Part one · Read the signalResistance is a label, not a diagnosis
A service team is told to use a new AI assistant. Usage stays low. One manager calls the team resistant, another schedules more training and the programme adds an adoption target. Yet agents can resolve an ordinary request faster than they can copy its history into the assistant. Their conduct is real. The proposed motive is not.
Organisations routinely compress refusal, delay, argument, silence, selective compliance, workarounds and withdrawal into one word. The compression feels useful because it locates the problem inside a person or group. It also destroys the information needed to manage the change. The same avoidance can indicate poor workflow fit, unreliable evidence, unclear liability, excessive checking work, loss of professional discretion, intrusive monitoring, inaccessible design, a broken consultation process, justified distrust after an earlier incident or simply a use case with no recurring value.
The managerial task is not to make every objection true. It is to convert an ambiguous signal into a claim that can be examined without punishing its source. That requires three separations. Observed behaviour is what happened: a clinician did not accept a recommendation, a dispatcher kept a parallel spreadsheet, a worker stopped reporting exceptions. Interpretation is a provisional account of why. An adjudicated condition is the finding that survives evidence, rival explanations and competent challenge. Senior repetition cannot promote an interpretation into the third category.
This paper defines the unit of success as a correctly interpreted signal that improves or stops the work system without suppressing voice. That unit is stricter than adoption. It permits the conclusion that the tool should be repaired, authority should change, a refusal should remain protected, or the proposed change should end. The aim is not critic satisfaction. It is a proportionate decision about a specified task, population and consequence.
A signal can be mixed. A middle manager may fear a loss of status and correctly observe that the escalation route is defective. An operator may dislike a new schedule and also possess precise evidence that its optimisation criterion produces unsafe handovers. Treating the personal element as proof against the operational claim commits a category error. Equally, finding a process defect does not sanctify every response. Threat, sabotage, discriminatory obstruction, deliberate data poisoning and unauthorised disclosure remain governable conduct.
Thought experiment: identical conduct, opposite causes
Two analysts ignore the same forecast. The first has better local evidence and prevents a costly stockout. The second is reacting to one memorable model error and creates excess inventory. A dashboard records two overrides. A resistance narrative records two difficult employees. A proper inquiry finds one protective correction and one calibration problem. Behaviour alone cannot choose the intervention.
Part one · Read the signalTwelve objects behind the conduct
The Resistance Signal Taxonomy classifies the possible object of a signal, never the personality of its source. Record what was observed, for which task and population, and with what consequence before selecting one or more provisional classes.
Value or premise challenges whether the use case solves a recurring problem worth its full cost. Task and workflow fit concerns handoffs, context capture, latency, exceptions and the practical sequence of work. Evidence and trustworthiness asks whether performance, limitations and comparative evidence are adequate for the local population and consequence. Authority and accountability concerns who may rely, override, accept exposure and answer for the result.
Workload, capability and support includes checking burden, transition work, staffing, learning and recovery capacity. Professional, ethical and rights boundaries cover duties, privacy, dignity, safety, fairness, accessibility and contractual limits. Identity, status and meaning concern discretion, craft, recognition and belonging. Distribution, job quality and income asks who gains, who absorbs risk, and what happens to hours, pay, progression and security.
Process, voice and procedural fairness addresses disclosure, consultation, bargaining, appeal and whether input could influence the decision. Prior experience and credibility captures unresolved incidents, broken promises and missing remedy. Technical, contextual and accessibility fit includes device, language, assistive technology, connectivity, data quality and operating environment. Genuine non-use or obsolete case recognises that low volume, absent data, correction cost or a changed process can invalidate the original premise.
| Class | Clues | Evidence to seek | Plausible rival | Likely owner | Unsafe shortcut |
|---|---|---|---|---|---|
| Value or premise | No recurring need; correction costs exceed benefit | Volume, baseline, total effort, alternatives | Benefit exists but is poorly communicated | Business outcome owner | Mandate usage to validate the business case |
| Task and workflow fit | Duplicate entry, context loss, slow handoffs | Observed episodes, exceptions, time and rework | Temporary learning friction | Workflow and product owners | Offer generic training |
| Evidence and trust | Population shift, opaque limits, salient errors | Comparative outcomes, subgroups, failure modes | Aversion after a memorable miss | Model and outcome owners | Advertise aggregate accuracy |
| Authority and accountability | Nominal oversight, unclear liability, weak override | Decision rights, time, logs, escalation tests | Reluctance to accept a legitimate role | Accountable executive | Call a person the human in the loop |
| Workload and support | Hidden checking, queue growth, no recovery time | Full load, staffing, competence, support response | Poor local prioritisation | Operations and workforce | Add learning outside paid work |
| Professional, ethical or rights | Duty conflict, privacy or safety concern | Standards, law, contracts, affected-party evidence | Boundary asserted too broadly | Professional and rights owners | Pressure first, investigate later |
| Identity, status or meaning | Loss of discretion, craft or recognition | Role evidence, interviews, decision changes | Real system defect coexists | Line and workforce owners | Psychologise the claimant |
| Distribution, jobs or income | Hours vary, progression narrows, risk shifts | Pay, grades, scheduling, load and benefit allocation | Short transition effect | Employment and business owners | Use average productivity as consent |
| Process and voice | Late disclosure, nominal consultation, silence | Decision chronology, influence, retaliation reports | Consultation happened but could not decide | Change and employee relations | Count meeting attendance |
| Prior credibility | Earlier incident unresolved; promise not kept | Incident trail, remedy, commitments, recurrence | Current system materially differs | Sponsor and vendor owner | Demand a fresh start without closure |
| Technical, contextual or accessible | Device, language, connectivity or interface barrier | Representative tests in operating conditions | Configuration or support gap | Product and accessibility owners | Exclude the affected cohort |
| Genuine non-use | Rare event, absent data, obsolete task | Current demand, correction effort, counterfactual | Latent need not yet observed | Portfolio owner | Keep funding to protect sunk cost |
The taxonomy prevents three common errors. First, an aggregate adoption curve cannot reveal which condition is binding for which population. Secondly, one class does not mechanically prescribe one response. A workload signal may require integration, staffing, narrower scope or termination. Thirdly, the classification can remain uncertain. A useful record can say that workflow fit and professional authority are both credible while the available evidence cannot yet distinguish them.
Part two · Turn conduct into evidenceProtect, test rivals, route and revalidate
The Resistance Inquiry Protocol is: observe, protect, name the object, specify the claim, test rivals, route, decide, then close and revalidate. Its sequence matters. Asking for candour after a manager has publicly branded a team obstructive does not restore the evidence environment.
Observe records behaviour, task, population, timing and consequence without motive. Protect establishes confidentiality, non-retaliation, representation and an independent escalation route proportionate to the exposure. Name the object uses the taxonomy provisionally. Specify the claim converts “the tool is unsafe” into a proposition such as: for urgent minority-language cases, the assistant omits a material exception that reviewers cannot detect within the available handling time.
Test rivals asks what else would generate the same conduct and what observation would weaken each account. A memorable failure may cause excessive rejection; a real population shift may make that rejection sensible. Route sends each claim to someone with evidence access and authority over the implicated condition. Decide chooses a proportionate repair, boundary or stop, states what was not chosen and protects interim operations. Close and revalidate means demonstrating that the condition changed and reopening the finding after material change.
The Resistance Evidence Matrix holds the inquiry trace. It records the behaviour and reporting channel; object of concern; testimony and protection; task, population and consequence; provisional classes; testable proposition; competing explanations and disconfirming evidence; technical, operational, professional, workforce and affected-party evidence; relevant history; decision, evidence, voice and challenge owners; deadline, interim protection and escalation; and closure plus revalidation.
Keep three columns visibly separate: observed behaviour, interpretation and adjudicated condition. A direct quotation belongs with testimony, not proof of a system effect. A usage log belongs with behaviour, not motive. An experiment may reject one rival without establishing every cause. A decision may be necessary under uncertainty, but the record must preserve the uncertainty rather than converting urgency into false certainty.
Useful inquiry is time-boxed. Low-consequence workflow friction may need a short observation and reversible test. A rights, safety, employment or material customer claim may need immediate restriction, representation and independent expertise before normal testing continues. Time-boxing prevents endless consultation, but it cannot be used to outrun a statutory, contractual or professional process.
Part two · Turn conduct into evidenceWhen refusal is part of the control system
A person cannot provide meaningful oversight merely because a process diagram places them after the model. The Protective Refusal Test asks whether nine operating conditions exist. A material absence means refusal may be a control response rather than adoption failure.
The conditions are: the requested action is legally, professionally and contractually permitted; evidence is adequate for the specified population, consequence and environment; the reviewer is competent, informed and holds real authority; errors are sufficiently detectable with time to intervene and recover; privacy, dignity, fairness, accessibility, safety and other rights are protected; an independent, confidential and non-retaliatory route exists; action is reversible or has effective recourse and remedy; the person has the time, information, tools and capacity to act; and the organisation retains ownership rather than shifting responsibility to the last human.
The test is conjunctive only where a condition is material. A harmless drafting aid does not require the same evidence or recourse as clinical triage. Proportionality prevents bureaucracy from swallowing routine work. It also prevents low statistical frequency from being mistaken for low consequence. The record should state which conditions apply, what minimum evidence supports them and who can authorise an interim boundary.
| Condition | Minimum evidence | If missing | Interim protection | Competent authority |
|---|---|---|---|---|
| Permitted action | Applicable law, contract, policy and professional standard | Person may be asked to breach a duty | Do not compel the action | Legal and professional owner |
| Adequate evidence | Relevant population, consequence, environment and version | Unknown error can be transferred to operations | Narrow population or pause | Outcome and model owners |
| Competence and authority | Role mandate, knowledge, override and escalation trial | Oversight becomes ceremonial | Restore real decision rights | Accountable executive |
| Detection and recovery | Observable error, intervention time and tested recovery | Human receives liability without control | Require second channel or restrict automation | Operations and safety owner |
| Rights and safety | Privacy, dignity, fairness, access and safety assessment | Benefit rests on unexamined harm | Protect affected population | Rights and risk owners |
| Independent route | Confidential channel, non-retaliation and escalation | Silence can masquerade as consent | Accept anonymous or represented evidence | Independent challenge owner |
| Reversibility and remedy | Rollback, appeal, correction and redress tested | Consequences can persist without recourse | Preserve manual path | Service and recourse owner |
| Time and capacity | Work sampling, queue load, tools and staffing | Checking displaces core service | Reduce volume or add capacity | Operations owner |
| Organisational ownership | Named owner for design, monitoring, incidents and remedy | Responsibility falls on the last user | Reject liability transfer | Executive outcome owner |
A protective boundary is not necessarily permanent. Evidence can improve, workflow can change and authority can be repaired. Nor is refusal automatically protective. Someone who has adequate evidence, capacity, authority and recourse may still reject a bounded tool because of habit or an exaggerated memory of error. Comparative testing, calibrated override review and a reversible operating trial can distinguish that case without humiliating the person or erasing professional expertise.
Part three · Examine the change makerThe programme is part of the causal system
The Change-Agent Contribution Audit asks what sponsors, product teams, vendors and managers did or failed to do before locating the difficulty in recipients. It is a symmetrical test, not an excuse for misuse or obstruction.
Audit inflated benefit promises, selective evidence and inconsistencies between stated principles and actual incentives. Check whether monitoring, job, liability or workload consequences appeared only after commitment. Determine whether consultation occurred early enough to change the design. Look for use pressure imposed before product repair, repeated failures left unacknowledged, feedback accepted but never closed, checking and liability shifted downstream, dissent discouraged, vendor dependence hidden and remedy absent.
A programme can manufacture the conduct it later calls resistance. Compulsory adoption targets encourage copied outputs and hidden manual work. A manager who rewards visible usage but penalises slower handling tells staff to conceal checking. A consultation that invites comments after the contract is signed teaches participants that voice has no effect. A vendor incident without remedy converts a technical event into an organisational credibility deficit. These are not communication gaps. They are design contributions.
| Action or omission | Likely signal | Evidence | Required repair | Accountable sponsor |
|---|---|---|---|---|
| Benefits inflated or evidence selected | Distrust, counterexamples, public challenge | Business case versions and excluded findings | Correct claim and publish limits | Business sponsor |
| Words conflict with incentives | Copied use, hidden checking, gaming | Targets, rewards, handling time and sanctions | Remove the conflict and protect disclosure | Operating executive |
| Monitoring, jobs or liability disclosed late | Opposition, withdrawal, representation request | Decision chronology and communications | Pause, disclose, consult or bargain | Change sponsor |
| Consultation cannot alter design | Cynicism, attendance without contribution | What remained open and what changed | Reopen material decisions | Programme owner |
| Use pressure precedes product repair | Workarounds and parallel processes | Defect backlog, deadlines and usage demands | Repair or narrow the workflow first | Product owner |
| Errors unacknowledged; feedback unclosed | Repeated escalation or silence | Incident, response and closure trail | Acknowledge, remedy and verify | Service owner |
| Checking work and liability shift downstream | Refusal, backlog, professional objection | Before-and-after task and authority map | Resource checking and retain ownership | Outcome owner |
| Dissent suppressed | Artificial unanimity and late surprises | Retaliation reports, meeting and case records | Independent protected route | Executive and assurance |
| Vendor dependency or no remedy | Refusal to trust releases or incidents | Contract, exit, service and remedy terms | Remediate, substitute or exit | Vendor owner |
Thought experiment: the perfectly compliant launch
Imagine every worker meets a compulsory usage target. Outputs are copied into the approved field, then corrected through an unrecorded manual process. Error reporting falls because raising a defect delays the individual. The launch dashboard is green while cost, risk and silence rise. Now remove the target, protect reporting and publish unresolved defects. Usage falls and useful evidence increases. Which state is closer to successful change?
The audit does not reverse accountability. A manager can acknowledge a defective incentive and still address an employee who leaked protected data. A union can raise a valid distribution claim without controlling technical validation. A vendor can repair a serious defect while the buyer corrects a configuration failure. Symmetry means examining contributions on all sides and assigning each condition to the authority capable of changing it.
Part three · Decide and prove closureRespond to the condition, not the emotion
The Resistance Response Menu prevents training from becoming the default. Select among product repair, workflow redesign, evidence, authority, capacity, protection, transition, vendor action, preserved boundaries and termination only after the proposition and rivals are visible.
Product or model repair may change data, reliability, interface, accessibility, explanations or failure handling. Workflow action may change integration, sequence, exceptions, manual paths and recovery. Evidence action may require local validation, monitoring, comparison, explanation or a bounded trial. Authority action may clarify decision rights, liability, override, escalation and who accepts consequence. Capacity action may fund staffing, protected learning, support and checking work.
Employment and protection responses include consultation, bargaining, privacy limits, accessibility, recourse, remedy, job transition and income protection. Incentive redesign removes conflicting targets. Role and identity work can preserve valued expertise, status and progression while changing tasks. Vendor action includes remediation, substitution and exit. A justified refusal, manual route or professional discretion can remain part of the operating design. A failed premise should trigger pause, rollback or termination rather than a campaign for enthusiasm.
| Adjudicated condition | Proportionate response | Owner | Closure proof | Revalidation trigger |
|---|---|---|---|---|
| Capability or workflow creates avoidable burden | Repair model, interface, integration, exceptions or support | Product and workflow | Representative work episodes show lower burden without hidden harm | Release, integration or task change |
| Authority, workload or incentives conflict | Change rights, staffing, targets, override or escalation | Operating executive | People can act within time and are not penalised for control work | Role, volume or reward change |
| Evidence, protection or recourse is insufficient | Validate, restrict, add confidential challenge, appeal and remedy | Outcome, rights and assurance | Relevant evidence and tested recourse cover affected populations | Population, consequence or legal change |
| Professional or rights boundary is justified | Preserve refusal, discretion, manual route or second channel | Professional and service owner | Boundary is accessible, workable and not retaliatory | Evidence, standard or recovery change |
| Use-case premise fails | Pause, roll back, replace or terminate | Portfolio and business owner | Cost and exposure cease; affected work has a supported route | Only new evidence that changes the premise |
| Identity concern with a real escalation defect | Repair the defect and redesign role meaning separately | Operations and workforce | Escalation works and role contribution is explicit | Role or workflow redesign |
| Vendor credibility damaged by unresolved incident | Remedy, independent assurance, substitution or exit | Vendor and service owner | Incident is corrected, parties remedied and recurrence tested | Vendor release or contract renewal |
The Resistance Decision Record makes the choice reviewable. Include the case identifier and source protection; behaviour and object; intended outcome, workflow, population and consequence; classification with uncertainty; proposition and rivals; evidence, limits and dissent; the contribution audit; the refusal test; options, non-action and any reversible experiment; accountable owner and rights; selected response, resources, interim boundary and deadline; communication; and closure, counterfactual plus revalidation.
Closure is not an action completed. Training delivered, a meeting held or a defect ticket closed proves activity. Closure must show that the adjudicated condition changed for the specified population and that consequence did not migrate elsewhere. Where a programme rejects a claim, closure includes the tested rival, reasons, dissent and appeal. Where it accepts a claim, closure includes the repair and its operating evidence. The source should receive a protected explanation unless law or confidentiality prevents it.
The Resistance Signal Panel therefore keeps seven views separate: objects, conditions, voice safety, response, distribution, closure and learning. Do not calculate a resistance, readiness, sentiment or compliance probability score. An average can combine improved workflow with deteriorating voice safety and hide both. The panel asks whether consequential signals are reaching authority, whether affected populations are represented, whether interim protections work, and whether repairs survive revalidation.
Part four · Apply the inquiryTen signals, five legitimate routes
The following cases show why visible conduct is insufficient. Each starts with a signal, states a testable account and routes the condition rather than the person.
1. Customer service
Agents rarely open an assistant because copying context takes longer than resolving ordinary requests. Observation confirms duplicate entry and exception delay across representative episodes. Route: repair integration and the exception path. Training cannot remove a handoff the product created.
2. Clinical oversight
Clinicians refuse nominal oversight when liability is unclear, explanations do not support error detection and workload allows seconds rather than review. The refusal test fails authority, evidence and capacity. Route: restrict use, repair those conditions and preserve professional refusal meanwhile.
3. Warehouse scheduling
Workers oppose optimisation associated with variable hours and intrusive monitoring. The inquiry separates privacy, job quality, income, business need and technical performance. Route: use protected social dialogue and employment process while testing each proposition independently.
4. Forecast rejection
Forecasters reject a useful model after one memorable error. Comparative results show better aggregate decisions inside a bounded envelope, while experts remain stronger for identified shocks. Route: run a bounded override test, preserve expertise and address aversion through shared evidence.
5. Compulsory usage
An adoption target produces copied outputs, hidden manual repair and silence. The programme reads target compliance as success. Route: the contribution audit removes the target, protects reporting and reopens workflow evidence before any further scale decision.
6. Union objection
A union objects because roles, progression and checking load were redesigned without negotiation. Model accuracy cannot answer the distribution or procedural claim. Route: disclose impacts, enter the applicable consultation or bargaining process and protect interim conditions.
7. Field accessibility
Field workers using assistive technology or a minority language cannot reliably operate the interface. Aggregate usage hides the excluded cohort. Route: treat this as accessibility and contextual failure, test with the population and repair before imposing use.
8. Managerial status and defect
Middle managers fear a loss of discretion and correctly identify a broken escalation route. Neither fact cancels the other. Route: repair escalation, define their new judgement role and assess identity concerns without psychologising the operational claim.
9. Vendor credibility
A new release is distrusted after an earlier incident produced no remedy. Current test results may be sound, yet prior accountability remains unresolved. Route: demonstrate closure, provide independent challenge and require contractual remediation or substitution.
10. Small business non-use
A small firm lacks recurring volume, usable data and benefit after correction cost. The implementation is not waiting for motivation. Route: stop the change because the premise fails, preserve a simpler process and avoid turning sunk cost into obligation.
Rival view: inquiry can romanticise obstruction
Yes. If every complaint becomes a veto, loud and secure participants can capture consultation while contractors, quiet professionals and affected customers remain unseen. Historical studies of change response and laboratory studies of algorithm use do not prove the cause of a local signal. The answer is disciplined, time-boxed inquiry: specify the claim and counterevidence, protect minorities and silent populations, distinguish authorised refusal from sabotage or misuse, and preserve the ability to decide after a fair process. Appeal, incident reporting and revalidation remain open. The measure is changed condition and outcome, not whether a critic becomes enthusiastic.
Part four · Use the instrumentAI Resistance Inquiry Workbench
The workbench is a reasoning aid, not an automated adjudicator. Choose a scenario or alter the inputs. Its five outputs are routes, not scores; the accountable owner must review evidence, rights, dissent and local obligations.
Design the inquiry
Repair the AI capability or workflow
The workbench deliberately refuses to rank people. Its result preserves the observed behaviour, object, taxonomy classes, protected route, proposition, rivals, refusal conditions, programme contribution, evidence gaps, affected population, distributional effect, owners, response, closure and revalidation. That trace is more demanding than a resistance score and far more useful to the person who must act.
ConclusionListen hard enough to change the decision
Resistance to AI is informative only after the organisation gives up the fantasy that one visible behaviour has one hidden cause. Refusal, delay, workaround and silence are entry points to inquiry. They are neither guilt nor virtue.
The operating discipline is compact. Describe conduct without motive. Protect the person and the affected population. Name possible objects with the twelve-class taxonomy. Write a falsifiable proposition and credible rivals. Test whether refusal is compensating for missing evidence, authority, rights, recourse or capacity. Audit how the change programme contributed. Route each condition to competent authority. Select among repair, changed operating conditions, evidence and protection, a preserved boundary or termination. Return closure evidence and reopen it after material change.
This discipline can reach uncomfortable findings. A technically sound model may still require negotiation because job quality and income are changing. A sincere professional objection may fail comparative evidence. A celebrated implementation may be held together by invisible manual work. A low-use tool may have no economic reason to exist. None of those findings is available to a programme that defines success as compliance.
The practical test is simple: did the signal alter a condition or stop a failed premise while voice remained safe? If it did, disagreement strengthened the system. If it disappeared only because reporting became costly, the organisation has not overcome resistance. It has blinded itself.
Evidence baseReferences and scope
Governance sources are applied within their stated purposes and jurisdictions. Workplace, change, silence, implementation and algorithm-response research is treated as bounded evidence, not as proof of a local cause.
- National Institute of Standards and Technology. AI RMF Core.
- National Institute of Standards and Technology. Appendix A: descriptions of AI actor tasks.
- National Institute of Standards and Technology. Appendix C: AI risk management and human-AI interaction.
- UK Government. AI Playbook for the UK Government.
- International Labour Organization. Global case studies of social dialogue on AI and algorithmic management.
- OECD. Using AI in the workplace.
- OECD. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market.
- European Union. Regulation (EU) 2024/1689, consolidated text.
- Ford, J. D., Ford, L. W. and D'Amelio, A. Resistance to change: the rest of the story.
- Piderit, S. K. Rethinking resistance and recognising ambivalence.
- Oreg, S., Vakola, M. and Armenakis, A. Change recipients' reactions to organisational change.
- Morrison, E. W. and Milliken, F. J. Organisational silence.
- Detert, J. R. and Edmondson, A. C. Implicit voice theories.
- Edmondson, A. Psychological safety and learning behaviour in work teams.
- Klein, K. J. and Sorra, J. S. The challenge of innovation implementation.
- Damschroder, L. J. and colleagues. The updated Consolidated Framework for Implementation Research.
- May, C. and Finch, T. Implementing, embedding and integrating practices.
- Dietvorst, B. J., Simmons, J. P. and Massey, C. Algorithm aversion after seeing an algorithm err.