Research-backed answers on AI adoption, job transformation, and employee resistance. What the data reveals about layoffs, automation, and workforce planning.
Earlier this month, I sat in a leadership strategy meeting and watched the room divide cleanly in half. One side was all about AI productivity gains, the other doing math on headcount reductions. Neither side had certainty.
So I did what any reasonably anxious executive would do, and spent two weeks reading every credible study on the impact of AI on the workforce I could find. Reports from the OECD, Stanford, the International Labour Organization, BCG, LinkedIn, Indeed, and the AI labs themselves. I was looking for answers to the questions keeping many of us up at night.
Let me address the elephant in the room: are we about to execute waves of AI-driven layoffs over the next five years?
Recent reports and research, based on data collected during the past months and years, doesn't support the idea of mass, across-the-board job loss. But it also doesn't suggest we can pretend nothing will change.
The International Labour Organization looked at this globally and found that about one in four workers are in occupations with GenAI exposure. Most of these jobs contain tasks that still require human judgment and input. The ILO's position is that transformation is far more likely in the foreseeable future than outright replacement.
That said, Stanford researchers found something concerning when they analyzed US administrative data. Young workers, aged 22 to 25, are the most AI-exposed jobs, which saw a 13% decline in employment after widespread GenAI adoption. The impact was concentrated at entry levels, while more experienced workers in the same fields remained stable or even grew.
The OECD's review across multiple countries hasn't shown an AI-driven collapse in white-collar employment yet, but they're watching carefully for signs that certain groups are being locked out of AI-related opportunities. Meanwhile, Indeed's analysis suggests that while AI's potential impact on the workforce is widespread, only about 26% of U.S. job postings could be "highly" transformed, and actual outcomes depend heavily on how quickly businesses adopt these tools and whether workers get the reskilling they need.
Here's a surprising result: BCG surveyed workers and found that 41% think their job will probably or certainly disappear entirely within the next decade. Whether or not that fear is justified, it's tangible, and already shaping morale and retention.
My advice: Don't plan for waves of layoffs as an inevitability. But do plan for reallocation, reskilling, and scenario-based contingencies. Know which of your roles are most exposed to automation versus augmentation, because the difference matters enormously.
Roles that are heavily digitized, cognitive, and information-processing-heavy are the ones that will see the most augmentation.
The pattern keeps showing up across multiple studies.
Clerical occupations consistently rank as the most exposed in the ILO's refined GenAI exposure index. Software development is another occupation that Indeed flags as highly exposed to significant transformation. By contrast, nursing (which requires physical presence and human connection) will see AI mostly change administrative tasks rather than the core work.
The "why" comes down to task structure. Where work involves problem-solving and cognitive processing but doesn't require physical presence, GenAI can handle a much larger share of routine work and shift humans toward oversight, exceptions, edge-cases and judgment calls.
Indeed's framework is particularly useful here. They found that 46% of skills in a typical U.S. job posting are headed for "hybrid transformation", meaning GenAI does the bulk of routine work, but humans still manage exceptions and oversee the process. This is precisely the “redesign instead of replacement” dynamic we mentioned before.
Stanford's data reinforces this. They found that in occupations where AI use is more augmentative, employment actually grew. It's only where AI is used to automate that employment declined. What’s interesting here is that the distinction between augmentation and automation is actually showing up in hiring patterns, and it isn't just semantics.
LinkedIn's survey adds color to this. Employees are already using AI tools for innovation and brainstorming (72% reported this), automating repetitive tasks (70%), and simplifying work processes (58%). These are classic augmentation patterns: AI handling the tedious parts while humans focus on the interesting problems.
The ILO also notes that some strongly digitized occupations have seen increased exposure as GenAI's capabilities expand into more specialized tasks. So even roles we thought were "safe" because they required expertise are now candidates for significant transformation.
Your biggest augmentation opportunities are probably in roles where people are drowning in routine cognitive work. Look for jobs that involve a lot of drafting, templating, testing, analyzing, and formatting. Especially if those people then apply judgment to the output. That's where augmentation creates the most value.
Despite all the anxiety, there's very little evidence that entire professions are about to be automated away by GenAI.
The ILO is explicit about this: most jobs will be transformed rather than made redundant because they contain tasks that genuinely require human input. Indeed initially found zero skills that were "very likely" to be fully replaced by GenAI. In their latest analysis, they identified 19 skills, but that's only 0.7% of the roughly 2,900 skills they track.
Now, Stanford's findings add important context. Employment declines are measurable, but they're concentrated in specific segments of certain professions. For example, entry-level workers in highly exposed occupations. The substitution pressure appears first in roles where firms can directly replace routine tasks with AI, and it hits junior people harder than experienced ones.
The OECD makes an important distinction that often gets lost: occupations highly exposed to AI aren't necessarily the ones at highest risk of automation overall. When they looked at automation risk from all technologies (not just AI), the top three occupations were fishing and hunting workers, food processing workers, and textile/apparel workers. These are largely manual, repetitive jobs. Interestingly, 12% of male workers versus 6% of female workers are in high-automation-risk occupations, and the risk is much higher among less educated workers: 22% for those with lower education levels.
So yes, some skills and tasks will be replaced. But the evidence suggests this will be uneven, gradual, and concentrated in specific pockets rather than sweeping entire professions off the map.
What matters for your planning: focus on task-level redesign rather than assuming whole job families disappear, since significant AI impact on the workforce happens at the task and workflow level, not at the occupation level. The bigger risk isn't that entire departments become obsolete, but that we fail to redesign positions fast enough and get stuck with expensive, inefficient hybrid states.
This question might be the most urgent one for leaders, because teams that resist adopting AI solutions can disrupt top-down innovation efforts, which also affects the ROI of these initiatives. However, this resistance is often justified.
BCG's survey highlights parts of this issue. Only 36% of employees are satisfied with their AI training. Just 25% of frontline employees say they received sufficient leadership support on how and when to use AI. And 54% said they would use unauthorized AI tools if corporate solutions fall short.
That’s right: more than half of your workforce may already be using shadow AI because you haven't given them the right tools and guidance, or changed systems and processes without consulting them first.
But here's what makes this even more complicated: not all AI usage is created equal.
A joint study from Harvard, MIT, and BCG involving over 200 consultants revealed how people use AI matters infinitely more than how much they use it. The researchers didn't train anyone on "correct" usage, just watched what happened naturally when people integrated AI into their work.
Three distinct patterns emerged:
1. some delegate everything to AI, and copy-paste without review
2. others engage in constant back-and-forth
3. while the most productive group only uses AI for specific tasks in their domain
The workflow that actually builds skills looks like this: prompt comprehensively, challenge the output, ask it to critique your own approach, iterate until you reach consensus, then make the final decision yourself. This helps retain and build critical thinking and helps evolve in pre-existing skills.
The most important takeaway for leaders is that using AI correctly once per week might produce more value than using it 40 hours per week without proper domain knowledge and review. Volume doesn't equal impact. The mode of collaboration is everything.
The OECD case studies provide additional insight: in manufacturing contexts, older workers were sometimes described as skeptical toward AI and less willing or able to adapt. But interviewees emphasized that training could override negative attitudes and address skills gaps, especially when training is guided rather than self-directed.
CEEMET's paper (representing European technology and metalworking industries) stresses the importance of transparency, information, and consultation when implementing AI systems. They call for worker representatives to be involved and for a human-centered approach where human intelligence has the final say in algorithm-driven decisions.
The pattern across all these sources is consistent. Resistance comes from fear, uncertainty, and lack of support and trust.
BCG's data shows that employees with more than five hours of training are much more likely to become regular AI users. Those with clear leadership support have dramatically higher adoption rates: 82% versus 41% among frontline workers who lack that support. LinkedIn frames this as an opportunity: engage employees in "test and learn" approaches so AI becomes a defining skill rather than a threat. When people see concrete benefits, like 76% reporting they can shift saved time to more strategic work, resistance softens.
My advice would be to treat resistance as a predictable response to insufficient enablement, not as irrational fear. Address it with genuine leadership engagement, substantial training (not a single webinar), access to approved and actually useful tools, and transparent consultation with your workforce about how internal systems and processes could be changed for the better.
LinkedIn's data suggests three process entry points: innovation and brainstorming, automating repetitive tasks, and simplifying existing processes. These align with where employees are already finding value and where the time savings can be reinvested in strategic work.
BCG argues that organizations need to progress through stages: first "deploying" tools, then "reshaping" workflows, and eventually "inventing" new business models. About 50% of companies say they're starting to reshape processes, but BCG's analysis shows that companies creating the most AI value concentrate 80% of their investment in reshaping and inventing, focusing on a few core processes rather than scattering pilots everywhere.
As mentioned before, with half of workplace skills headed for hybrid transformation, the opportunity is to redesign processes so GenAI handles routine portions while humans manage exceptions and oversight. The already discussed exposure results point clearly toward clerical and digitized tasks, which suggests administrative, information-handling, and software development / data professional work are prime candidates. Think about all the places in your organization where people are moving information between systems, creating reports, drafting communications, processing documents, or working with code. Those are your high-value targets.
CEEMET adds an industrial angle: AI can support occupational safety, productivity and cost optimization by analyzing data in real time and detecting anomalies. It can also substitute for repetitive, monotonous, or dangerous tasks. If you have operations where people are exposed to physical risk or mind-numbing repetition, or maintain expensive equipment, AI may do more than improve efficiency.
My advice here: pick a small number of end-to-end processes that are important to your business and genuinely redesign them. Make sure they're processes where work is highly digitized and routine-heavy within knowledge workflows. Then concentrate your investment there rather than trying to AI-wash everything all at once.
Understanding barriers matters a lot, because you can't fix what you don't diagnose correctly.
As discussed above, BCG's results show that while 72% of workers are regular AI users, frontline adoption has stalled at just 51%. The main blockers are insufficient training, weak leadership guidance, and lack of access to the right tools. Only 36% feel properly trained, and only 25% of frontline employees got sufficient leadership support.
The governance issue is particularly thorny. When corporate solutions fall short, 54% of employees will use unauthorized tools anyway. This creates security risks and undermines your ability to manage AI adoption in a controlled way, amplifying unintended AI impacts on the workforce through fragmented, unsanctioned usage. It's also a waste, as you're paying for enterprise AI tools that people aren't using properly because they're not good enough or well-integrated enough.
Then there’s the supposed “95% failure rate” everybody talked about in the second half of 2025.
The comprehensive MIT & NANDA study from which this metric originates analyzed 300 public AI deployments. They found that 95% of enterprise GenAI pilots never made it to production. To contrast this, the same study showed 80%+ adoption for individual productivity (shadow AI). It is custom enterprise solutions that struggled. These PoCs pass through initial evaluation 60% of the time, then pilot (20%), but then almost never get deployed (only in 5% of cases). This isn't about model quality or regulation, but the cost of innovation when having to do internal R&D with a new technology instead of buying software off-the-shelf.
What made all the difference in some of the successful deployments was about what MIT researchers call "the learning gap."
Most enterprise AI systems don't retain feedback, adapt to context, or improve over time. This stems from a technical limitation of current-generation LLMs themselves. They're static tools in dynamic environments. Most systems can't remember previous interactions or learn from feedback, every session starts from zero.
So the core technical roadblock is memory. Solving it, even in the crudest and most inelegant way yielded massive added value. But this would need targeted custom development on the company’s part, and that’s not what’s happening most of the time. As one CIO put it: "We've seen dozens of demos this year. Maybe one or two are genuinely useful. The rest are wrappers or science projects." This is why the same person using ChatGPT for brainstorming won't trust it for high-stakes work.
The MIT NANDA research also revealed that purchasing specialized AI tools or relying on external help succeeds about 67% of the time, while purely internal builds succeed only 33% of the time, challenging conventional wisdom in sectors like financial services where firms are building proprietary systems. The barrier to scaling isn't infrastructure, regulation, or talent, it's whether systems can actually learn and adapt.
Indeed frames this differently but arrives at similar conclusions: realized impact depends entirely on whether and how businesses actually integrate GenAI tools. Many firms lack the foundational level of digitalization needed to deploy AI effectively. Model choice matters too, as picking the wrong model for a specific process leads to more unreliable outputs and erodes trust.
The OECD highlights unequal capacity to adapt. Low-skilled workers are 23 percentage points less likely to participate in job-related training than those with medium or higher skills. The people who most need reskilling are engaging least in training activities. AI systems can also amplify bias and often operate as "black boxes," making it unclear what's driving decisions in sensitive areas like hiring, firing, and performance evaluation.
CEEMET adds structural barriers specific to the European industrial context: excessive red tape, regulatory uncertainty, and a constrained investment climate that makes it harder for companies to invest in AI and see returns.
BCG's data shows another barrier that's easy to overlook: only 13% of respondents see AI agents integrated into broader workflows. Most AI usage is still tool-based and disconnected. Until AI becomes part of how work actually flows and internal system capabilities that make sense functionally, not just a thing people do on the side, you won't capture the full value.
The bottom line: your biggest controllable barriers are training quality, integration depth, leadership guidance, and whether you've given people sanctioned tools that actually work.
The evidence is clear: the "mass layoff" narrative is largely a myth, but the "business as usual" approach is actually more dangerous. We are entering a period of profound transformation where the real risk isn't total job loss, but a failure to redesign roles fast enough to keep pace with entry-level displacement and shifting skill requirements. Success will be found by leaders who stop sprinkling "AI dust" across every department and instead double down on reshaping a few core, information-heavy processes where augmentation (not just automation!) can drive genuine growth.
Ultimately, your biggest barriers to adoption are entirely within your control. Which is why leadership decisions will shape the long-term impact of AI more than the technology itself.
Employee resistance is rarely irrational. It is typically a response to poor training, inadequate tools, and a lack of leadership transparency. By moving past the pilot phase and investing in high-quality enablement, you can shift your workforce from using "shadow AI" in silos to driving measurable value within a secure, integrated framework.
The transition is already happening. Your job is to ensure it is guided by strategy rather than reacting to panic.
This post synthesizes findings from research by Stanford's Digital Economy Lab, the International Labour Organization, the OECD, BCG, LinkedIn, Indeed, CEEMET, Harvard, MIT, Microsoft, and Anthropic's Economic Index. Full list of sources:
Explore more stories
From SQL DAGs to Databricks Jobs: How We Automated Snowflake Task Migration
Migrating Snowflake Tasks to Databricks Jobs used to be tedious manual work. We built a simple automation tool that converts SQL DAGs to Databricks YAML workflows, saving 80% of orchestration migration effort.
Context is King: UX Design for Private Equity AI Tools
A tool that fits one private equity firm might completely derail another. UX designers must keep specific users in mind while fulfilling specific requirements of a highly sensitive industry. Here’s how we do it.
Databricks Asset Bundles on Isolated Networks
Network isolation requirements in regulated industries create unique challenges. Learn how to implement secure, maintainable deployment patterns with Databricks Asset Bundles.