handling technology well were rarely the earliest movers. "You didn't need to be the first to adopt internet technology to get value out of it," explains Kelly. "The front runners tended to be overtaken in time by people who later adopted the same technology in a more strategic, thoughtful, and deliberate way."

Picking the first project

Sequencing starts with a sorting exercise in which projects are assigned expectations for value, risk, and complexity.  "You want to figure out what counts as a high priority, low risk, low complexity start," advises Kelly.  

For value, Kelly separates everyday company work like hiring, payroll, calendars, and sick days, from the work that produces what customers pay for. Handling the first category unusually well can lower costs enough to become an advantage, the way low turnover shrinks a training budget. Something that returns hours to a team qualifies as something that returns revenue.  The second category holds the parts of a business a leader should be most careful about automating. "The things that make your company special, that deliver value for customers, that make your business more effective, are the things you really want to double down on," Kelly says. It is possible that AI can transform this part of the business, but it should be done intentionally, which means it is usually not the place to start when it comes to AI experimentation. Projects requiring data a company can't safely use, licenses it does not hold, or security exceptions it would rather not grant at all come off the list.  "Complexity always makes things take much longer than you think they're going to."

Finishing a project is the step companies usually skip. A team gets one project underway, and within weeks another two hundred products are being pitched to the same people, while colleagues ask for five more deployments. Thirty adjacent projects start before the first one produces anything worth examining, and nobody gets to look back at what the original choice taught them. For Kelly, scaling AI is a leadership problem as much as a technical one. "Starting there and actually allowing space to finish the work before you pivot to start the next thing is one of the biggest things a leader can do right now," he says.

Authority before architecture

Ahead of any technical question, Kelly recommends clients consider staffing. The choice determines how quickly the project reaches a result. Whoever takes it on will have to prioritize one group's interests over another's along the way. "The most important decision is who's going to be in charge of the process and what sort of authority and room to operate will they be given," says Kelly. "You have to have some level of familiarity and institutional credibility to drive the process."

Whoever owns the project then has to bring in the company’s exiting functions, ideally in the form of a centralized digital risk oversight function. Security review, information governance, and HR sign-off for employment matters each have clear owners and established processes. Organizations that invest heavily in responsible AI report materially higher earnings impact, and much of that work already belongs to teams a company has. "If you reinvent the wheel for AI, you're wasting time and not going to be as effective," Kelly warns. "You have to implement it in a way that integrates with your existing model."

The technical questions follow from there. Data requirements, on-premise or cloud hosting, third-party management or internal ownership, all of them depend on what the project is meant to accomplish, how sensitive the underlying data is, and how quickly results are needed. Kelly says supporting the person making those calls shortens the timeline more than any single architectural choice. "If you pick the right person and you're willing to back them, you'll be able to see a result sooner than if you're constantly second guessing everything they're doing in the middle of the project."

Beyond the five-star pilot

A pilot that goes well sometimes proves less than it appears to. Five enthusiastic users awarding five stars tells a company almost nothing about how a system behaves once everyone has it, and the conditions that let a small test succeed are often the ones a company cannot reproduce at full scale. Setting a few thresholds before the work starts, from a modest result to a strong one, gives a team something to check the outcome against. "It's pretty easy to make something that works once in one scenario," says Kelly. "It's really hard to make something that works consistently in a variety of scenarios that you didn't see coming, particularly in the hands of users who may not be excited about it."

Getting a usable answer out of a finished project puts pressure on the people closest to it. Months of effort produce an attachment to the outcome, and the finding most useful to the next project is often the one nobody wants to deliver. Saying it out loud is the step that most often gets skipped. "It requires a level of honesty for people who put a whole bunch of time into something who might need to say, 'look, this just didn't work,'" Kelly notes. "That's just the input to the next potential success."

The speed of adoption depends on how quickly people change what they do at work, which is slower than the release schedule of any model. The change reaches a business through phones, computers, and email, and it counts for nothing until someone works differently from the way they did last year. Finishing one project, examining what it produced, and letting that shape the next investment is a method that works on a timeline measured in years. "It's going to happen slowly, incrementally, and eventually massively over time," Kelly concludes.