AI at Work: Why Deeper Enterprise Use Produces Stronger Returns

Everyone’s deploying AI, and those going deepest are seeing the highest returns.

Almost all U.S. enterprises have adopted AI in some form. Across 60 large companies spanning software-as-a-service (SaaS) and platforms, financial infrastructure and digital payments, and cybersecurity and data privacy, AI is used in an average of seven out of eight business functions. How deep that adoption goes is the differentiator.

Some firms use the technology in a more focused way than others. The survey, which tracked 437 function-level deployments to gain a comprehensive picture of how firms use AI, found that only 20% of those deployments have reached the embedded stage, where AI becomes a core part of the workflow. Another 37% sit at broad deployment, in use across multiple teams without being woven into daily operations, and the remaining 43% are still in limited deployment or early exploration.

The deeper the deployment, the more likely a firm is to see a financial return. Among enterprises with no embedded functions, 25% report that AI is delivering a return. For those with one or two, 55% say the same. Among enterprises with three or more embedded functions, 93% report making back their investment. That payoff comes when a firm really commits to AI across functions.

That said, an enterprise using AI broadly but shallowly hasn’t necessarily stumbled. It may simply be earlier in the process or deliberately sequencing its investment differently than an enterprise racing ahead. Both approaches to adopting AI could be valid strategies depending on a firm’s needs and culture.

Firms are already investing heavily in AI, though not all are seeing the returns yet.

Budgets aren’t holding enterprises back from deploying AI deeper. Sixty-two percent of enterprises in the survey spent more than $10 million on new forms of AI in the past 12 months, with the median falling between $10 million and $25 million. If anything, enterprises with no embedded functions are slightly more likely to spend above $10 million than those with one or two functions embedded, at 58% versus 55%. Firms in the deepest deployment tier (those with three or more functions embedded) do spend somewhat more, at 79%, so money isn’t entirely irrelevant, but it’s not what’s holding the other enterprises back. They’re investing, but they either can’t or are choosing not to put that investment toward depth.

Depth and breadth aren’t interchangeable. Enterprises with one or two embedded functions use AI across an average of 41 of the 75 surveyed tasks. Enterprises with three or more embedded functions use it across 40 tasks, essentially the same number. Yet the first group reports a return 55% of the time and the second reports one 93% of the time.

It’s not that breadth doesn’t matter. Enterprises using AI across more tasks generally do report somewhat stronger returns than enterprises using it sparingly. But once an enterprise clears a moderate level of task coverage, adding more tasks doesn’t do much on its own. Depth is what pays off past that point.

Most firms are following their preexisting AI playbooks.

Firms’ past use of AI provides a pretty good clue as to how they’re deploying newer forms of AI. The survey asked how many functions embedded first-generation AI, meaning rules-based systems built up before 2022, when large language models first took off with OpenAI’s launch of ChatGPT. The answer predicts today’s depth of deployment almost as clearly as anything else in the data.

Among enterprises that had embedded no first-generation AI tools anywhere, 91% have embedded none of today’s more modern AI, either. For enterprises that had embedded first-generation AI in a single business function, most (63%) have now embedded new AI in one or two functions. Among enterprises that had embedded old AI in two or more functions before 2022, nearly half (47%) now have new AI embedded in three or more functions.

This is a snapshot rather than a tracked history, so it can’t prove that early depth prompted later depth. Still, the pattern points toward organizational habits. Enterprises appear to be running the same playbook they wrote for the last wave of AI, whether it favored going deep or staying broad. That’s not necessarily a verdict on which playbook is right. It does suggest that enterprises largely established their AI posture before the current generation of tools arrived, which matters for anyone trying to change that posture now.

Deep AI deployment doesn’t create new problems. It exposes the ones already there.

Enterprises with three or more embedded functions report an average of 5.6 barriers to AI adoption, compared with an average of three barriers among enterprises with nothing embedded. While at first glance, that may seem like a case against going deeper, it isn’t. Depth doesn’t appear to manufacture these problems, so much as expose them. Embedding AI into daily operations tests whether the underlying data is fit for use and whether the systems and institutional readiness around it can support it. An alternative reading is that deeper adopters are simply more attuned to their constraints. Either way, it’s the point at which always-present problems finally become visible.

Only 4% of enterprises with no new AI embedded cite internal skills as an obstacle compared with 64% of enterprises with three or more functions embedded. Regarding unclear ownership and accountability as hurdles, those shares are 4% to 50%, respectively. No one in the least embedded group and half of those in the most embedded group cited the absence of leadership alignment as an obstacle. Data quality and integration with existing systems follow similar patterns. These are readiness issues an enterprise only discovers once it tries to make AI part of day-to-day operations.

Going deeper means absorbing more organizational friction.

It isn’t that AI becomes less efficient as enterprises go deeper, or that depth is a trap. It’s that going deeper requires an organization to absorb friction, and not every enterprise is equally prepared to absorb that right now. Some enterprises in this survey may be holding back from embedding further precisely because they know their data pipeline, ownership structures or internal alignment isn’t ready yet, and they’re choosing to fix that first rather than push ahead. That may be the more disciplined strategy for some firms.

AI is already delivering measurable returns, and depth is clearly where those returns are concentrated. The real variable separating enterprises is organizational readiness: whether a company has the people, the data and the internal clarity to absorb what deeper AI deployment demands.

Read More

PYMNTS Intelligence is the leading provider of information on the consumer trends driving innovation in consumer finance, digital payments and financial inclusion. To stay up to date, subscribe to our newsletters and read our in-depth reports.

• The Enterprise AI Payback Curve: Adoption Accelerates as Returns Take Shape
• Wholesale Writes the AI Playbook: How Goods Firms Are Scaling Intelligence Across the Enterprise
• New Data Shows How Tech Sectors Are Turning AI Into Strategy

Methodology

“AI at Work: Why Deeper Enterprise Use Produces Stronger Returns” draws on a survey of 60 verified senior technology executives at U.S. enterprises with at least $1 billion in annual revenue, fielded July 14–22, 2026. Respondents were drawn equally from three technology sub-industries: SaaS and platform businesses, financial infrastructure and digital payments, and cybersecurity and data privacy. The survey defines AI as tools released after 2022.

The survey did not directly ask questions about reach, depth and breadth. They are measures built from the responses. Executives first reported whether AI was in use in each of eight business functions, then rated how far that deployment had progressed, from exploring to embedded, and finally reported which of 75 specific tasks within those functions used AI. The number of functions an enterprise has embedded sets its depth tier for every comparison in this report.

Companies self-reported their deployment stage, and the survey instrument didn’t set a fixed line between broad deployment and embedded use. Because of that, every depth-tier comparison in this report was also tested against a looser definition that merges those two stages, and the pattern held. Given the sample size, comparisons across the smallest depth tiers should be read as directional rather than definitive.

^{1. This finding contrasts sharply with a July 2025 study from MIT showing that despite up to $40 billion in enterprise investment in generative AI tools, 95% of organizations report zero return. The two studies measure different populations and outcomes.↩}