Is the AI ROI gap actually a workflow depth problem?

← Back to Blog

“Stop asking whether the pilot succeeded. Ask what authority it was given.”

Diana Wolfe

Fifty-seven percent of firms that have adopted AI use it in three or fewer business functions, according to U.S. Census Bureau Business Trends and Outlook Survey data covering December 2025 through May 2026 (Emerj Artificial Intelligence Research, 2026). That figure sits next to a more famous one: roughly 95% of enterprise pilots show no measurable effect on profit and loss (MIT NANDA Initiative, 2025). The two numbers are usually reported as separate findings. I think the first explains the second.

Consider a hypothetical mid-size insurer with forty AI deployments: campaign copy, sales call summaries, a support chatbot, an internal knowledge search tool. Every one works as specified. None touches claims triage, reserving, underwriting, or fraud investigation, because those workflows carry regulatory exposure and nobody wants to own the first bad automated decision. At year end the CFO looks at combined ratio, sees nothing, and concludes the technology underdelivered. The technology did what it was permitted to do.

My hypothesis: the pilot failure statistic describes a decision-rights gap rather than a tooling gap. Most firms confine AI to peripheral functions. The minority that grant it authority inside core operational workflows capture most of the measurable value. The question was never whether the systems work. It was what organizational authority they were given.

The Research

Large-scale administrative data has shown for several years that AI adoption clusters in a narrow band of functions and firm types rather than spreading across the operating model (McElheran et al., 2024) (Established). The 2026 Business Trends and Outlook Survey figure is that same pattern, updated.

This is where Human Systems Architecture earns its keep. HSA treats organizational structure, workflow design, and decision rights as the designed layer that determines whether a capability produces value or sits idle. Under that lens, which functions AI touches reads as an architecture measure rather than a technology one. Shrestha, Ben-Menahem, and von Krogh (2019) supply the vocabulary: full human decision-making, full delegation to algorithms, hybrid sequential structures, aggregated human-AI structures. Each allocates authority differently, and the allocation is a governance choice made by people with names (Emerging). Raisch and Krakowski (2021) describe that choice as a standing tension between automating a task and augmenting the human who holds it, resolved case by case rather than once (Emerging).

The depth pattern is not an AI peculiarity. Brynjolfsson, Rock, and Syverson (2021) argue that general-purpose technologies show no measured productivity gain until firms make the complementary organizational investments that move the technology into core processes, which means deployment counts mismeasure value during the transition (Emerging). Babina, Fedyk, He, and Hodson (2024) find the same shape in AI specifically: investment depth, not adoption breadth, predicts firm growth and product innovation (Emerging).

What This Means in Practice

Peripheral placement is rarely a decision anyone remembers making. It is the residue of many small avoidances. Marketing copy has a low blast radius, so the pilot lands there. Underwriting has a regulator, so it does not. Parker and Grote (2022) argue that work design outcomes follow from deliberate choices about task allocation, and that defaults take over when nobody designs (Emerging). The default here is caution shaped like a strategy.

Take the insurer again. Moving a model into claims triage raises questions a marketing pilot never did. Who can override the model, and does an override require a written reason? What confidence threshold routes a case to a human adjuster? When the model and the adjuster disagree, whose call stands, and who reviews the disagreement rate at quarter end? What is the audit trail when a regulator asks how a decision was made? Those are decision-rights questions, and a firm can answer them before it has a working model.

That reframing changes who owns the problem. If the constraint is tooling, the answer is procurement. If the constraint is authority allocation, the answer sits with whoever can redraw accountability lines, which is generally not the team running the pilot. That mismatch may be the most common reason pilots stall at the boundary of a core workflow (Speculative).

Three Things to Take Away

Audit which functions AI touches before you diagnose an ROI problem

Count functions, not deployments. If AI touches three or fewer, as it does for 57% of adopting firms (Emerj Artificial Intelligence Research, 2026), a flat financial result tells you about scope rather than capability. Census-based adoption research has shown this concentration pattern for years (McElheran et al., 2024).

Treat expansion into core workflows as a governance decision

Assign decision rights before scaling, not after the first incident. Shrestha, Ben-Menahem, and von Krogh (2019) show that different structures for combining human and algorithmic judgment carry different accountability implications, so the structure has to be chosen rather than inherited. Write down who decides what, and what happens when the model and the human disagree.

Measure value capture by workflow depth, not by pilot count

Depth predicts outcomes; volume does not. Babina et al. (2024) find that firm-level AI investment depth predicts growth and innovation, and Brynjolfsson et al. (2021) explain why count-based metrics understate value during a general-purpose technology transition. Report how much operational authority AI holds, and where.

My Two Cents

The 95% figure has been read as a verdict on the technology. I read it as a verdict on organizational design, and the distinction matters because the two diagnoses point at different budgets. One sends you back to vendors. The other sends you to the operating committee.

What concerns me is the mechanism underneath. Moving AI into a core workflow requires slack: people who can redesign the work, governance capacity to define new decision rights, and tolerance for a period where the numbers get worse before they get better. Firms with that slack go deep and capture the returns. Firms without it buy more tools, stay peripheral, and add to the failure statistic. If that holds, the distance between the deep adopters and everyone else is self-reinforcing, and better models will not close it (Speculative). It closes, if it closes, through governance capacity that most organizations still book as overhead.

Read to Learn More

Academic: McElheran, K., Li, J. F., Brynjolfsson, E., Kroff, Z., Dinlersoz, E., Foster, L., & Zolas, N. (2024). AI adoption in America: Who, what, and where. Journal of Economics & Management Strategy, 33(2), 375–415. The Census-data picture of how narrowly AI adoption is distributed across firms and functions, and the empirical lineage behind the 2026 survey figures.

Industry: Emerj Artificial Intelligence Research. (2026). The conditions that turn AI pilots into enterprise value. Source of the Business Trends and Outlook Survey analysis showing 57% of adopting firms limit AI to three or fewer functions, plus the conditions that separate value-capturing deployments from stalled ones.

References