By Reginal Campbell · August 11, 2026 · 10-minute read

Nobody Asked What the Saved Time Was For.

AI is creating capacity faster than most organizations know what to do with it.

Ask most program leaders whether their teams are using AI, and the answer is yes. Ask what has actually changed as a result, and the answer gets vague fast: faster first drafts, quicker email, fewer hours lost to routine reporting. Ask what the organization is doing with those recovered hours, and most leaders cannot say. Nobody redirected them. Nobody redesigned the work around them. They simply evaporated back into the workday.

That gap is not anecdotal. A six-month field experiment tracking 7,137 knowledge workers across 66 firms found that employees with active access to generative AI saved close to two hours a week on email.^[1] Researchers found no meaningful shift in what those workers actually did with the rest of their time. The composition of the work stayed the same. It just happened faster.

That finding gets to the heart of the problem with enterprise AI. AI can compress a task. It cannot decide what the freed capacity is for. It cannot remove an obsolete approval process, redefine a role, change a performance measure, or stop leaders from filling every saved hour with more of the same work. Those are operating-model decisions, and no tool makes them on its own.

Over more than fifteen years leading enterprise programs, I have seen this pattern repeatedly: a tool saves time, leadership treats the time savings as the outcome, and nobody redesigns the system around the capacity that was created. The organization gets faster. The work does not necessarily get better. AI is not failing to create efficiency. It is creating efficiency inside operating models that were never rebuilt to capture its value. That is a leadership problem, not a software problem.

Enterprise AI Is Everywhere and Still Not Very Deep

AI adoption statistics are cited as if they settle the transformation question. They do not. Adoption measures breadth. Transformation requires depth.

The U.S. Census Bureau found that 18 percent of firms used AI in at least one business function during a recent reference period, rising to 32 percent when weighted by employment. Among firms using AI, though, 57 percent had deployed it in three business functions or fewer.^[2] McKinsey’s international executive survey found a much higher headline number: 88 percent of respondents reported regular AI use somewhere in their organizations, yet only about one-third said they were scaling AI across the enterprise.^[3] Different populations, different definitions of “use,” same conclusion: logging into an AI tool is not organizational transformation. A faster status report or a cleaner first draft can be useful without proving the organization has changed how decisions get made or how low-value work gets retired.

A program manager can produce twelve AI-assisted status reports each week and still be operating inside the same governance structure, with the same bottlenecks, the same unclear ownership, and the same unresolved dependencies. That is adoption without redesign.

AI Did Not Create the Data Swamp

When an enterprise AI system produces a confident but wrong answer, the model is an obvious target. Sometimes the model is the problem. But in enterprise settings, the failure often originates in the information environment surrounding it: source material that is outdated, two departments using the same term differently, a system retrieving the wrong version of a document, or a business rule that lives only in the memory of an employee who was never involved in the implementation.

AI did not create those conditions. It made them harder to ignore. Enterprise data environments accumulate disorder quietly through acquisitions, local taxonomies, drifting definitions, and ownership that disappears during reorganizations. Humans have always worked around these problems: they ask a colleague, or they know which spreadsheet is actually trusted. AI systems do not possess that institutional instinct unless the organization deliberately captures it.

A major academic review of more than 140 papers on AI data readiness describes readiness as a continuing lifecycle involving data inventory, quality, metadata, integration, governance, and evaluation.^[4] NIST’s generative AI guidance similarly treats provenance, testing, and human oversight as ongoing responsibilities rather than one-time setup activities.^[5] Many organizations still talk about AI readiness as though it were a technical sprint before launch: clean the data, connect the system, move to the next use case. That is not how enterprise information behaves. Data changes, permissions change, regulations change, and source processes deteriorate. AI readiness is not a cleanup event. It is an operating obligation.

Deloitte found that data-related problems caused 55 percent of surveyed leaders to avoid certain generative AI use cases entirely.^[6] That does not mean those leaders rejected AI. It means they recognized that their information foundation could not yet support the capability safely or reliably.

For program leaders, the practical lesson is straightforward: do not diagnose every bad AI output as a model failure. Ask what source was used, who owns it, when it was updated, what it means, who authorized its use, and whether the system retrieved the right context. What appears to be a model problem may be a governance, ownership, retrieval, or data-quality problem wearing the model’s name.

The Pilot Was the Easy Part

AI pilots often succeed under conditions the enterprise cannot reproduce. The document set is curated, the user population is small, permissions are simplified, and edge cases are handled manually. Someone quietly cleans the data, removes duplicate files, and explains the exceptions before the demonstration reaches leadership. That work rarely appears in the pilot business case; it is absorbed by employees who want the pilot to succeed.

The model may have genuinely demonstrated that it can perform the intended task. But the pilot may also have demonstrated something narrower: that a small, highly involved team can create ideal conditions for an AI system. Production removes those conditions. The capability now has to function across business units, data domains, security boundaries, and inconsistent operating practices, and the manual judgment that was invisible during the pilot must be formalized, automated, funded, or abandoned.

This is why the gap between experimentation and enterprise scaling matters. Only about one-third of McKinsey respondents reported scaling AI broadly across their organizations.^[3] Deloitte has similarly reported that many organizations do not expect most current experiments to scale in the near term.^[7] A successful demonstration proves that a capability can work under the conditions of the demonstration. It does not prove that the organization can operate it sustainably.

Before approving scale, leaders should ask what information was manually selected or cleaned, which edge cases were excluded, who reviewed the outputs, which permissions were simplified, which business units or data domains were not represented, what work was absorbed outside the project budget, and what changes when the user population increases by a factor of ten or a hundred. The question is not whether the pilot worked. The question is what made it work.

AI Moves Human Work Backstage

AI is often described as a labor-saving technology, and that is partly true. It can reduce drafting, summarization, search, and first-pass analysis; a controlled project-planning study found that AI assistance helped novices approach the performance of unassisted professionals on a bounded planning task.^[8] But reducing visible production work does not eliminate the human contribution. It often relocates it. Someone still has to determine whether the answer is correct, reconcile conflicting sources, and identify the exception the model smoothed over.

A pilot study of AI-supported risk management found that AI-assisted teams produced more structured and comprehensive outputs, while teams working without AI demonstrated stronger contextual and stakeholder sensitivity.^[9] The lesson is not that humans are always better or that AI makes judgment worse. Different forms of work are being redistributed: AI improves the breadth and structure of a first pass, while human beings still have to supply context, challenge, and accountability.

A study of 319 knowledge workers described a similar shift: information gathering increasingly became verification, problem solving became response integration, and task execution became stewardship.^[10] The danger appears when organizations automate the visible work but fail to recognize, staff, or reward the verification work that replaces it. The meeting summary gets automated, but nobody owns checking whether the commitments were captured correctly. The executive update becomes more polished, but the leader receiving it is now further removed from the underlying disagreement. That work belongs in the plan, the budget, and the role design, or it becomes invisible labor performed by whoever cares enough to prevent the system from failing.

Program Managers Got a Tool. Many Still Did Not Get Authority.

The expectations placed on program and portfolio leaders are changing. They are increasingly expected to understand AI, enterprise data, governance, model risk, and the relationship between technical outputs and strategic decisions. That evolution makes sense: enterprise AI cuts across business operations, technology, data, legal, security, and workforce adoption, and someone has to connect those pieces. Program management is well positioned to play that integrative role.

But the role does not become strategic merely because the expectations become more strategic. In many organizations, program managers remain accountable for outcomes without meaningful control over data ownership, funding for remediation, access decisions, or the authority to stop an unready initiative.

The research does not yet provide strong direct evidence that program managers broadly are receiving greater authority as AI adoption expands. What it does show is that AI increases the need for verification, cross-functional governance, data fluency, and contextual judgment. That creates a serious organizational risk: expectations can rise faster than authority. When that happens, the program manager becomes responsible for coordinating the cleanup, integration, governance, validation, and adoption work that nobody included in the original AI business case. They are expected to produce the outcome but cannot compel the organization to fix the conditions on which the outcome depends.

That is not an AI literacy problem. It is a role-design problem. If program leaders are expected to own AI-enabled outcomes, they need the authority to challenge readiness assumptions, surface hidden costs, escalate unresolved ownership, and recommend that an initiative be paused or stopped. Accountability without decision rights is not empowerment. It is exposure.

Rebuild the Operating Model, Not Just the Metrics

Many AI programs are still evaluated through measures that are easy to collect and difficult to interpret: number of users, prompts, licenses activated, pilots launched, or hours reportedly saved. Those measures tell leaders that something was used. They do not establish that anything important improved. A program can generate more reports without making better decisions. An organization can save thousands of hours while those hours simply disappear into increased workload. The correct question is not whether AI saved time; it is what organizational value was created with the time that was saved. That requires measures such as decision cycle time, error and rework reduction, benefit realization, and sustainability after the pilot team disengages. Time savings are an input. Value is the outcome, and the two are not the same thing.

Fixing this requires decisions that cannot be delegated to the software.

Decide What Work Will Stop

Do not assume saved time will automatically become strategic capacity. Name the reports, meetings, and approvals that will be removed, and define what the reclaimed capacity is for. Without subtraction, AI becomes an additional layer of work rather than a replacement for existing work.

Price the Work the Pilot Concealed

Data discovery, remediation, validation, security and legal review, and ongoing maintenance belong in the business case. Do not bury them under “implementation support.” Make the full labor visible before approving scale.

Separate Feasibility From Readiness

Use the pilot to expose permission conflicts, ownership gaps, and failure conditions rather than to conceal them. A working demo proves the model can perform the task. It does not prove the organization is ready to run it at scale.

Give Accountable Leaders Stop Authority

Someone must be empowered to say the data is not ready, the ownership is unresolved, or the expected value no longer justifies the cost. A readiness gate without stop authority is theater.

Program managers have their own work to do here. They do not need to become data scientists, but they do need to stay close to source evidence rather than letting a polished AI-generated summary become the only version of reality they see. They should make hidden labor visible by putting data preparation and governance into the work breakdown structure, interrogate what was manually curated in every pilot before accepting its results as a baseline, and translate data debt into program language, explaining how an undefined term becomes rework and an unresolved owner becomes schedule risk. Most importantly, they should refuse to accept ownership of an outcome without documenting the decisions, access, and escalation rights required to produce it. The future value of program management will not come from generating more polished artifacts. It will come from helping the organization see the system beneath the artifact.

The Time Has Been Saved. Now Capture the Value.

The tools are moving faster than the operating models built to receive them. That gap will not close because a more capable model is released next quarter. It closes when leadership redesigns work, removes obsolete requirements, funds the hidden labor, strengthens the information foundation, and gives accountable people the authority to act on what they discover.

The NBER field experiment, the adoption gaps documented by the Census Bureau and McKinsey, and the data barriers reported by Deloitte are not separate observations. They are different views of the same organizational problem: companies are becoming better at deploying AI tools than at redesigning the systems those tools operate inside.

AI can save time. It cannot decide what the time is worth. It cannot repair years of data debt without ownership, funding, and human effort. It cannot turn a successful pilot into a sustainable enterprise capability, and it cannot grant a program manager the authority the organization withheld. Those are leadership decisions. In many workflows, the time is already being saved. Whether it becomes strategic capacity or simply disappears into the same workload at a faster pace is still up to us.

Sources

Dillon, E. et al., “Shifting Work Patterns with Generative AI”, NBER Working Paper 33795, May 2025, revised November 2025. A six-month randomized field experiment covering 7,137 workers across 66 firms. ↑
U.S. Census Bureau, “The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks”, 2026. Nationally representative Business Trends and Outlook Survey AI supplement, November 2025 to January 2026. ↑
McKinsey & Company, “The State of AI”, 2025. Survey of 1,993 respondents in 105 countries. Consulting-sponsored and self-reported; not representative of all firms. ↑
“Data Readiness for AI: A 360-Degree Survey”, ACM Computing Surveys 57(9), April 2025. A review of more than 140 papers. ↑
National Institute of Standards and Technology, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile”, NIST AI 600-1. ↑
Deloitte, “State of Generative AI in the Enterprise, Third Quarter”, 2024. Consulting-sponsored, self-reported global enterprise survey. ↑
Deloitte, “State of Generative AI in the Enterprise”, 2025. Consulting-sponsored, self-reported global enterprise survey. ↑
Holzmann et al., “Bridging the Expertise Gap: The Role of Generative AI in Supporting Project Planning Tasks for Novices and Professionals”, Creativity and Innovation Management, June 2025. ↑
“How Generative AI Supports Novice Project Managers in Creative Reasoning: Evidence from a Pilot Experiment in Risk Management”, International Journal of Managing Projects in Business, 2026. ↑
Lee et al., “The Impact of Generative AI on Critical Thinking”, CHI 2025. A study of 319 knowledge workers describing 936 first-hand examples; self-reported and correlational. ↑

About the author

Reginal Campbell writes about enterprise technology, AI governance, leadership, and the systems organizations build to make consequential decisions.