An organization can need workflow redesign and employee training at the same time. The sequencing question becomes easier when leaders distinguish what each activity is meant to resolve.
Employees need enough understanding of AI to participate meaningfully in designing a change. They need a sufficiently clear proposed process before they can practise their particular responsibilities. Experience with that process can then reveal something that both the design and the training missed.
So, should workflow redesign come first? Start by understanding the work and giving people the knowledge needed to examine it. Specify a bounded change, practise the affected roles and use the findings to revise the proposal. A complete redesign need not precede every learning activity.
The research discussed below does not establish one universally superior order. It does show why access, training and a prescribed way of working should be examined as different interventions.
“Training” can mean several different things
An introductory session might explain what a system can do, where its outputs can be unreliable and which uses the organization permits. That knowledge can help employees contribute to an informed discussion about changing their work.
A role-specific practice session serves another purpose. It might ask someone to inspect an AI-assisted response, resolve a missing input or decide when a case needs review. The exercise depends on knowing how the proposed process should operate.
Ongoing support addresses something else again: a new exception, a changed tool or an unexpected pattern in use.
These activities do not all belong at the same point. Delaying basic understanding until every process is settled can exclude employees from the design conversation. Launching detailed role training while responsibilities remain undecided can teach a process that nobody has actually agreed to support.
The useful question is what uncertainty the next activity should reduce.
Access alone does not specify a new way of working
In the revised abstract of their 2025 working paper, Shifting Work Patterns with Generative AI, Dillon and colleagues describe a six-month field experiment across 66 firms and 7,137 knowledge workers. Individuals were randomly selected for access to AI integrated into familiar work applications.
The researchers report changes in individual time use, without detecting broader shifts in the quantity or composition of tasks from individual-level provision. That finding concerns access to a particular tool in a particular period. It does not show that access has no value, or that coordinated redesign would necessarily produce a better result.
It does leave an important distinction visible: making assistance available to an individual does not itself determine how colleagues will coordinate their work.
A prescribed workflow is not automatically an improvement
A 2026 preprint by Farach and colleagues examined 388 employees at one retailer, all with access to the same AI tool. It investigated a structured protocol for paired AI use and, in a separate task, training that framed AI as a thought partner.
The imposed paired protocol was associated with poorer document outcomes than unstructured use. The training task's primary continuous quality result was not statistically significant; a more favorable top-score finding was exploratory and vulnerable to selective attrition.
The study has substantial limitations, including different session times, attrition, grading issues and comparisons across different tasks. It cannot establish that training should always precede redesign, or that workflow redesign generally harms performance.
Its practical relevance is narrower: a process described as AI-first still needs to be tested for the people and task involved.
An older official resource points toward a similarly contextual approach. Singapore's 2020 guide to job redesign in the age of AI discusses employee consultation and implementation phases shaped by readiness and resources. It predates today's generative AI tools and offers practical guidance, rather than experimental evidence for a universal sequence.
Follow the request before choosing the course
Consider a hypothetical internal equipment service. Employees request approved devices, an operational team checks the information and a responsible manager confirms the next step. This is an illustrative scenario, not a Learning Coach client case.
An AI assistant can help draft replies. Staff learn to use it and produce clearer messages more quickly.
Yet some requests still move repeatedly between teams. The intake form leaves out a detail required later. Different colleagues have different views about who can resolve an exception. The response can be well written while the request remains stuck.
More practice with drafting may improve the messages. It will not, by itself, decide which information the form should collect or who has authority to resolve the ambiguous case.
The organization can investigate a small set of prepared examples with the people who handle the work. Where does a request arrive? What makes it complete enough to proceed? Which decision is made next? Why does a case come back?
That investigation might reveal a learning need, a process decision or both. It might also reveal that the proposed AI use addresses only a minor part of the problem.
Give employees enough knowledge to help design the change
The people examining the equipment process need a workable understanding of the proposed assistance.
They should know, for example, that a plausible draft may contain an unsupported assumption. They should understand which information is appropriate for the approved tool and what remains a human decision.
This foundational learning can happen before the future workflow is fully specified. It equips staff to ask better questions about the proposal.
A useful activity could compare an ordinary prepared request with one that lacks an important detail. Ask the group where AI assistance might help and what would still need to happen before the request could proceed.
The objective is informed participation. The activity does not need to promise proficiency in every feature, or require people to master a process that is still under discussion.
Learning Coach's article on human learning capabilities alongside AI skills develops this wider capability question.
Specify enough of the proposal to make practice meaningful
The team does not need to redesign the entire organization before trying a bounded change. It does need to state what participants are being asked to do.
For the fictional equipment service, the proposal could specify which requests are in scope, what information is needed, what the assistant may draft and who checks the result. It should identify what happens when a case does not fit.
These decisions form the basis for role practice.
The intake colleague may need to recognize an incomplete request. The reviewer may need to detect an unsupported assumption in a draft. The manager may need to decide how an exception should be handled.
If those responsibilities are still disputed, the exercise can expose the disagreement. It should not quietly turn one trainer's interpretation into company procedure.
The distinction helps L&D too. A request for “AI training” becomes a more precise question about what people will need to demonstrate in the proposed process.
Use practice to examine the design
Role practice can reveal problems that a process diagram leaves hidden.
Give the fictional team a prepared request that appears ordinary, then introduce a missing input or an unusual requirement. Observe where participants hesitate and what information they seek.
A colleague may understand the tool but be unable to proceed because no one has specified the approval route. Another may know the route but struggle to recognize an inappropriate AI suggestion.
Those observations call for different responses.
The first may require a process decision. The second may require explanation, examples and practice. A third case may show that the application should have a narrower scope.
These are editorial suggestions for investigating a proposal, not a validated implementation method. The aim is to learn what needs to change before extending use.
Let the observed obstacle determine the next step
A conditional decision table can keep the conversation focused.
| What the team observes | What to investigate next | A possible response |
|---|---|---|
| People cannot explain what the proposed AI assistance does | Foundational understanding | Provide a short, relevant introduction before asking for design judgments |
| Colleagues disagree about who acts next | The process and decision rights | Resolve the handoff before teaching it as an established routine |
| The process is clear, but people miss unsuitable outputs | Role-specific knowledge and judgment | Practice with prepared cases and feedback |
| The same exception repeatedly defeats the proposal | Scope, inputs or assumptions | Revise or narrow the process, then update the practice |
| The bounded use works in the tested conditions | What changes in the next conditions | Extend cautiously and examine the new requirements |
The table is a discussion aid. It cannot decide whether a specific AI application is technically suitable or sufficient for a consequential use.
It does help prevent every difficulty from being labelled a training gap. It also prevents “we need to redesign work” from becoming a reason to postpone all useful employee learning.
Avoid freezing the training too early
A detailed course can become an expensive commitment to a process that is still changing.
For the fictional service, it may be more useful to begin with a small set of prepared requests, a clear account of the proposed responsibilities and a way to record unresolved questions. The learning material can grow as the team learns which distinctions matter.
That does not mean improvising without preparation. It means matching the maturity of the training to the maturity of the proposal.
A stable, well-understood change may support clear procedures from the start. A new and uncertain arrangement may need more observation, discussion and revision before a large rollout.
In either case, someone should keep the learning material aligned with the actual process. Employees should not have to reconcile conflicting instructions from a course, a manager and the tool interface.
Treat employees' experience as design information
The people doing the work can identify practical conditions that a central team misses.
Perhaps the request arrives through a channel the proposal did not consider. Perhaps the information is available, but only to a different role. Perhaps the new draft creates an additional check that the reviewer has no time to perform.
Invite those observations while the process can still be changed. Ask what happened in the case, rather than treating all hesitation as resistance.
Managers also need to make room for practice and feedback. A process can appear reasonable while offering employees no realistic opportunity to learn it.
Learning Coach's perspective is that the organization should learn through the change. Employee learning helps people understand and improve the proposal; experience with the proposal helps clarify what they need to learn next.
This is a management position to test in context, not a claim that one sequence guarantees adoption or business results.
A better brief for a leadership event
A corporate AI event can begin with a concrete unresolved decision.
Choose one fictional or appropriately prepared workflow. Ask operational, technical and L&D participants to identify what employees need to understand before contributing to its design, which responsibilities need a decision and what should be practised before use expands.
The exercise can reveal that the organization is trying to solve different problems under one label. Some participants may be discussing tool familiarity, while others are discussing an approval route or an unsuitable task.
A keynote can establish the distinctions. A facilitated session can help participants apply them. Subsequent workplace work must examine whether the proposed changes are usable.
The guide to designing a corporate AI learning summit explores how to connect that shared conversation with practical follow-through.
Decide what needs to become clearer next
The sources reviewed here were checked on 28 October 2026. They include field experiments with different interventions and limitations, and older practical guidance. They do not directly compare complete redesign-first and training-first transformation programs.
The equipment-service scenario, sequence and table are original editorial proposals. They offer a way to examine the decision, rather than evidence of a completed Learning Coach intervention.
The next step should make the work clearer: enough understanding for employees to participate, enough specification for relevant practice and enough feedback to revise either one.
Before commissioning another course or approving a redesigned process, which uncertainty does the team need to resolve next?
Bring that question to your next leadership or workplace learning event. Explore Learning Coach's international keynotes and request a tailored proposal.
References
Dillon, Jaffe, Immorlica and Stanton, Shifting Work Patterns with Generative AI. NBER Working Paper 33795, 2025, revised November 2025; individual-level access experiment and stated scope.
Farach and colleagues, Scaffolding Human–AI Collaboration. April 2026 preprint; protocol and training findings with design and measurement limitations.
IMDA and PDPC, A Guide to Job Redesign in the Age of AI. Singapore, 2020; practical guidance on task change, consultation and implementation.
