Separate HR's own AI-assisted work from its contribution to organization-wide adoption by giving each a distinct output, owner and acceptance test. An HR team might improve an induction FAQ while also helping colleagues learn to check AI-assisted work. Those efforts can share people and lessons, but a successful FAQ does not prove that another department has the access, judgment or support it needs.
For an HR director or adoption sponsor, the immediate task is to agree these two commitments without making HR responsible for decisions it cannot control. HR can help the organization change how people work while remaining accountable for the quality of its own services. Neither responsibility requires pretending that HR owns every tool, workflow or business outcome. The wider roles and industries guide explains why a department label is only the starting point for choosing work.
Name the two outputs before assigning the work
- Internal HR adoption changes an HR service or task. The output might be a checked onboarding answer or a draft interview guide. HR checks whether it serves the recipient, preserves approved sources and follows the relevant process.
- Organization-wide enablement helps colleagues use AI within their work. The output might be supported practice or an agreed change to a manager's responsibilities. HR can contribute learning design and workforce planning; the receiving department still needs someone who can judge its work and commit resources.
McKinsey's June 2026 discussion of HR's dual mandate argues for transforming HR while supporting the enterprise. That is consulting advice about organizational design. It does not establish that the same reporting structure works everywhere, or that HR must finish its own transformation before helping another team.
The distinction is useful precisely because both efforts can proceed together. What must remain separate is the evidence used to accept them. A better HR draft is evidence about that draft. It is not a substitute for observing whether warehouse staff can recognize an incorrect stock summary.
Keep employee service, learning and recruitment distinct
“AI in HR” can hide very different consequences. Begin with the actual output and the person affected, rather than approving a whole function at once.
| HR work to consider | Output someone can inspect | Boundary to keep explicit |
|---|---|---|
Employee service | A draft answer grounded in the current induction pack | Explaining a published process does not decide an employee's entitlement or resolve a personal case |
Learning and development | Practice material tied to an approved task and its checking rules | Producing a course does not show that learners can perform the work independently |
Recruitment preparation | Draft questions based on the approved role requirements | Preparing material does not authorize scoring, ranking or rejecting applicants |
In an October 2026 Reddit discussion, a self-described sole HR practitioner at a business with about 110 employees asked what to try beyond administration, communications and templates. They already had enterprise tools available. Their question makes a practical constraint visible: access to a tool does not identify the next useful piece of HR work. The employer, arrangements and results were not independently verified.
For a small HR team, choose a task whose source and finished output someone has time to check. Do not select sensitive or consequential work merely because it consumes many hours. An approved general-purpose tool is not, by itself, permission to upload every employee record or make every employment decision.
The UK's Responsible AI in Recruitment guidance asks organizations to define a system's purpose and outputs, maintain effective oversight, examine performance and consider affected applicants. Use those questions with the relevant HR, privacy, legal and technical owners. A person clicking approval does not establish that the underlying assessment is appropriate or fair. Requirements depend on the actual use and jurisdiction.
Try the distinction on an induction FAQ
Suppose Jamie Lee coordinates HR services at a wholesaler. New starters repeatedly ask where to collect their access badge, which induction session to attend and whom to contact when their arrangements change. Jamie proposes drafting an FAQ from the company's current approved induction pack. No individual employee files are needed for this exercise.
Claire Bennett, the HR service owner, checks each proposed answer against the pack. A useful test includes a question the pack does not settle. If a new starter cannot attend the scheduled induction, the draft should preserve the need to contact the named coordinator rather than invent another session or promise an exception.
For this trial, Claire can inspect:
- Source accuracy. Does each answer preserve the approved instruction and identify its source?
- Unresolved questions. Does the draft recognize missing information and point to the actual person who can help?
- Service effort. What drafting, checking and correction work is required before the FAQ can be used?
Now suppose the same company asks HR to arrange AI practice for warehouse coordinators. The proposed exercise is checking a draft stock-discrepancy summary against approved sample records. The FAQ trial supplies experience with source checking. It does not supply the warehouse's records, operating judgment, approved access or time away from dispatch work.
One HR team, two commitments to accept
Induction FAQ
Claire checks answers against the induction pack, including the route for a question it cannot settle.
Stock-summary practice
The warehouse manager must accept the task and reviewer; IT confirms access and L&D prepares practice.
The role-specific learning guide develops the exercise itself, including a different independent attempt. Here the decision is whether the people responsible for that exercise have agreed to do their part.
Make shared ownership specific enough to act on
Joint leadership can be useful, but the label leaves many operating decisions unanswered. Who can release warehouse coordinators for practice? Who checks an inaccurate summary? Who changes the exercise when the stock records change?
In a May 2026 account on X, Alex Lieberman described a conversation with an unnamed consumer-business HR chief. He reported joint CHRO and CTO ownership alongside widely available chat tools, limited changes to ways of working and inconsistent standards for acceptable AI work. This is a secondhand account, not a verified assessment of that company. It illustrates the questions that a shared ownership label can leave open.
For the wholesaler, agree the contributions before promising a rollout:
- HR and L&D prepare the learning approach, make employee concerns visible and identify changes to skills or responsibilities that need attention.
- The warehouse manager accepts the practice task, supplies a capable reviewer and commits staff time within the operation.
- IT and the relevant data owners confirm the tool, permitted inputs, access and route for technical problems.
- The executive sponsor resolves competing priorities and resource commitments that the working group cannot settle.
These are proposed responsibilities for this case. Adapt them to the company's actual authority. If one colleague holds several roles, keep the commitments visible and check that the time is available. Do not count the same hour once for HR service improvement and again for enterprise coaching.
An IBM survey release from September 2026 also discusses unclear accountability, coordination and the work of validating AI output. Its surveys covered 1,500 senior people leaders and 8,800 employees across different geographies. These are reported experiences and associations, not proof that assigning adoption to HR causes better results.
If the unresolved issue is who has cross-company authority, compare the leadership arrangements. Adding a training obligation to HR will not supply a missing executive decision.
Review HR service quality and workforce capability separately
HR's own use can teach valuable lessons about checking, exceptions and support. It should also remain open to criticism from the people receiving the service.
IBM's account of its AskHR transformation describes early fragmentation across separate bots, consolidation into one entry point and later refinements that helped managers reach a live expert. The company also says it would pilot and iterate if starting again. This is IBM's account of its own service, not an independently established result for another employer. The lesson to investigate is the complete support experience, including what happens when an answer is insufficient.
Keep two short review records for the wholesaler:
- Claire's HR service review shows accepted FAQ answers, corrections, unresolved questions and whether a new starter can reach the right person.
- The warehouse learning review shows participants' attempts, errors they recognized, help they needed and what the manager permits next.
Neither record needs to become a large reporting system.
CIPD's April 2026 guidance on reskilling as AI changes work recommends connecting skills planning to how AI is actually being used, rather than treating it as a one-off training exercise. Its discussion draws on survey and action research; it does not establish that a particular course guarantees capability.
A useful next decision might be to continue the FAQ trial while postponing the warehouse session until its reviewer and practice time are available. It could also be the reverse. Separate records let you make that choice without calling all HR activity either successful or stalled.
Questions about HR's role in AI adoption
Can HR and IT share responsibility for AI adoption?
Yes, if shared responsibility identifies the decisions each party can make and the work each will provide. HR might coordinate learning and workforce implications while IT supplies approved access and technical support. Business managers still need to accept their workflows and commit review time. Name who resolves a disagreement instead of leaving “joint ownership” as the only answer. The wholesaler responsibilities show one possible arrangement.
Must HR finish adopting AI internally before helping other departments?
No. Internal HR trials and organization-wide support can proceed together when each has a bounded scope and sufficient capacity. Experience improving an HR service can supply useful lessons, but it does not establish readiness in another department. Agree the receiving team's task, sources, reviewer and access separately. The induction FAQ example shows which lessons transfer and which commitments remain open.
Is drafting an interview guide the same as using AI to select candidates?
No. Preparing questions from approved role requirements produces material a recruiter can inspect. Scoring or ranking applicants introduces a different decision, different information and consequences for individuals. Even draft questions need appropriate review for relevance and fairness. Treat a proposed selection use as a separate scope for the responsible HR, privacy, legal and technical owners; approval of a drafting exercise does not authorize it.
Should HR measure enterprise enablement through course attendance?
Attendance can show who took part, but it does not establish that people can perform or check the work. Retain evidence from an approved practice task, including mistakes recognized, unresolved questions and a different independent attempt. Review the result with the operational owner before deciding the permitted next use. Keep access problems and missing practice time visible rather than interpreting every incomplete attempt as a skills deficit.
Does saving time inside HR prove that the wider adoption program works?
No. A checked reduction in HR drafting or service effort concerns that task and its conditions. Organization-wide enablement needs evidence about the receiving teams' work, capability and support, including any extra checking or maintenance. Keep the two review records separate, then examine their actual resource commitments together. An internal HR improvement can inform the wider program without being counted as its outcome.
