Deloitte's 2026 State of AI in the Enterprise reports that 84% of companies have not redesigned jobs around AI capabilities. On the process side the picture is different. Deloitte splits companies three ways on their current approach to AI transformation: 37% using AI at a surface level with little or no change to underlying business processes, 30% redesigning key processes, and 34% deeply transforming. So a majority report changing how the work runs while a large majority have not changed the job that runs it.
Access moved faster than either, rising 50% in 2025.
Two things are true at once. Companies are genuinely changing how work flows, and they are leaving the roles, the headcount model and the performance criteria more or less where they found them.
What the Deloitte survey measured, and when
Deloitte's AI Institute fielded the seventh edition between August and September 2025 and announced it on January 21 2026. The sample is 3,235 business and IT leaders across 24 countries and six industries, at director through C-suite level, all with direct involvement in their company's AI initiatives, split equally between IT and line of business leaders.
The fielding window matters. The data describes late 2025 and is read in late 2026, so treat it as a baseline.
The figures that carry the argument, all read from Deloitte's own report page:
- 84% of companies have not redesigned jobs around AI capabilities
- 37% are using AI at a surface level, 30% are redesigning key processes, 34% are deeply transforming. Deloitte notes on the chart that figures may not sum to 100% because of rounding
- Education, not role or workflow redesign, was the leading way companies adjusted their talent strategies in response to AI
- Worker access to AI rose 50% in 2025
- 25% of respondents have moved 40% or more of their AI pilots into production, with 54% expected to reach that level by May 2026
A note on how we read this report
Two handling decisions, both of which cut against the easier version of this story.
The three transformation figures belong to one question, so all three get reported. Quoting 37% and 30% while dropping 34% would make surface-level use look like majority behavior. It is the largest single category and it is not a majority.
The frequently quoted range "from fewer than 40% to around 60% of workers now equipped with sanctioned AI tools" appears in the key-takeaway summary of Deloitte's announcement, not in the report body, which says only that worker access to AI rose 50% in 2025. We use the body figure.
Why does an adoption dashboard miss all three?
Access is a distribution measure. It answers how many people could use the tool. Usage is a frequency measure. It answers how often they do. Neither describes the process or the job, which are the two objects that determine whether anything reaches a financial statement.
This persists for structural reasons rather than lazy ones. Entitlement and telemetry data arrive free with the contract, while process and role data have to be instrumented deliberately. So the easiest metric to produce is the one that gets reported, against a question it was never built to answer. The same error appears one level down in AI usage counts as a review metric.
Surface-level use and deep transformation draw the same chart
Consider two teams with identical adoption numbers. The first drafts documents in a chat window and runs the original process unchanged around the output. The second rebuilt its intake, routing and review steps so the model sits inside the workflow. Both show full seat coverage and rising session counts. Only one changed anything that will reach a quarterly result.
That is Deloitte's 37% and its 34% sitting on top of each other in one dashboard. The 37% are not absent from their own reporting, which would at least show up as a gap. They are recorded as a success. This is why an operating system view of the workforce has to sit above tool telemetry rather than inside it.
The denominators are not interchangeable
One caution before anyone builds a slide from these numbers. The 84%, the three transformation figures and the 25% are shares of companies or respondents. The access growth is a rise in the share of workers with access. They are not on one scale, they cannot be subtracted, and there is no single headline gap to quote.
One more limit. Deloitte publishes a year-over-year change for access and for transformative impact, which doubled, but no prior-year figure for job redesign. So 84% is a level, not a trend. Nothing here says job redesign stalled. It says most companies have not done it yet.
Do leaders already know redesign is the lever?
At the top of the house, yes. Mercer's Global Talent Trends 2026 preliminary findings, released January 14 2026 and drawing on nearly 12,000 business executives, HR leaders, investors and employees worldwide, report that 63% of C-suite leaders believe redesigning work for AI and automation will yield the highest people-related return on investment in 2026. Only 46% of HR leaders say the same. Mercer calls that a significant alignment gap, and the honest reading is that the belief is strongest one level above the function that would have to execute it.
Different survey, different sample, so these figures and Deloitte's are not a before and after of one population. Read together they establish something narrower and more useful. The constraint is not awareness. It is the instrumentation that would tell a leader whether the lever had been pulled.
What would you have to measure to tell the difference?
| Question | What answers it | Why adoption data cannot |
|---|---|---|
| Do people have the tool? | Sanctioned seat coverage by team and role | Nothing. This is the question adoption data answers well |
| Do people use the tool? | Sessions, prompts, active days | It answers this too, and then stops |
| Did the steps of the work change? | Cycle time, handoff count, rework rate and queue time, recorded before and after, on the process rather than the person | Usage records the tool that was touched, not the route the work took |
| Did the job change? | Task composition, decision rights and skill mix derived from demonstrated work, compared against the role as written | An entitlement record describes a license, not a role |
| Did output change because of the tool? | Outcome measures for adopting teams against comparable non-adopting teams over the same period | One group's numbers cannot separate the tool from everything else that moved that year |
Three implications follow.
- Keep access and usage on the input side of the model. They are real signals of coverage, rollout health and cost forecasting. The moment they are asked to be evidence of return, the measurement stops working.
- Measure the process and the job separately. They move independently in Deloitte's data, so a single "transformation" number collapses two things a CFO and a CHRO need to see apart.
- Hold a comparison group. Without one, every number is a single observation from a year in which everything else also changed.
Where this sits in the Levos view
This is the gap the AI Impact signal family was built for. It surfaces where AI adoption is high but productivity is flat, as an enablement gap for the CHRO rather than an adoption metric for the CIO, and it derives skills from demonstrated work in connected tools rather than self-reported assessments. That last point is what makes the job question measurable at all, because a role as written is a document and a role as performed is a signal. The comparison comes from controlled cohort analysis with confidence scoring, which compares adopting teams to non-adopting teams while controlling for tenure, role and tool stack, and discloses the limits.
The privacy posture is part of the method. Individual data flows to a person's direct manager rather than routinely up the chain, with a logged escalation exception for urgent cases, and team-level views require 5 or more people. That is a measurement rule as much as a privacy one, because a group of one has no error bar. Why one vendor's dashboard cannot produce this comparison is the longer argument in the AI productivity paradox.
Frequently asked questions
How many companies have redesigned jobs around AI?
Few. Deloitte's 2026 State of AI in the Enterprise reports that 84% of companies have not redesigned jobs around AI capabilities. The survey covers 3,235 business and IT leaders across 24 countries, fielded August to September 2025 and announced in January 2026, so treat it as a baseline rather than a present-day reading. Deloitte also reports that education, rather than role or workflow redesign, was the leading way companies adjusted their talent strategies in response to AI.
What is the difference between redesigning a process and redesigning a job?
A process is the route the work takes. A job is the bundle of tasks, authority and accountability assigned to a person. They can move independently, and in Deloitte's data they clearly do. On the process axis Deloitte splits companies three ways, with 37% using AI at a surface level with little or no change to underlying business processes, 30% redesigning key processes, and 34% deeply transforming. Deloitte notes that figures may not sum to 100% because of rounding. On the job axis, 84% have not redesigned jobs around AI capabilities. A company can reroute the work and leave the role description, the headcount model and the performance criteria where they were.
Why can an adoption dashboard not tell you whether AI changed the work?
Because the shallowest and the deepest tiers produce the same chart. A team that pastes drafts into a chat window and a team that rebuilt its intake, routing and review steps around a model both register as active seats with rising session counts. Deloitte's 37% surface-level group and its 34% deeply transforming group are indistinguishable on seat coverage and usage frequency. The dashboard is not silent about the 37%. It reports them as a success.
Is work redesign recognized as the priority by leaders?
At the top of the house, yes. Mercer's Global Talent Trends 2026 preliminary findings, released January 14 2026 and drawing on nearly 12,000 business executives, HR leaders, investors and employees worldwide, report that 63% of C-suite leaders believe redesigning work for AI and automation will yield the highest people-related return on investment in 2026. Only 46% of HR leaders say the same, a split Mercer itself describes as a significant alignment gap. Different survey and different sample from Deloitte's, so the two sets of figures are not a before and after of one population. What they establish together is that the constraint is not awareness.
What should an organization measure instead of adoption rate?
Measure adoption too, and keep it on the input side of the model. Then add two things it cannot supply. The first is process measures taken on the work rather than on the person, such as cycle time, handoff count, rework rate and queue time, recorded before and after the change. The second is a comparison, so that outcome measures for adopting teams are set against comparable non-adopting teams over the same period, controlling for tenure, role and tool stack. Publish a confidence level with every number, and aggregate before interpreting, so that team-level views require 5 or more people.
Start the measurement before the next renewal
If your adoption number went up this year and you cannot say what changed in the process or in the job, you are holding the same evidence as the 84%. That is an instrumentation problem, cheaper to fix before the renewal conversation than during it.
Request a Demo, or read the measurement methodology first if you would rather check the method before the product.
Levos is accepting design partner applications from US organizations of 150 employees or more with an active AI rollout. Large organizations typically start with one function or division, measured against comparable teams that have not adopted yet. That is a stronger attribution design than a company-wide before-and-after.
