Short answer
IBM’s September 2026 CHRO study can prompt a closer look at AI-enabled work, but it is not independent proof that skills are declining or that one governance model causes better results. Start with one workflow: identify its human owner, required checks, override and escalation route, and whether validation work is visible, planned, and recognized.
What IBM reported, and what it does not establish
IBM announced a global CHRO study on 21 September 2026. It said that, with Oxford Economics, surveys ran from April to June, covering 1,500 workforce-strategy executives and 8,800 full-time employees. IBM reported 71% of CHROs prioritized supervision, validation and override skills, against 29% of employees ranking judgment important. It also reported employee concerns about skill erosion, accountability, challenge safety, HR participation and unrecognized review work. Source
This is IBM-issued research summarized in a press release, not an independent causal evaluation of AI outcomes. The supplied release does not provide the complete questionnaire, response rates, weighting, country-level results, uncertainty ranges, or underlying data. Its findings can identify perceptions and patterns worth testing locally. They do not establish that AI has objectively reduced critical-thinking ability, that a governance practice causes confidence, or that another organization will reproduce the reported associations.
Read the apparent priority gap carefully
The 71% and 29% comparison is striking, but it should not be treated as one clean measure of a shared priority gap. The groups differ, and the terms differ. Supervising, validating and overriding AI outputs describes operational responsibilities. Judgment can describe a broader personal capability.
The release does not provide enough questionnaire detail to show whether respondents received comparable terms, options, scales, or question order. A leader should therefore avoid concluding that employees do not value oversight, or that CHROs understand the issue while employees do not.
A more useful interpretation is a local hypothesis: workers may experience oversight as added work, exposure to blame, or responsibility without enough time, authority, or recognition. That hypothesis can be tested in a specific workflow rather than assumed from a headline statistic.
Practical artifact: a decision-rights worksheet
Proposed worksheet, not an IBM model: use this before expanding one AI-enabled workflow. Select a bounded task, such as drafting a performance-summary note, triaging an HR service request, or preparing a manager's initial response to an employee question.
| Worksheet field | Questions to answer | Record before the test |
|---|---|---|
| Task and decision | What work is the system helping with? What decision follows? | A one-sentence task definition and the decision it may influence. |
| Human owner | Who remains accountable for the outcome? | A named role, not “the employee” or “the system.” |
| AI role | Does the system retrieve, classify, draft, recommend, or execute? | Permitted and prohibited actions. |
| Validation | What must a person check before use? | Accuracy, context, assumptions, exceptions, and source quality. |
| Override authority | Who may reject, change, or pause an output? | Authorized roles, the route to use, and a no-retaliation expectation. |
| Escalation | Which cases require another person or function? | Triggers, receiving role, and response expectation. |
| Failure responsibility | If an output contributes to a poor result, who investigates? | Incident owner, documentation location, and decision on continued use. |
| Evidence review | What will decide whether the workflow continues? | Observations, error patterns, challenge use, and worker feedback. |
Checking and owning are different responsibilities. A worker may validate an output without being solely responsible for system design, access settings, training data, or a business rule imposed elsewhere. The worksheet makes those boundaries explicit before an AI output becomes routine.
For a related application, see AI 360 Feedback: Human Controls for Synthesis. It helps readers apply the same owner, review, and override questions to feedback synthesis rather than assuming every AI use case needs identical controls.
Fictional illustration
Fictional example: A people-operations team tests an AI tool that drafts replies to routine leave-policy questions. A trained HR adviser owns the final response and checks the policy version, employee location, and exception requests. If the draft relies on an unapproved policy source, the adviser stops use for that case and escalates it to the policy owner. The team records corrections and reviews recurring failures after two weeks. This is a proposed operating design, not evidence of a customer result.
Practical artifact: an invisible-work audit
Proposed audit: observe the work around an AI output, not only the time needed to generate it. For each use of the selected workflow, record whether the worker had to:
- check facts, policy, calculations, or missing context;
- correct an error, rewrite unsuitable language, or discard an output;
- provide context the system did not have;
- manage an exception or seek a second opinion;
- explain the output to another person; or
- document an override, failure, or escalation.
Where feasible, record minutes, frequency, role, delay, and whether the activity appears in workload planning. Do not turn this audit into individual performance scoring. Its purpose is to reveal workflow effort and decision conditions, not rank workers.
A local observation may show added work, displaced work, unchanged work, or a mix across cases. That evidence is more useful than assuming that a generated answer represents the full cost of a workflow.
A bounded action plan for CHROs
- Select one workflow with a clear human decision and a manageable user group.
- Map ownership, validation, override authority, escalation, and failure review before changing the process.
- Observe review work in ordinary cases and exceptions.
- Ask participants whether they can challenge outputs, know who decides, and see review work recognized.
- Set stop conditions in advance, such as unresolved harmful errors, unclear ownership, inability to escalate, or impractical review effort.
- Review the evidence with the task owner, HR, technology owner, and affected workers before extending use.
This sequence does not promise a better AI outcome. It creates a disciplined way to determine whether a specific workflow protects judgment, makes responsibility explicit, and treats validation as real work.
Working glossary
Accountability: In this article, accountability is the named role or function responsible for investigating a poor outcome and deciding what happens next. It differs from asking a worker to click approve. A credible arrangement identifies who owns the decision, who can contribute evidence, and who can pause use when the process fails.
AI-assisted workflow: A workflow in which a system contributes information, a draft, a classification, or a recommendation while a person retains a defined review or decision role. The label alone is insufficient. The practical questions are what the person must review, what authority they hold, and what occurs when they disagree.
Challenge route: A documented way for a worker to question, reject, or escalate an AI output. It can include a named recipient, an escalation trigger, and a record of the issue. It is meaningful only when people can use it without accepting responsibility for problems outside their authority.
Critical thinking: A broad term that can include framing a problem, examining evidence, identifying assumptions, considering exceptions, and deciding whether an output is fit for purpose. It should not be used here as if the study supplied an objective test of the skill or proved that the skill has declined.
Decision rights: The allocation of authority across a workflow. Decision rights answer who may make, validate, override, escalate, and stop a decision. They should be explicit where an AI recommendation could influence another person, a policy interpretation, or a material work action.
Human-led, AI-assisted, AI-executed: Labels for different degrees of system participation in a workflow. They can be useful starting points, but labels do not themselves create safety or quality. A local team still needs to define the task, controls, responsible roles, and circumstances in which use must stop.
Invisible work: Work that may not appear in an output count or automation metric, including checking recommendations, correcting errors, adding context, handling exceptions, and documenting a decision. An audit turns this broad concern into observable activities in one workflow.
Override: A decision to reject, alter, or pause an AI output. An override right has little practical value if workers do not know when to use it, lack authority to exercise it, or face pressure to accept a recommendation.
Validation: A human check that an output is accurate, relevant, sufficiently complete for the task, and suitable for the circumstances. Validation is not a generic approval step. It should specify what must be checked and which cases require escalation instead of routine use.
FAQ
What did IBM's September 2026 CHRO study examine?
The study is an IBM-reported account of workforce views on AI-enabled work, including skills, judgment, accountability, HR participation, and review effort. It is useful context for designing local tests, but it is not independent proof that a particular AI practice causes a specific outcome.
Is the reported CHRO and employee comparison proof that employees do not value judgment?
No. The comparison involves different respondent groups and differently framed concepts. It can support a question for local investigation, such as whether workers experience AI oversight as useful authority or as unrecognized responsibility. It cannot establish employee motivation or priorities on its own.
What should a CHRO do before scaling an AI-enabled workflow?
Run a bounded review of one workflow. Define ownership, validation requirements, override authority, escalation routes, and responsibility for failures. Observe checking and correction work, test challenge routes, set stop conditions, and review local evidence before expanding use.