Workers should help define the problem, the kind of assistance worth trying and what would make that assistance useful before an AI use case is selected. Give them a way to challenge the proposed purpose and revise the criteria used to judge it. Asking for comments after the task and success measures are fixed leaves much less open to influence.
Keep a visible connection between what workers ask for and the proposal that reaches the decision-maker. A facilitator's polished summary can sound reasonable while changing the work people wanted help with.
This is guidance for an adoption lead choosing work with employees. It complements the everyday-work discovery method, which helps identify a candidate from an actual task. Formal consultation, bargaining and approval requirements remain separate; use the worker-representation guide to establish the applicable process.
Begin before the candidate has a score
Suppose Alex Morgan leads customer service at an industrial-parts distributor. The team spends time preparing explanations for late deliveries. Alex proposes an assistant that drafts those explanations faster.
Dispatch coordinator Claire Bennett and service adviser Sam Taylor see a different difficulty. A dispatch log may show when goods left the warehouse without establishing what happened on the carrier's route. A customer's report of a missed delivery may arrive before the carrier's attempted-delivery record. They do not want a fluent explanation that fills that gap with an assumption.
Their proposed help is to identify the questions that must be answered before explaining the delay. The working packet for a rehearsal contains a dispatch entry, a customer query and the missing carrier record marked as unavailable. Nobody has established that the warehouse or carrier caused the problem.
That changes the candidate. Instead of producing a concluded explanation for a customer, the trial would produce investigation questions for the service adviser. Faster updates can remain a business objective, but they cannot be the only test of this form of help.
One delivery problem can lead to different kinds of help
Draft an explanation faster
Assumes the available records already establish why the delivery was late.
Prepare questions about missing evidence
Preserves the gap in the carrier record before anyone explains the delay. Trial authorization remains open.
Ask workers what the first proposal has assumed about their job. In this case, it assumed that the explanation was already known and that writing it was the bottleneck. Claire and Sam's input changes the proposed AI role before anyone compares tools or estimates a saving.
Let workers define what useful assistance means
An August 2026 study by Anna Kawakami and colleagues offers a concrete example of workers shaping that definition. Across eight workshops, 19 participants at a small Pennsylvania school-social-work organization helped develop an evaluation for AI assistance. They wanted reflective questions that challenged their assumptions, rather than an assistant that simply supplied solutions or agreeable responses.
The researchers helped scope the use case. Workers then examined examples and revised evaluation criteria, including rejecting or substantially changing refinements proposed by the researchers. Their contribution extended beyond rating answers against a finished standard.
This was an exploratory study at one organization of about 30 workers. Participation varied, and the last workshop had one participant. It did not establish improved downstream work performance or a causal benefit from participation. The useful distinction is between supplying examples for someone else's definition of success and helping shape that definition. Eight workshops are not a recommended enterprise formula.
For Alex's team, the equivalent question is what the assistance should help Sam notice. “A professional response” is too vague. “Questions that expose missing delivery evidence before we explain the delay” identifies a different contribution, with different requirements.
Compare examples before fixing the criteria
People may find it easier to explain a preference when they can point to an example. Use permitted practice material and make clear that the exercise concerns the proposed work, rather than an assessment of the employees taking part.
For the delivery-review candidate:
- Show contrasting responses to the same packet. One might confidently explain a carrier delay. Another might ask for the missing attempted-delivery record without deciding what happened. Include a credible current-method alternative too.
- Ask what each response would change in the job. Claire may notice an unsupported account of the route. Sam may notice that a technically correct question gives him no useful next contact. Keep both observations.
- Draft the criterion from those reasons. For example, investigation questions must distinguish recorded events from assumptions and identify which missing evidence would help resolve the customer's query.
- Return the wording with another example. Ask whether the criterion still captures the intended help when a carrier record is present but contradicts the customer's account. Revise it when it rewards an answer the team would not use.
These are steps for examining a proposal locally, not a validated workshop protocol. A small exercise can reveal an unclear objective; it cannot establish how every team works or authorize using private records in an unapproved service.
Keep the purpose of the exercise open to correction. If workers conclude that a clearer carrier-contact route would address the problem without AI, preserve that alternative. Participation should be able to change the proposed solution as well as its wording.
Return the translated criterion to the people who supplied it
Facilitation involves interpretation. A note saying “help us investigate” might become “generate a comprehensive explanation” as it moves into a project brief. Both sound helpful, but they ask the system to do different things.
Keep the worker's observation, the proposed interpretation and any correction close enough to compare. Give each responsibility a clear home:
| Responsibility | Contribution to the selection decision | Boundary to preserve |
|---|---|---|
Workers who perform and receive the work | Explain desired help, consequential mistakes and useful counterexamples | Participation does not imply agreement with every later revision |
Facilitator or adoption lead | Translate observations into a testable proposal and return changes for correction | A tidy summary is not evidence that its meaning was accepted |
Authorized sponsor and specialists | Decide what can be funded and which permissions or conditions apply | A popular idea does not replace required approval |
Do not equate a quiet room with an agreed criterion. In a September 2026 Reddit discussion, one commenter reported raising concerns at work, receiving a cold reaction from multiple managers and subsequently going along with the prevailing view. This is an anonymous, unverified account in a critical discussion, not evidence of how common that experience is. It illustrates why an invitation to contribute and a usable opportunity to disagree are different things.
Keep two checks separate when returning the proposal:
- A usable opportunity to disagree. Offer a way to correct the summary outside the group discussion, and explain who will see the input. Record unresolved differences instead of averaging incompatible preferences into a broad word such as productivity.
- A clear account of the revision's reviewers. If only a few people can return, say who examined the changed criterion and whose perspective is missing. Decide what further input selection needs; an attendance total is not a permission threshold.
Carry the changed proposal into selection
The decision-maker needs to see what worker input changed. For Alex's team, a useful proposal would preserve:
- The revised purpose. Prepare investigation questions for delivery queries, rather than assert a cause that the records do not establish.
- The meaningful check. Questions distinguish known events from assumptions and point to evidence that could resolve the query.
- The remaining disagreement or gap. Record any unresolved view about the proposed help, missing participant perspective or unavailable source.
- The decision still required. Name who can commit trial capacity and settle the relevant tool, information and operating permissions.
Take that proposal into the steering decision if it needs shared resources or authority beyond the local owner. Once the intended use is sufficiently clear, the output-quality method helps test criteria and reviewers on comparable work. A worker-informed definition of usefulness is a starting point for that evaluation, not proof that the assistant meets it.
Questions about worker participation in AI selection
Should workers make the final choice of AI use case?
Workers should have a meaningful opportunity to shape the problem, proposed assistance and criteria for usefulness. Who makes the final decision depends on the organization's actual authority and applicable representation arrangements. Explain which decisions their input can change and who controls funding, information permissions and rollout. The responsibility comparison separates those contributions without treating participation as automatic approval.
Is asking workers to rate AI outputs enough?
Rating outputs can help, but it leaves the intended task and definition of success untouched if those have already been fixed. Ask whether workers can challenge what the AI is meant to help them do. In the delivery example, a well-written explanation is the wrong output when the team first needs questions about missing evidence. Use contrasting examples to examine the proposed purpose before turning preferences into scores.
Who should participate when teams want different kinds of help?
Include people who perform the work and people who receive or check its result, with attention to relevant shifts, locations and working arrangements. Give them a way to explain differences without requiring an immediate shared answer. Record which perspectives informed each revision and which remain missing. The aim is a sufficiently informed selection decision, not an attendance target or an assumption that one enthusiastic group speaks for everyone.
How can a leader show that worker input changed the proposal?
Preserve a short before-and-after record with the reason for the change and the unresolved decision. For the distributor, that means showing how a request for faster explanations became a trial of investigation questions, what evidence those questions must preserve and who still has to approve the trial. Return that record to affected workers for correction. A count of suggestions or workshop attendees does not establish what changed.
Can AI decide which worker-defined criteria are best?
AI can help organize candidate wording or prepare comparisons within permitted information boundaries. It cannot establish whose preferences should govern, settle a genuine disagreement about the work or provide the authority to proceed. People must examine whether the proposed criteria preserve the intended help. If automated scoring will influence a later trial, validate the configured grader against the actual acceptance decision rather than assuming agreement with a score proves usefulness.
