Dialogic Leadership in AI Adoption: A Pilot Decision
Nour
Founder, Coachful
Dialogic leadership in AI adoption helps people examine a specific proposal, question the assumptions behind it and contribute to the decision. It does not establish that a system is reliable, that information can be used lawfully or that everyone has agreed to a workplace change.
The useful shift is from “Are we for or against AI?” to “What work are we considering, what would count as acceptable evidence, and who can authorize the next step?” Someone may welcome experimentation while rejecting an unsupported answer, an inappropriate upload or hidden review work.
This guide combines dialogue and a polarity lens with an original, completed evaluation case. The final decision preserves useful learning while refusing live deployment. The errors, additional work and unresolved source conflict remain in the record.
Understand what dialogue and polarity thinking contribute
William Isaacs describes listening, respecting, suspending certainty and voicing a considered view as practices of dialogic leadership. Suspension means making your thinking available for examination, rather than suppressing it or treating it as unquestionably correct. These are conversational practices, not an AI assurance method.
Barry Johnson’s polarity approach asks whether a difficulty continues over time and involves interdependent needs. Learning and dependable service can be such a tension: a team needs to improve its work and continue serving people well. Committing entirely to either a permanent freeze or unexamined experimentation can create problems.
That does not make every disagreement a polarity. A broken permission setting needs correction. An unsupported statement needs examination. A legal restriction needs an appropriate response. “Innovation versus privacy” is not an instruction to compromise an obligation halfway. A choice of vendor or a decision about one release can also be a finite decision within a wider recurring tension.
Use the polarity management guide for the general approach. The questions, meeting structure and records here are original applications, not a proprietary map or validated AI-readiness assessment.
Make four candidate tensions concrete
| Recurring needs to investigate | Useful contribution from each | Warning to examine |
|---|---|---|
| Learning and dependable service | Learn from new possibilities while keeping a reliable current route | A demonstration becomes live advice before its limitations are understood |
| Local usefulness and shared controls | Workers define meaningful tasks; responsible owners maintain information and system boundaries | A general policy ignores the task, or a local experiment bypasses required approval |
| Participation and clear responsibility | Affected people contribute evidence; named people still decide and answer for the work | An open meeting is described as consent or nobody owns an unresolved concern |
| Timely action and considered evidence | Decide within a useful period without pretending unknowns are resolved | Either endless discussion or a deadline used to overrule an unmet condition |
Test these labels with the actual participants. A worker may say their issue is not participation but insufficient time to perform review. That correction changes the proposal. Do not praise a beautifully balanced diagram while leaving the concrete condition untouched.
For each relevant tension, ask what the team wants to preserve, what excessive reliance on one side would look like, and what observable signal would require a response. An example is a reviewer reaching the allocated limit with cases still unassessed. The signal must lead to a decision about time or scope, rather than an instruction to finish invisibly.
Keep the technical and governance evidence separate
The NIST AI Risk Management Framework is voluntary guidance for incorporating trustworthiness considerations into AI design, use and evaluation. NIST’s current page states that AI RMF 1.0 is being revised. Its July 2024 generative-AI profile is a companion resource, not a product certificate or a universal checklist proving compliance.
The generative-AI profile discusses confidently presented false content and recommends checking sources and citations. It also warns against extending conclusions from narrow, unsystematic assessments. A fluent answer with a link can still omit an important condition.
Those points support a careful question: what was tested, against what source, by whom and under which conditions? The qualified organization owners must address the actual system, rights, information handling, affected people and applicable requirements. A coach can help the leader prepare that discussion without becoming the system’s safety assessor or legal adviser.
The case below is deliberately narrow. Its local review rules illustrate a workplace decision; they are not claimed to be NIST’s prescribed thresholds or a complete risk assessment.
Worked case: change a public-information evaluation before it starts
All people, documents, counts and events in this case are fictional. Jules directs a community-centre service. Noor works at its front desk, Evan works part time, Sana owns the published visitor information, Leo manages the proposed technical environment and Mina is the experienced information reviewer. Jules has individual coaching with Beth.
The proposed system would generate answers with references to nine approved public pages about rooms, opening arrangements and visitor services. The centre already has an index linking those pages and a route to ask its information team. No staff, visitor, booking or payment records are in scope.
The initial proposal says “a quick pilot will save staff time.” Neither the saving nor the review demand has been established. In coaching on 2 November, Jules says Noor is resistant to change. Beth asks what Noor actually said. Jules checks his note: Noor had asked who would verify an answer during a busy shift.
Jules revises his preparation: “I treated a workload question as a position against AI. I need to ask what review would involve and who could do it.” That is a change in his account, not evidence that Noor supports the proposal.
1. Invite a decision conversation with a usable route
Jules sends the other five participants an invitation for a forty-five-minute meeting on 3 November. It explains that the open decision is whether to authorize an offline evaluation, not live use or replacement of their roles. Participants can reply to the existing operations email thread before the meeting; Jules will include their attributed points in the shared discussion record. Personal concerns can be raised directly with him outside that shared thread.
“Please identify a visitor-information question that the current index handles poorly, a condition an answer must preserve, or work needed to evaluate it. You can reply to this thread or bring it to the meeting. We will record the proposed scope, outstanding conditions and who can decide. Participation in this conversation is not approval of deployment or employee monitoring.”
Jules does not promise anonymity to six known colleagues. Evan replies in advance because his shift ends before the meeting: generated answers must distinguish normal hours from holiday exceptions. Jules acknowledges the point and arranges for Evan to join the full meeting as paid work.
The meeting uses five minutes for purpose and authority, ten for actual work examples, fifteen for possible scope and alternatives, ten for conditions and capacity, and five to read back decisions. That is forty-five minutes, with no hidden presentation added to the agenda.
2. Let the objection alter the proposal
Jules: I assumed the pilot would free up time. We have not measured the work of checking it.
Noor: I do want an easier search. I do not want a visitor waiting while I verify a generated paragraph.
Mina: An offline assessment would let me compare each response with the current page. That is different from front-desk staff approving answers during service.
Evan: Please include exceptions, not just the normal-hours question that works in the demonstration.
Jules: Then the proposal is a bounded evaluation away from the desk, with the current index still used for visitors. We need explicit permission and allocated review time before it starts.
Listening has produced a factual distinction between finding information and verifying an answer. Respect allows Noor’s question to stand without requiring enthusiasm. Jules voices his preference to investigate while making its assumptions examinable. There is still a finite approval decision.
The group retains an ordinary alternative: improve the index’s labels and visibility of exceptions. It will not have to prove that AI is unusable before anyone can fix a confusing index.
3. Complete the brief and obtain actual permission
| Field | Completed evaluation brief |
|---|---|
| Question | Do twenty selected generated responses accurately preserve the relevant conditions in the approved public information? |
| Inputs | Nine specified public pages and twenty written fictional visitor questions, including exception questions; no individual records or free-form staff uploads |
| Environment | Leo’s approved restricted evaluation environment, not the visitor-facing website or operational booking system |
| Prohibited use | No answers sent to visitors, no booking decisions, personal data, safety advice, staff ranking or expansion to other collections |
| Reviewer | Mina checks the complete response against the current source, including exceptions; uncertain cases remain uncertain |
| Initial allowance | Twenty minutes of reviewer setup, eight minutes average initial review for twenty responses, plus four additional fifteen-minute escalations |
| Pause rule | No further generation in an affected category after an unsupported answer; no additional review beyond the allowance without a decision |
| Decision owner | Jules can authorize this limited staff allocation; technical and information permissions must also be confirmed by Leo and Sana |
| Record | Question, generated answer, exact source/version, support judgment, omitted condition, time, escalation and decision in the organization’s evaluation register |
On 3 November, Sana explicitly approves the nine-page collection and excludes personal records. Leo confirms the environment, access and handling arrangements for this bounded task. Mina accepts the review method and time. Noor and Evan agree to contribute selected questions under these conditions; neither is committed to deploying the system.
Jules then authorizes the offline evaluation and the scheduled work time. The participants receive and acknowledge the saved brief. These are stated fictional permissions for a specific setting, not proof that public availability alone permits every use or that a short meeting exhausts procurement and legal review.
Retain the unsupported response and the exhausted allowance
Leo produces the twenty responses on 4 November before Mina begins review. A later pause therefore prevents new generation in the affected category; it does not destroy the existing material needed to understand the failure.
On 5 November, Mina finds that response seven says a meeting room is available on every weekday. Its cited page says weekday booking is available during staffed opening periods and lists a holiday closure. The answer has dropped the condition. A relevant citation has not made its unconditional claim supported.
Mina records the original answer and source passage in the approved register. Jules confirms the category pause that day; Leo acknowledges it and disables further generation for that category in the evaluation. Noor continues to use the normal index and information-team route for visitor work. No generated answer has been released.
After the initial review of all twenty responses, the record has fourteen supported as written, four unsupported and two uncertain. Mina spends the four planned escalation slots examining the four unsupported cases and confirms their unsupported status. The two uncertain cases remain unassessed at that deeper level.
The time record is explicit: twenty minutes of setup, 160 minutes of initial review in total and sixty additional escalation minutes. That is 240 minutes, or four reviewer-hours. Mina stops at the agreed allowance. Describing all twenty as fully resolved at this point would hide two open cases.
On 9 November, Jules requests that Mina use another thirty minutes to examine the two uncertainties. Mina confirms availability. Jules reallocates thirty minutes from a nonurgent internal presentation review and authorizes the extra work before it occurs; the presentation owner acknowledges the postponement. This is a real trade-off, not an instruction to absorb more work.
On 10 November, Mina uses the additional thirty minutes. One case becomes a fifth unsupported answer because the generated wording removes a booking condition. The remaining case has conflicting source statements that Sana has not yet resolved. It stays uncertain, with a named information-owner action.
| Review point | Supported as written | Unsupported | Uncertain |
|---|---|---|---|
| Initial review and four escalations | 14 | 4 | 2 |
| After the authorized extra review | 14 | 5 | 1 |
The final classifications still total twenty. Reviewer time is now 240 + 30 = 270 minutes, or four hours thirty minutes. The additional work has improved the record without converting an unresolved policy conflict into a reliable answer.
Make a completed decision without overstating the evidence
At the 17 November review, Jules initially calls the result “seventy percent accurate.” Mina corrects him: fourteen of twenty selected responses were judged supported as written under this review. The exercise does not establish a stable accuracy rate across future questions, people or model versions.
Jules accepts the correction in the meeting record. The twenty questions were deliberately selected, not a representative random sample. No comparable timed baseline was run, so the evaluation cannot establish time saved. The unsupported answers and the uncertain source remain reasons not to authorize the proposed live use.
Jules records this decision on 18 November:
“Live deployment is not authorized. The affected generation category stays paused and the offline evaluation ends here. The current index and information-team route remain authoritative. Sana will resolve the conflicting page statements and bring an approved correction for review on 2 December; she has accepted that review date. Any further AI evaluation requires a new scope, evidence plan and time allocation.”
He shares it with all six participants. Each acknowledges the decision. Noor says the process addressed the verification-work question. Evan still prefers the index to generated answers, and Jules still sees possible future value in AI. The decision does not require those preferences to converge.
The index improvement proceeds separately. Noor proposes adding the source date and a visible holiday-exception label to its relevant entry. Sana approves the exact wording on 19 November; Noor saves it on 20 November, and Sana checks the live index entry that day. This completes a modest current-service repair without claiming to resolve the AI proposal.
Sana’s deeper source-conflict correction remains due for the future review. The case does not invent its completion or a later deployment. Jules’s coaching closes on 23 November with those facts intact.
Count the work that the conversation made visible
The four-hour review estimate was only one part of the effort. The completed case records these separate contributions:
- Six people in the forty-five-minute decision meeting: 270 staff-minutes.
- Mina’s setup, initial review and all escalation work: 270 minutes.
- Leo’s collection preparation, generation and pause implementation: ninety minutes.
- Jules and Sana’s brief permissions and final decision administration: sixty minutes combined.
- Noor’s index edit and Sana’s check: thirty minutes combined.
That is 270 + 270 + 90 + 60 + 30 = 720 minutes, or twelve staff-hours. The sixty-minute permissions and decision entry excludes their meeting attendance, and the index-check time excludes Mina’s response review, so those rows do not double-count the same work.
The total still excludes private coaching and any future source correction. It measures observed work within this fictional record, not every organizational cost. No wage rate, financial return or avoided visitor error is inferred. The point is to make the chosen investigation’s human cost visible before proposing expansion.
If reliable review needs more time than the original task, that may change the business case. It does not automatically make evaluation useless. Learning that a proposal is not ready can justify a bounded investigation while leaving the claim of operational savings unproven.
Use a repeatable decision record for the next proposal
- Name the work. Describe the task and actual users, not a department-wide ambition to “use AI.”
- Separate tensions and requirements. Identify recurring needs worth managing, finite defects to fix and conditions that cannot be negotiated away.
- Invite evidence through an actual route. State the audience, the open decision and how contributions outside the meeting receive a response.
- Confirm authority and time. Obtain the relevant owners’ permissions and the reviewers’ accepted capacity before starting.
- Preserve the original output. Keep unsupported and uncertain cases, source versions and any later correction distinguishable.
- Decide and communicate. State what is authorized, blocked or deferred, with accepted owners and dates for remaining actions.
For the next proposal, Jules can reuse this short template: task; affected roles; current alternative; permitted inputs; excluded use; reviewer and standard; planned time; pause condition; decision owner; observed evidence; remaining uncertainty; actual decision and acknowledgments. Reusing fields does not mean reusing the previous approval.
Discuss what happens if someone cannot assess an answer, if a reviewer is unavailable, or if the source itself is wrong. “Human in the loop” does not answer those questions. A person needs usable evidence, appropriate competence, time and actual power to pause or escalate.
Support the leader’s preparation in Coachful
Coachful is our #1 recommendation for organizing the individual leader’s coaching preparation, agreed actions and reflection around this decision. Keep the organization’s authoritative AI permissions, system details, test outputs and incident records in its approved systems. Coachful is not presented here as the AI evaluation environment or a compliance certification service.
Beth is Jules’s sole coach in a private coaching workspace. She assigns an original preparation worksheet through Clients > Jules > Worksheets & resources > Assign. Its prompts separate the leader’s observation, assumption, contribution and question for the meeting. Jules shares only the permitted summary, rather than copying employee concerns or technical test data.
They use ordinary dated tasks for requesting the decision meeting and revisiting the actual outcome. At review, an authorization task is complete only when the relevant permission exists; a future correction remains pending. The pictured goal view can organize that follow-through without judging whether an AI system is reliable.
Jules and Beth agree one brief clarification in their one-to-one Messages conversation by 4 November at 12:00, Europe/London, with Beth’s response due by 17:00. Jules asks at 09:00 whether agreement to the offline exercise permits a live demonstration for visitors. Beth replies at 13:00 that the authorization covers the stated offline scope; he needs the appropriate new decision before changing it. No continuing technical monitoring or incident response is promised.
Unshared journal and takeaway pages remain private to the client. Where enabled, Share with coach lets the client deliberately share an individual page with the coach. Keep sensitive personal words unshared. Jules chooses what to share. For any summarized session, Beth checks the automatic recap default, Approve before sending and session-type overrides before use. She reviews held content against the agreed audience. A sponsor or worker does not automatically receive Jules’s private reflections.
The digital-transformation guide addresses wider steering decisions. Use Coachful’s features for the coaching work that supports this specific decision.
Frequently asked questions
Does dialogic leadership mean everyone must agree to AI adoption?
No. It supports examination of views and evidence, including the leader’s assumptions. A responsible decision may proceed narrowly, remain conditional or stop. Record disagreement honestly instead of treating participation or silence as consent.
Is caution about AI always one side of a polarity?
No. A concern may identify an ordinary defect, an unmet requirement or a specific unsuitable use. Use the polarity lens only where recurring interdependent needs actually help explain the situation. It does not dilute obligations.
Does a citation make a generated answer trustworthy?
No. The source must support the actual claim, including relevant conditions. In the fictional case, a correct page reference accompanied an answer that omitted a closure exception. The review preserved that failure.
Can a small pilot prove an AI tool saves staff time?
Only appropriate evidence can support the specific claim. Count preparation, review, escalation and maintenance, and use a meaningful comparison with quality held in view. The case has no timed baseline and therefore makes no savings claim.
What should a coach do when technical or legal expertise is needed?
Help the leader identify the decision, responsible owner and appropriate consultation. Remain clear about your competence and role. A coaching conversation, worksheet or software task does not supply technical assurance or legal authorization.
