Skip to content
All writing

GenAI in the workplace: What 51 successful deployments show

What research on enterprise AI suggests about oversight, data and adoption, and how it connects to the systems I once had to take over in Corporate IT.

Essay · AI · Leadership · April 16, 2026 · 9 min read

Engraving: a row of closed boxes, the last one opened.

In advisory sessions, I have been hearing less “should we use AI?” and more “we adopted it, so why are the results not there?” The second question is harder to answer because it asks what changed in the work after the tools arrived.

Stanford’s 2026 AI Index reports AI use in 88% of surveyed organizations and rapid growth in individual adoption. Those figures describe use, rather than the return an organization can expect. The Enterprise AI Playbook, published by Stanford’s Digital Economy Lab in April 2026, approaches the question through 51 deployments that delivered measurable value.

That selection matters. These are accounts of successful deployments, based largely on interviews and company-reported results. They offer useful examples of how organizations worked through difficulties, but they cannot tell us how likely any new project is to succeed. The Playbook also discusses preliminary MIT NANDA findings on pilots that had not produced measurable financial impact.

The organizational issues in the cases are familiar from my work with manufacturing, insurance, finance and public-sector organizations. They also remind me of an earlier part of my career.

The handover cliff

When I ran Corporate IT, we sometimes inherited systems built by capable external teams that our organization could not comfortably operate. The technology was impressive; our understanding of it was much less complete. We ended up restricting what a system could do because we did not know it well enough to trust it.

That stayed with me. A project could be considered delivered while the people responsible for running it were still working out what they had received. I wrote about this handover problem because I see the risk again in AI work.

The productivity J-curve research by Erik Brynjolfsson and his co-authors helps explain the broader issue. Technology investment often needs substantial investment in processes, skills and organizational change before the gains appear. The Playbook cites earlier research suggesting that complementary investment can be several times the technology spend, in some cases up to ten times as much. I would not use that as a universal budget multiplier. I would use it as a reason to look carefully at what the business case leaves out.

In the Playbook, 77% of the hardest reported challenges concerned change management, data quality, process redesign and similar work around the technology. A delivery plan needs to make room for those costs and for the people who will operate the result.

Deciding where human review belongs

Chapter 3 distinguishes three arrangements: escalation, where AI handles most work and people review exceptions; approval, where a person reviews each output before action; and collaboration, where people and AI work together throughout the task.

The reported median productivity gain was 71% for escalation and 30% for each of the other two arrangements. Collaboration had a higher mean, at 43%. These are comparisons across different tasks and organizations. They do not establish that removing approval would improve the same workflow by a similar amount.

The useful question is what each review is for. High-volume work with recoverable errors may suit exception handling. A clinical document or a consequential financial decision may need review of every output, because of the risk involved or a specific legal requirement.

I am wary of approval steps that ask someone to review so many outputs that they have no time to exercise judgment. It is easy to design a process that records a click without establishing what the person checked. I want the reviewer to have enough context, time and authority to make the decision we are assigning to them.

Where agents helped in the cases studied

Chapter 8 looks separately at agentic implementations. They represented about 20% of the cases and had a median reported productivity gain of 71%, compared with 40% for the study’s high-automation category. This is a different classification from the oversight comparison above, and again the tasks differ.

The successful agentic cases tended to involve repeated decisions, clear criteria and errors that people could detect and correct. One supermarket chain used an agent to make purchasing decisions across its stores using inventory, sales and supplier data. The company reported that its EBITDA margin doubled. That is a result from one deployment, not an expected return for autonomous procurement in general.

The same report describes settings where continued human review was useful, including document review and clinical documentation. For a new use case, I would start with the consequences of an error and the evidence that someone could catch it in time. The label “agentic” does not settle that decision.

Working with imperfect data

Only 6% of the Playbook’s implementations began with data the researchers classified as fully ready. In 88% of cases, models helped make previously inaccessible data usable, including documents, transcripts and legacy code. That is more specific than saying AI fixed the data: making information accessible does not by itself establish that it is correct.

This challenged advice I had sometimes repeated about getting the data ready before starting. It suggests a more iterative approach. Test whether the model can extract something useful, examine the errors, improve the input or process and test again.

It also leaves a selection question. A study of successful deployments will miss some cases where poor data made the work impractical. I would not postpone every experiment until all the data was clean. I would identify which problems the proposed system could help with and which still required someone to resolve them.

Work that becomes feasible

The cases that interested me most involved work that had previously been difficult to justify or undertake.

A fintech serving more than 100 million customers needed to migrate millions of lines of legacy code. Its earlier estimate was 18 months with more than 1,000 engineers. After introducing coding agents, business units began completing migrations in weeks. The report describes progress by those units; it does not establish that the entire original program was completed on that timescale.

An insurance firm reported completing a system rewrite in 600 hours with three people, against an earlier estimate of 5,000 hours with seven. Those figures depend on the firm’s original estimate and its account of the finished work. They are a reason to revisit what is feasible, rather than a reliable multiplier for the next project.

A call-center services company put agentic resolution into its service and attributed more than 20 new project wins to AI. The Playbook’s narrative refers to around thirty new projects and reports that an independent assessment placed the company among the top four for AI capabilities in customer relations, alongside three companies built around AI. The commercial significance was that the service could compete for work it had previously struggled to win.

I get frustrated when an AI strategy only asks how much faster existing work can be done. The cases above suggest another useful conversation: which work have we left undone because it was too expensive, too slow or beyond the team’s capacity? Some of it may deserve another look. The costs of integration, checking and ongoing operation still belong in that discussion.

Making room for a failed attempt

In a 2025 survey by MIT Technology Review Insights and Infosys, 83% of executives said psychological safety measurably improved AI success, and 22% said fear of failure or criticism had made them hesitate to propose or lead an initiative. These are executives’ assessments, but they raise a practical question about what people expect will happen if an experiment disappoints.

Steven Kerr examined a related problem in “On the Folly of Rewarding A, While Hoping for B”, first published in 1975 and reprinted in 1995. Organizations can reward behavior that conflicts with their stated aims: individual achievement while asking for teamwork, or good-news reporting while asking for candor.

I recognize that tension when prototypes receive attention and funding while integration, testing and adoption are treated as work to find a budget for later. A team can do what it was rewarded for and still leave the organization without a usable service.

The Playbook describes sponsors who stayed involved, helped resolve obstacles and gave people permission to learn from attempts that had failed. That is a more useful expectation to set than asking teams to experiment while expecting every experiment to work. People also need to be able to report an error or a weak result without first turning it into a success story.

The judgment I still rely on

In advisory work, I often need to recognize when an answer that looks technically sound does not fit the organization. That depends on understanding its history, constraints and people. I want AI output examined with that context in mind.

Curiosity matters too. When something does not add up, someone has to keep asking questions and test the explanation. Producing a plausible response is easier than deciding whether it answers the problem in front of us.

And there are moments when the useful contribution is knowing when to push, when to listen and when to leave a little space. In the engagements I have worked on, understanding the people involved has sometimes changed what I thought the next step should be.

What I would ask a leadership team to do

I would ask them to decide how work will be reviewed before broad deployment, and to make sure the people responsible can carry out that review. I would give teams a bounded way to test imperfect data, with explicit checks on what comes out. And I would look at work the organization has not previously been able to undertake, alongside the opportunities to save time.

I would also manage the initiatives together. AI can arrive through SaaS suppliers, platforms, open-source software and tools bought independently by business units. Those choices can overlap or conflict. Someone needs a view of what the organization is using, what it costs and who will take care of it.

The skills required extend beyond using a chatbot. Teams that connect AI to operational systems need people who can build, evaluate and govern those systems. Training should follow the responsibilities people are taking on.

For Austria, Microsoft’s January 2026 diffusion report estimated adoption at 31.4% of the working-age population in the second half of 2025. That is a dated estimate of use, not evidence that an organization is ready to run an AI service. The operational questions still need to be answered locally.

European regulation also needs to be considered against the rules that apply to the particular system. Following the 2026 amendments, the European Commission lists December 2027 and August 2028 deadlines for different categories of high-risk AI, while other obligations already apply. I would use the Commission’s implementation guidance to plan the work, rather than assume a single deadline covers every initiative.

A CDO in financial services once put the organizational question to me this way: “I finally stopped asking whether AI works and started asking whether we work, with AI.” It is a question I would want the people operating the system to help answer.

Sources