Companies Froze Job Design and Called AI a People Problem

Deloitte’s 2026 State of AI in the Enterprise survey finds that 84% of companies have not redesigned jobs around AI. Executives still name a skills gap as the main barrier, and the most common talent move is a fluency course.

MIT’s Project NANDA, in July 2025, found 95% of organizations getting zero return from generative AI spend of $30 billion to $40 billion. The same research showed workers at more than 90% of firms already using personal chatbots for work.

The 95% Line Measures the Wrong Failure

The NANDA paper, The GenAI Divide: State of AI in Business 2025, was written by Aditya Challapally, Chris Pease, Ramesh Raskar, and Pradyumna Chari. They reviewed more than 300 public AI initiatives, interviewed people at 52 organizations, and surveyed 153 senior leaders from January through June 2025.

Headlines turned that work into “95% of AI pilots fail.” The authors’ own wording is narrower. Most organizations saw no measurable P&L impact. About 5% of integrated pilots were pulling out millions in value. Generic tools such as ChatGPT and Copilot were widely tried: over 80% of organizations had explored or piloted them, and nearly 40% reported some deployment. Those tools lifted personal output. They did not move the P&L.

Custom and vendor-built systems stalled in a steeper chute. Sixty percent of organizations evaluated them, 20% reached a pilot, and 5% reached production. The authors wrote that most of those efforts failed because of brittle workflows, a lack of contextual learning, and a poor fit with day-to-day operations.

THE CUSTOM-TOOL DROP-OFF

Generic chatbots looked healthier on paper, with a pilot-to-implementation rate of about 83%. Users told the researchers those tools were flexible and familiar. The same people called the custom pitches brittle, over-built, or out of step with how work actually ran. A CIO in the interview set said the team had seen dozens of demos that year and maybe one or two were useful, with the rest wrappers or science projects.

The report’s own myth list is blunt. Model quality, legal review, and data risk are not what held most programs back. The core barrier, the authors said, is learning. Most deployed systems do not keep feedback, adapt to context, or improve with use. Talent, infrastructure, and regulation sit behind that gap in their ranking. On a 1-to-10 frequency scale of obstacles to scale, resistance to new tools led the list.

Kerry Brown, lead transformation evangelist at process-mining firm Celonis, has boiled change work into a simple equation: expectations plus accountabilities. That diagnosis treats “unwillingness to adopt” as the disease. NANDA’s fieldwork points at tools that forget the last session, then at companies that never rewrote the job those tools were supposed to enter.

Workers Already Crossed Without Permission

While only 40% of companies in the NANDA set had bought an official large language model subscription, workers from over 90% of the surveyed firms said they used personal AI tools for work, often several times a day. The authors called this a shadow AI economy. It is the unofficial stack: a personal ChatGPT or Claude tab, used on real tasks, while the sanctioned pilot sits in a slide deck.

That split explains a lot of the so-called people problem. Staff who already know what a responsive model feels like have little patience for an internal bot that asks them to re-enter context every time. For quick work such as drafting mail, 70% of users in the study preferred AI over a colleague. For mission-critical work, 90% still wanted a human. The authors said the dividing line is not raw model IQ. It is memory, fit, and the ability to learn a workflow.

A mid-market manufacturing COO told the researchers that LinkedIn hype said everything had changed, while operations had not, aside from processing some contracts faster. Seven of nine sectors in the study showed little structural change. Technology and media were the exceptions. Enterprises with more than $100 million in revenue ran more pilots and still converted them more slowly. Mid-market teams that did get through reported about 90 days from pilot to full use. Large firms took nine months or longer.

External partnerships succeeded at twice the rate of internal builds. Buyers who got through asked for process-specific customization and judged tools on business outcomes, not software benchmarks. Highest-performing groups cited lower business-process-outsourcing spend, less use of outside agencies, and, in some cases, better retention and conversion from automated follow-up. Those wins sat in the back office, which is not where most of the budget went. NANDA flagged an investment bias toward visible, top-line functions over the repeatable work that actually shows up on a P&L.

Why 84% of Jobs Still Look Pre-AI

Deloitte’s 2026 survey of 3,235 leaders across 24 countries, fielded in August and September 2025, finds that companies haven’t redesigned jobs to fit AI even as they talk about automation at scale. Executives still pick insufficient worker skills as the biggest barrier to putting AI into existing workflows. Less than half say they are making a serious change to talent strategy.

The job freeze sits next to aggressive forecasts. Thirty-six percent of companies expect at least 10% of their jobs to be fully automated within a year. Eighty-two percent expect that level within three years. Worker access to approved tools rose by 50% in 2025. Fewer than 60% of workers who have access actually use AI in the daily flow of work.

On transformation, Deloitte splits the field into three rough thirds. Thirty-four percent say they are using AI to deeply change products, processes, or the business model. Thirty percent are redesigning key processes and leaving the model intact. Thirty-seven percent are using AI at the surface, with little or no change to how work runs (the three shares sum to 101 because of rounding). Twice as many leaders as in the prior year report a transformative effect. Two-thirds, 66%, already claim productivity or efficiency gains. Revenue is still mostly a hope: 20% say AI is already growing it, and 74% want that later.

Confidence is up. Eighty-four percent of organizations are raising AI investment, and 78% of leaders say they trust the technology more than before. Forty-two percent call their AI strategy highly prepared. Preparedness falls once the question moves from the slide to operations, data, risk, and talent.

HOW FIRMS SAY THEY ARE CHANGING TALENT STRATEGY

Education is the easy lever. A completion rate looks like progress in a quarterly review. Rewriting a procurement role, or an accounts-payable desk, so that a model drafts the routine and a person owns the exception, is slower and more political. Deloitte’s own write-up of the 2026 study says most firms are focused on teaching staff, and far fewer are rebuilding roles, workflows, and career paths.

Experts Cheer. The Public Does Not.

That freeze has a public face. Pew Research Center, in April 2025, asked AI specialists and U.S. adults what the next 20 years of AI would do to the country. Fully 56% of the experts surveyed said AI would have a very or somewhat positive impact on the United States. Among the general public, that share was 17%. Thirty-five percent of adults expected a negative effect, against 15% of experts.

The widest gaps sit on work. Seventy-three percent of experts expected a positive effect on how people do their jobs. Twenty-three percent of the public agreed. On the economy, the split was 69% against 21%. Sixty-four percent of U.S. adults said AI would lead to fewer jobs over 20 years, and 5% said more. Experts were split: 39% expected fewer jobs, 19% more. Forty-seven percent of experts said they were more excited than concerned about AI in daily life. That fell to 11% among the public.

EXPERTS AND THE PUBLIC, NEXT 20 YEARS

Companies keep asking staff to live with that gap. Brown’s point about executives who cannot say how jobs will change is fair as far as it goes. NANDA also found limited layoffs from generative AI so far, and no consensus among executives on hiring over three to five years. Silence is still a signal. When the official story is “we do not know,” unofficial use fills the vacuum, and Pew’s 17% is what the rest of the country hears.

Most Talent Budgets Still Buy a Course

The top talent response in Deloitte’s 2026 survey is a course. Fifty-three percent of organizations are trying to raise overall AI fluency. Forty-eight percent have an upskilling or reskilling plan. Redesigning career paths sits at 33%. Incentives for using AI, checks on skill supply and demand, org-chart experiments, and trust surveys each sit at 30%. Changing the mix of full-time, contract, and gig work is 19%.

WHAT GETS FUNDED INSTEAD OF A NEW JOB DESIGN

• Fluency classes: A shared course can be counted, scheduled, and reported without touching a single role description.
• Specialist hiring: Thirty-six percent are buying AI skills on the open market rather than rewriting the work those hires will join.
• Path experiments: A third say they are looking at career mobility, which still leaves the current job as it was.
• Contract mix: Fewer than one in five are changing how they split full-time, contract, and gig work around the new tools.

The World Economic Forum’s Future of Jobs Report 2025, published in January 2025, found that employers already say 50% of their workforce has completed training as part of long-term learning, up from 41% in the 2023 edition. Skill instability, the share of a worker’s skill set expected to be transformed or outdated from 2025 to 2030, slowed to 39% from 44%. Employers still call skill gaps the biggest barrier to transformation, at 63%. Eighty-five percent plan to put upskilling first. Seventy percent expect to hire people with new skills. Forty percent plan to cut staff as current skills fade, and 50% plan to move people from shrinking roles into growing ones.

If the world’s workforce were 100 people, the Forum’s survey says 59 would need training by 2030. Twenty-nine could be upskilled in the job they hold. Nineteen could be trained and moved inside the same firm. Eleven are unlikely to get the training at all. That is a planning number, not a headcount forecast, and it still leaves a large group with no map. Fluency training does not tell an invoice clerk whether the role becomes exception handling, vendor risk, or a smaller headcount.

The Back Office Is Where Returns Show Up

NANDA’s high performers did not win by turning every knowledge worker into a prompt engineer. They aimed at repeatable sequences: invoices, order-to-cash, procurement, support deflection, the work that already has a cost baseline. Selective workforce effects showed up in customer support, software engineering, and admin, including roles that used to sit with an outsourcer. The authors were careful: most implementations they saw did not drive broad headcount cuts.

That is the boring layer Brown wants mapped. Official process charts rarely match the workarounds, hidden handoffs, and exception piles that keep a company running. A model dropped onto the documented process will miss the real one, then get blamed for bad output. Process intelligence and a so-called digital twin of the firm are one vendor answer to that mess. The underlying need does not depend on the brand: if the company cannot see which steps are stable and which need a person, it cannot tell staff which skills will still be paid for.

THE FORUM’S 100-WORKER SKETCH, 2025 TO 2030

• No major training: 41 of 100 workers, on the employer view, will not need a big reskill by 2030.
• Upskill in place: 29 could be trained in the job they already hold.
• Train and move: 19 could be reskilled and shifted inside the same organization.
• Left out: 11 are unlikely to get the training, which the Forum treats as rising risk to their employment.

Automation of “the boring stuff” still needs a human on the loop for the case the model cannot see, a geopolitical shock that rewrites vendor risk, a one-off invoice, a customer who is not in the script. NANDA’s users already vote that way: AI for the simple pass, a person for the work that has to be right. Companies that leave 84% of jobs in their pre-AI shape are asking the model to sit beside a role that was never redesigned to use it, then scoring the result as a failed pilot.

A Thin Slice Burns Most of the Tokens

A year after NANDA, the 95% line is still the open of vendor pitches. The part of the report that did not travel as far is the 83% implementation rate on generic chat tools, and the shadow stack sitting under the official program. Seat counts keep rising. Output often does not.

Vasuman, who leads the enterprise-agent firm Varick and previously worked on AI at Meta, described that pattern in August 2026 after watching large operations rollouts. In a typical Claude Cowork-style launch, he wrote, about 5% to 10% of staff become power users, about 20% use the tool a little and poorly, and about 70% barely touch it. Dashboards still register adoption. The operations leader who bought the licenses sees no speed-up. In another company that had just made an eight-figure license commitment, about 10% of people burned 90% of the tokens.

People are not in the market for a tool that helps them get the work done. They just want the work done.

Vasuman, CEO of Varick, previously on AI at Meta

Prompt class, in that view, is a small slice of the work. The rest is deciding which flows should never touch a model and which should run every day inside the systems staff already open, Salesforce, NetSuite, a payables queue, with a person only on the second glance. Power users have little reason to close the gap on their own. Their edge is the gap.

That is the bind the 84% freeze creates. Companies buy tools, fund fluency, and report access. They leave the job as it was, then read low official uptake as fear or stubbornness. Workers who already live in a personal chatbot are not refusing AI. They are refusing a sanctioned system that does not remember, a role that has not been rewritten, and a story from the top that still cannot say which parts of the job will still exist. Until those three things move together, the P&L line NANDA published in July 2025 will keep looking like a people problem, and it will keep being a design problem.