A deployment autopsy: what the failed rollouts of the last two years reveal about where L&D should put its energy.
Your people do not need more AI training. They need to know how their work changes.
That distinction is the whole game. Most organizations are getting it backwards, and the autopsy reports now say so plainly.
Deloitte’s 2026 State of AI in the Enterprise found that insufficient worker skills is the biggest barrier to integrating AI into workflows. The reasonable response is to train people, and 53% of leaders are doing exactly that, raising general AI literacy, with another 48% running upskilling or reskilling programs.
But here is the number that matters more: 84% of those same organizations have not rewired a single role or workflow around AI. They bolted the tool onto the old job and called it adoption.
McKinsey’s 2026 research completes the picture. Almost 90% of companies now use AI somewhere. Only about 6% see material impact on earnings. PwC’s 2026 operations survey found 89% of operations leaders say their technology investments have not fully delivered expected results.
The pattern is consistent: the technology arrived, the work did not change, and the returns did not follow.
The London School of Economics and Protiviti added the sharpest edge in October 2026. Across more than 3,000 employees, company-provided AI training got 93% of people using AI, against 57% without it. Yet the chance of actually saving one to three hours a week sat near 50% either way. Training drives adoption. It does not drive productivity. The payoff depends on what people do with the time saved, and that is a work design question, not a training question.
What the autopsies find
Read those numbers as autopsy findings from thousands of deployments and a consistent cause of failure emerges. The work was never redesigned. That is the first and largest finding. The 84% who never rewired a role or workflow bought exactly the outcome they measured: none.
The second finding is quieter. Almost no one set the kill criteria before the pilot started. There was no written definition of what success looked like. There was no date when the evidence was due. There was no named person with the authority to call it off. So nothing got called off. Pilots drifted into production on enthusiasm and vendor demos, and the results disappointed on schedule.
The third finding is the one that should sting our profession. Training was prescribed as the fix for a design problem. The LSE and Protiviti numbers show why that prescription keeps failing. Training gets people to use the tool. It does not change the work the tool sits inside. When the deployment fails anyway, the usual response is more training. That is not a remedy. It is the same mistake with a bigger budget.
Every failed deployment gets an autopsy, formal or not. The question is whether yours happens before the money is spent or after.
The trap: teaching tools instead of redesigning work
Generic AI literacy has a place. Everyone should understand what these tools can and cannot do. But literacy alone does not change how a team operates. A two-hour prompting workshop does not tell an instructional designer which parts of their job still need them.
This is where most “AI upskilling” content fails. It teaches people to use the tool inside the existing workflow. The workflow is the problem.
The organizations seeing returns did something different. They asked which work should disappear, which work AI should assist, which work agents can execute, which human capabilities become more important, and how they will know adoption is creating business value. Then they redesigned the work around the answers.
Training followed the redesign. Not the other way around.
What a working kill switch looks like
In a role where I chaired an education technology committee, we evaluated a third-party add-on for Microsoft Teams meant to lift engagement in web-based training. Engagement was not broken. Two departments were simply curious whether it could go higher.
We ran the tool through our evaluation filters. We assessed the risks with our security and technology colleagues in the room. We approved a limited trial: two computers, planned events only. After the trial the verdict was plain. The tool was clunkier than promised. It added no value. It created new friction for the people using it. We canceled it. It was never rolled out.
That was a governance success, not a failed pilot. A failed pilot is a rollout nobody stopped. The difference was decided before the trial started. We had evaluation filters up front. The risk assessment included the right people. The trial was small enough to contain the blast radius. And the exit criteria were ones we were prepared to act on.
Here is the discipline most AI deployments are missing. A disciplined kill is stronger evidence of governance than a rollout. It proves the process can say no. If your pilots have no written success criteria, no evidence date, and no one authorized to stop them, you do not have an adoption program. You have a spending program with good branding.
Why L&D has to go first, starting with itself
There is an uncomfortable truth here for our profession. L&D cannot lead enterprise AI adoption while its own operating model still runs on courses.
If your team’s response to AI was to produce courses faster, you proved the point. AI made content generation cheap. That did not make your team more valuable. It made the course factory more efficient, and the course factory was already the wrong model.
The L&D teams that will matter in the next three years are making a different move. Their instructional designers are becoming workflow architects. Their value is shifting from content production toward diagnosing work, judging output quality, and measurably improving performance. They decide which work humans should perform, which AI should assist, and which agents should execute. Then they build capability around those decisions.
That is not an upskilling program. It is an operating model change with a capability plan attached.
Run the autopsy on your own pilots first. If L&D cannot point to a tool it evaluated and killed on the evidence, it has no standing to tell the enterprise how to govern AI.
The five questions every function has to answer
Strip away the jargon and functional AI adoption comes down to five questions. They work for L&D, and they work for every function after it:
- What work should disappear?
- What work should AI assist?
- What work can agents execute?
- What human capabilities become more important as a result?
- How will we know adoption is creating business value?
They are also, in reverse, the five questions every deployment autopsy asks. The ones left unanswered at launch become the findings later.
Most teams cannot answer the first question honestly, which is why they never reach the fifth. Start there. List the work your team does in a month. Mark what exists only because the old process required it. You will find candidates for disappearance before you find candidates for AI assistance, and that order matters.
From questions to capability: the EDGE sequence
Questions diagnose. Building capability takes a sequence. In AI for Learning and Development I lay out the EDGE sequence: Extract what the work actually requires, Design the intervention around the redesigned workflow, Generate with AI where it earns its place, and Evaluate whether performance moved.
Note the order. Extraction and design come before generation. Most teams start with generation, which is why they end up with faster content and unchanged performance. Faster content is not better learning. It never was.
The diagnostic that accompanies this article applies the five questions to your L&D function directly. Ten questions, one page, an honest score. If your team cannot answer them, you do not have an AI training problem. You have a work design problem, and training will not fix it.
Going further
The method behind this series is the EDGE framework, Extract, Design, Generate, Evaluate. My book AI for Learning & Development works the full system with the prompts and workflow templates: https://www.amazon.ca/dp/B0D5YJ1QQG
The move
Stop asking how to upskill your team for AI. Start asking how their work changes, then build the capability the new work requires.
And before the next pilot starts, write its ending in advance. Name what success looks like. Set the date the evidence is due. Name the person with the authority to call it off. A pilot with a kill switch is an experiment. A pilot without one is a purchase order.
The organizations that figure this out will not be the ones with the most AI training hours. They will be the ones where the work itself was redesigned, and the people were prepared for the work as it is now, not as it was.
Your people do not need more AI training. They need to know how their work changes. Give them that, and the capability follows.
Dr. Ravinder Tulsiani is the author of AI for Learning and Development: Faster Content Isn’t Better Learning and Solve the Right Problem: A Practical Guide for L&D. Take the AI-Ready L&D Function Assessment.
