GenAI in the Workplace: What Actually Works

I keep a running list of the questions I get asked most often in advisory sessions. A year ago, the top question was "should we use AI?" That question has disappeared. Completely. It has been replaced by a more uncomfortable one: "we adopted AI, so why are the results not there?" AI adoption is no longer the question. Stanford's 2026 AI Index reports that 88% of surveyed organizations now use AI in some form, and generative AI reached 53% population adoption in just three years. The AI Index compares this to the personal computer and the internet, not mobile phones, to make the point about speed. But adoption is not the same as measurable return. A separate preliminary study from MIT's NANDA initiative argues that most GenAI pilot programs still fail to produce measurable financial impact within six months of launch. The Stanford Enterprise AI Playbook, which studied 51 cases where AI did deliver value, set out to understand what those organizations did differently. The answer, consistently, was not better models. It was better organizational design. I have spent the past year advising enterprises across manufacturing, insurance, finance, and public sector on exactly this gap. And the patterns are more consistent than you would expect.

The Handover Cliff

Before I get into the research, I want to share something from my own experience that I think explains why so many AI deployments stall. Back when I was running Corporate IT, we would get handed systems built by brilliant external teams. Technically impressive. Documented, sort of. And completely unabsorbable by the organization that was supposed to run them. We would end up restricting what the system could do because nobody on our side understood it well enough to trust it. The transformative initiative quietly becomes the thing running in a corner doing less than it should. I wrote about this handover cliff in my last post. It still haunts me. Erik Brynjolfsson calls the broader pattern the productivity J-curve. For every dollar you invest in AI technology, you may need up to ten dollars in complementary intangibles before you see the gains. New processes. Retrained people. Redesigned workflows. Cleaned-up data. Reorganized teams. The Stanford Playbook found that 77% of the hardest challenges in its 51 enterprise cases were not technical. They were change management, data quality, process redesign, and organizational readiness. The unsexy stuff that never makes it into the vendor pitch. The board sees "AI project delivered." The ops team sees "black box we are afraid to touch." Same project, two completely different realities.

How Oversight Design Changes Everything

The Playbook's Chapter 3 examined how human oversight is structured across AI deployments. They classified each case on a three-point scale: escalation, where AI handles 80%+ autonomously and humans review only exceptions; approval, where humans review and approve every AI output before action; and collaboration, where human and AI work together continuously on each task. The results are striking. Escalation-based models delivered a median productivity gain of 71%. Approval models came in at 30%. Collaboration models also landed at 30%, though with a higher average of 43%. The Playbook is careful to note that this partly reflects task selection. Escalation tends to get applied to high-volume, recoverable work. Approval models serve regulated or high-stakes contexts where human review is legally required. The right oversight model depends on error tolerance, regulatory requirements, and task complexity. Still, the practical implication matters. The default instinct in most enterprises, especially in regulated industries where I do a lot of my work, is to put a human in the loop on everything. It feels safer. But spreading human oversight across every single task dilutes it to nothing. The human becomes a rubber stamp. They stop applying judgment and start clicking "approve" four hundred times a day while their brain checks out somewhere around click fifty. That is theater dressed up as oversight. The escalation model works because it concentrates human attention where it matters. The AI handles the volume. The human handles the exceptions. It sounds obvious. Almost nobody designs for it.

Where Agentic AI Delivers, and Where It Does Not

The Playbook's Chapter 8 examines agentic AI separately from the oversight analysis above. Agentic implementations are currently a minority, about 20% of the cases studied. But even with immature scaffolding, agentic AI delivered higher median productivity gains: 71% versus 40% for high-automation approaches. The conditions that predict agentic success keep showing up. High volume. Clear success criteria. Recoverable errors. When all three line up, the results are hard to argue with. A supermarket chain deployed an AI procurement agent that negotiated with suppliers, optimized orders, and managed inventory autonomously. The result doubled their EBITDA contribution from procurement. But look at the conditions: massive transaction volume, well-defined rules for what constitutes a good deal, and errors that get caught and corrected in the next order cycle. Textbook fit. Contrast that with a legal team trying to use an agent for contract review. Low volume, ambiguous criteria, errors with potentially irreversible consequences. The agent added overhead instead of removing it. Same underlying technology. Completely different outcome. The job market tells a similar story about where things are headed. The Stanford AI Index shows that demand for agentic AI skills has surged enormously since 2024, even as basic chatbot-related postings level off. Deployment, however, remains in single digits across most business functions.

The Data Readiness Myth

This one surprised me, because it contradicts the conventional wisdom I have been hearing and, honestly, repeating for years. Only 6% of organizations in the Playbook study had their data fully ready for AI when they started. Six percent. And yet many of them succeeded anyway. In 88% of successful deployments, the AI was not just consuming clean data. It was fixing data. LLMs turned out to be remarkably good at normalizing messy inputs, extracting structure from unstructured sources, and making sense of the kind of chaotic real-world data that traditional systems choke on. "Our data is not ready" is the single most common reason I hear for delaying AI deployment. In many cases, it gets the causality exactly backwards. The AI can help clean the data. Waiting for perfect data before starting is like refusing to hire a translator until everyone already speaks the language. The relationship between data readiness and AI deployment is iterative, not sequential. You start with what you have. The AI helps you see what is broken. You fix it. The AI gets better. That loop is how the successful deployments actually worked. The ones that waited for clean data before starting? Most of them are still waiting.

AI Making Previously Impossible Work Feasible

The productivity framing dominates most AI conversations. How much time does it save? How many FTEs can we reduce? Valid questions. But they miss what I think is the more interesting story. Some of the most compelling cases in the Playbook are not about doing existing work faster. They are about doing work that was never on the roadmap, work that simply was not economically viable before AI made it possible. A fintech with over 100 million customers needed to migrate millions of lines of legacy code. The traditional estimate was 18 months with over 1,000 engineers. With AI coding agents, business units began completing migrations in weeks. AI made a plan possible that simply did not exist before. An insurance firm found that AI could rewrite legacy systems from scratch faster than refactoring them. A project originally quoted at 5,000 hours with a team of seven was finished in 600 hours with a team of three. This is the part that most AI strategies miss, and it frustrates me, because this is where the real money is. They frame AI as an efficiency play: same work, less cost. Neat. Safe. Easy to model in a spreadsheet. But the real opportunity is different work. Things your organization needs but could never justify building. One case made this concrete. A call center company, traditionally a pure cost center, embedded agentic AI directly into its product. Rather than using AI to make human agents faster, they redesigned the service so AI could resolve tickets end to end. The result was competitive repositioning. The company started winning deals it could not have competed for before. The Playbook reports over 20 new project wins attributed to AI, with the narrative referencing around thirty total. An independent technology assessment ranked the company among the top four for AI capabilities in customer relations. The other three were AI-native companies. A traditional call center, benchmarked against startups.

The Permission Problem

An MIT Technology Review/Infosys survey of global executives surfaced a finding I think is underrated: 83% of executives say psychological safety measurably improves the success of AI initiatives. And 22% admitted they had hesitated to lead an AI project because they were afraid of being blamed if it failed. One in five leaders self-selecting out of AI leadership because the organizational culture punishes failure. In a technology where experimentation is the only way to learn what works, that dynamic guarantees stagnation. Steven Kerr wrote a famous paper in 1995 called "On the Folly of Rewarding A, While Hoping for B." Organizations systematically reward behaviors that contradict their stated goals. They hope for long-term growth but reward quarterly earnings. They hope for teamwork but reward individual achievement. They hope for candor but reward people who report good news. Apply that lens to AI adoption and the pattern is painfully clear. Companies say they want AI in production. But they reward impressive demos. They reward innovation theater. They fund PoCs with great fanfare and then starve the production engineering that would make them real. The team that builds a flashy prototype gets the stage time. The team that does the unglamorous work of data cleaning, integration testing, change management, and production hardening? Invisible. The organizations actually shipping AI into production share one trait. Hard to quantify, but impossible to miss: they have made it safe to try things that might not work. Not reckless, but genuinely safe. The Stanford Playbook reviewed 51 enterprise AI deployments. In none of them was anyone punished for a failed initiative. A vast majority of top-performing organizations in the MIT study strongly encourage experimentation. The bottom performers talk about AI strategy in boardrooms and penalize the people who actually try to execute it.

Where Humans Still Hold the Edge in Enterprise Settings

As AI commoditizes execution, three capabilities keep surfacing in the engagements I work on as the places where humans still hold a decisive advantage. Discernment: the ability to look at something technically correct and know it is wrong for the context. An AI can generate a strategy document that hits every structural checkbox and completely misses the point. Discernment is pattern recognition trained on years of domain experience. AI does not have it. Curiosity: not information retrieval, but the kind of thinking that says "wait, that does not make sense" and spends a weekend pulling at the thread. In a world drowning in AI-generated output, the ability to ask better questions is becoming more valuable than the ability to generate answers. Human connection: the relational judgment that tells you when to push, when to listen, when to shut up. Every meaningful advisory engagement I have been part of had a moment where the technical answer was not the right answer, and the only way to know that was to understand the people involved. These are the skills that will not depreciate. Everything else might.

What I Would Actually Tell You to Do

If you are leading AI adoption right now, whether as a CTO, a CDO, or just the person who ended up responsible for "the AI thing" because nobody else raised their hand, here is what I would say. Design for the oversight model, not just the model. Escalation, approval, or collaboration. That architectural choice matters more than which LLM you are running. Stop waiting for clean data. Start with what you have. Use the AI to identify what is broken. The organizations that waited for perfect data are still in planning committees. Look for new work, not just faster old work. Ask your teams: what would you build if the cost of building it dropped by 80%? That question surfaces more value than any automation audit I have ever seen. Make it safe to fail. Psychological safety is infrastructure, same as your cloud environment or your data pipeline. Govern the portfolio, not the pilot. AI is arriving from five directions at once. SaaS vendors, platform providers, open source, shadow IT, business units buying their own tools without telling anyone. Sound familiar? This is cloud adoption all over again, except faster.

What Comes Next

Here is what I think the next two years look like. I want to be honest that this is projection, not certainty. The skills market is already shifting. Job postings are moving from basic GenAI literacy toward agentic systems, multi-agent orchestration, and tool use. The organizations that invested in training people to use ChatGPT are about to discover that the real capability gap is in people who can design, deploy, and govern autonomous AI systems. Austria, my home market, sits at 31.4% generative AI adoption among the working-age population according to Microsoft's AI Diffusion research. Below the global leaders, but not dramatically so. The gap is execution infrastructure, not awareness. The regulatory environment in Europe is tightening. Prohibitions and AI-literacy rules under the EU AI Act started applying in February 2025, GPAI and governance provisions in August 2025, and much of the remaining regime takes effect in August 2026. That creates an opportunity for organizations that build governance and trust infrastructure early. Compliance becomes a competitive advantage when your competitors have not built it yet. The gap between organizations that talk about AI and organizations that ship AI keeps widening. The ones pulling ahead are doing it with better organizational design, better measurement, better incentive structures, and the willingness to change how work actually happens. The technology is ready. It has been ready for a while. The question is whether you are willing to do the harder, uglier, less-glamorous work of actually changing how your organization operates. That is where the value lives. In the messy, human system around the model. One of the sharpest people I work with, a CDO in financial services, told me something last week that stuck. "I finally stopped asking whether AI works and started asking whether we work, with AI." I think that is the right question. And I think most organizations have not asked it yet.

Sources

Stanford HAI, AI Index Report 2026 (April 2026)
Pereira, Graylin, Brynjolfsson, Enterprise AI Playbook, Stanford Digital Economy Lab (April 2026)
MIT NANDA Initiative, preliminary findings on GenAI pilot impact (2025), as cited in the Enterprise AI Playbook
MIT Technology Review Insights / Infosys, Creating Psychological Safety in the AI Era (December 2025)
Steven Kerr, "On the Folly of Rewarding A, While Hoping for B," Academy of Management Executive (1995)
Brenna Spain, "What Are the Skills That AI Can't Automate?" (2026)
Microsoft, AI Diffusion research on generative AI adoption by working-age population