Change Management and Responsible Adoption

Chapter 14 named the roles an AI programme needs; this chapter concerns the much larger group of people who must use the systems those roles design, approve and oversee. An architecture that is compliant, observable and well evaluated still fails if case officers route around it, engineers distrust it, or a works council halts it on launch day. The chapter applies diffusion of innovation and Kotter's model of leading change to AI, adds employee consultation, expectation management and culture, and closes with an adoption dashboard as code, extending the treatment of organisational change in Architecture as Code (Chapter 17) to AI (66, Nordqvist, 2025).

Figure 15.1: The adoption curve for enterprise AI, with the chasm between pilot and production. The table below gives the programme response for each adopter segment.

Segment	Stance	Programme response
Innovators	Try what is new	Sandbox and shadow mode
Early adopters	Solve a known problem	Visible sponsors and measured wins
The chasm	Pilot to production	Golden path with guardrails and gates
Early majority	Need peer proof	Consultation and AI literacy
Late majority	Adopt under pressure	Controls in the pipeline and an adoption dashboard
Laggards	Adopt when there is no alternative	No separate programme response

Table for Figure 15.1: Adopter segments and programme response.

Adoption is the binding constraint

Most large organisations that stall with AI do not stall for lack of models, but because a pilot that impressed a steering committee never became part of how work is done: recommendations are accepted at a rate suggesting nobody reads them, or a drafting assistant is used daily by a third of the department and never opened by the rest.

These adoption failures are architectural in two senses. The architecture shapes adoption, because a system that explains itself, fails gracefully and leaves the user in control is adopted differently from one that does not. And adoption data is architectural evidence: the override rate, the share of recommendations that users reject or change, is the clearest signal of whether human oversight is real, and Chapter 17 shows how public authorities publish it in their algorithm registers. Adoption is therefore a property of the system that is designed, measured and governed, not a communications task bolted on at the end.

Diffusion of innovation applied to AI

Rogers (2003) described how innovations spread through a population in a predictable sequence of adopter segments: innovators, early adopters, early majority, late majority and laggards (76, Rogers, 2003). Early adopters solve a problem they already have and become opinion leaders, while the early majority wait for proof from peers. Rogers also identified five attributes that govern the rate of adoption (76, Rogers, 2003), each with a direct architectural translation for AI.

Attribute	Question the adopter asks	Architectural response
Relative advantage	Is this better than how I work now?	Measure task time and quality before and after; publish the numbers
Compatibility	Does it fit my tools and my judgement?	Integrate into systems of record; keep the user as decision-maker
Complexity	Can I understand what it does?	System-level explanation; plain-language model cards
Trialability	Can I try it without consequences?	Sandboxes with masked data; shadow mode
Observability	Can I see others succeeding with it?	Internal showcases; adoption dashboards visible to all

The chasm between pilot and majority

The gap between early adopters and the early majority is where AI programmes most often fall. Early adopters tolerate a model that is right eighty per cent of the time because they understand the trade-off; the early majority experience the same model as a tool that is wrong one time in five, and they have a job to do. Crossing the gap requires that the system change, not merely the message. Guardrails, evaluation gates and golden paths exist precisely so that the early majority receive a more reliable product than the innovators played with. The crossing therefore has an engineering budget and a governance decision attached: the pilot does not graduate until those controls are in place.

The pilot team should therefore not be the roll-out team, because innovators underestimate the cost of learning. Recruit early adopters from each professional community as visible sponsors, and give them the data to make an honest case.

Illustrative example. Nordvik Region formalised the early adopters for its internal staff assistant (NR-ALG-2026-007), rolled out to some 2,400 staff, as a champions network with a one-page charter in the programme repository. It provides one champion per 40 to 60 staff from the community they serve, excludes members of the pilot team in the first term, and grants four hours a month of protected time recorded as a contribution to the unit, never as a sales target. Champions route every error report to the feedback tracker within two working days. The charter also lists what champions do not do: approve use cases, interpret policy, or report on whether named colleagues use the assistant, so that the role cannot drift into surveillance. Its health measures reward reaching the early majority rather than counting enthusiasts.

Kotter's model and its AI-specific pitfalls

Kotter (2012) set out eight steps for leading large-scale change: establish a sense of urgency, create a guiding coalition, develop a vision and strategy, communicate the vision, empower broad-based action, generate short-term wins, consolidate gains, and anchor new approaches in the culture (75, Kotter, 2012). The model remains useful, but five of its steps fail in characteristic ways when the change is AI.

Urgency without hype. With AI, urgency arrives unbidden, manufactured by vendors, boards and the press, and the change leader must often moderate it. False urgency produces pilots chosen for visibility rather than value, and a backlash when they disappoint. Ground urgency in the organisation's own data, such as a growing backlog or lengthening case-handling time, because only urgency that survives scrutiny can carry a multi-year programme.

A coalition that includes the people who can stop you. The guiding coalition must include the functions with authority to halt the programme: the data protection officer, the CISO, legal counsel and, in most European organisations, the employee representatives. A coalition of enthusiasts reaches the deployment gate and discovers that the people who were not in the room have a veto; involving them early is slower at the start and much faster at the end.

A vision that names what AI will not do. "AI everywhere" is an absence of vision. Say which decisions remain with people, which tasks will be automated, and what the organisation will refuse to do whatever the technology permits. Chapter 4 proposed declared intent as a principle, and here it is also a change instrument, because people adopt a change whose edges they can see.

Short-term wins that are real. Demonstrations are easy and wins are hard. A demonstration on curated examples is not a win; a measured reduction in handling time on live cases, with error rates no worse than the baseline, is. Define the win before the pilot starts, record it as acceptance criteria in the use-case register of Chapter 16, and report against it.

Anchoring without governance theatre. Anchoring fails when governance becomes ceremony: a committee that approves everything, a register complete on launch day and never again. It succeeds when the controls are in the pipeline, as Chapter 9 describes, and the people doing the work can see that governance protects them rather than slowing them.

Employee consultation in European labour contexts

European organisations introduce AI into workplaces governed by information and consultation rights that have no equivalent in many other regions. These rights are part of the design process rather than an obstacle, and the AI Act has written one of them into the regulation itself. Article 26(7) of Regulation (EU) 2024/1689, the AI Act, requires deployers that are employers, before putting a high-risk AI system into service or using it at the workplace, to inform workers' representatives and the affected workers, in accordance with Union and national rules on informing workers (1, European Union, 2024). Because Annex III lists recruitment, task allocation and the monitoring and evaluation of performance among the high-risk categories, a wide range of workplace AI falls within this duty.

National law adds consultation and, in several countries, co-determination. German works councils hold co-determination rights over the introduction of technical systems capable of monitoring employee behaviour or performance under section 87(1) no. 6 of the Works Constitution Act (130, Germany, 2001), which in practice covers most workplace analytics. Swedish employers must negotiate with the trade unions before deciding on significant changes to operations or working conditions, and Dutch works councils hold a right of consent over regulations on the processing of employees' personal data and over systems that monitor attendance, behaviour or performance. The details differ, but in each case the architecture is part of what is consulted on, and a system whose design is fixed when consultation begins has lost the chance to be improved by it.

Consultation as a design input

Treat consultation as a design review with a different reviewer. Works councils and unions ask questions architects should want answered: what employee data the system uses, who sees the outputs, whether they can be used for performance management, how long logs are kept and what happens when the system is wrong. Record the answers as commitments in the system's AI Decision Record and enforce them in the permission model and retention policy, because a commitment in a presentation is a promise, whereas one in a policy-as-code rule is a control. The Article 4 literacy obligation of Chapter 14 is the positive counterpart, and the consultation is where that training is shaped (1, European Union, 2024).

Consultation across a group's jurisdictions

A multinational group meets these rights entity by entity, and who consults whom follows the employment relationship, not the platform. A claims assistant built by group IT in Hamburg and used in six markets is consulted on by each employing entity under its own national rules, with the platform team supplying the material. Article 88 of the GDPR allows member states to lay down more specific rules for employees' personal data, including through collective agreements (2, European Union, 2016), so the commitments can legitimately differ between entities for an identical system.

Three consequences follow. A shared service centre is consulted where its staff are employed, so a commitment made in Germany about logging does not automatically protect claims handlers in Poland. Where a matter concerns several German companies of a group and cannot be settled within each, the Works Constitution Act gives competence to the group works council (130, Germany, 2001), which turns a group-wide tool into one negotiation in Germany while other countries still consult separately. And entities outside the AI Act's reach, or in Norway's case outside it until incorporation into the EEA Agreement (187, EFTA, n.d.), are not exempt from consultation, because national labour law and group policy still apply. The rollout gate per entity of Chapter 12 turns a missing consultation into a blocked deployment.

Worked example: consultation commitments as controls

Illustrative example. Nordvik Region's staff assistant is not a high-risk AI system, so Article 26(7) does not strictly apply; national co-determination law does, and the Region chose to run the same information procedure it uses for high-risk systems. The commitments agreed with the employee representatives were recorded as a file naming, for each commitment, the control that enforces it and the test that proves it.

# consultation/NR-ALG-2026-007-commitments.yaml
system_ref: NR-ALG-2026-007
system_name: Staff assistant (drafting and guidance search)
consultation:
  counterparts: [employee representatives, local trade union branches]
  basis: national co-determination law; AI Act Art. 26(7) procedure applied voluntarily
  agreement_signed: 2026-04-22
  decision_record: adr/AIDR-0211.md
commitments:
  - id: C1
    commitment: Usage data is never used to assess individual performance
    control: policy/staff-assistant/usage-access.rego
    test: tests/policy/test_no_manager_usage_access.rego
  - id: C2
    commitment: Prompt and response logs are deleted after 90 days
    control: infra/logging/retention.tf
    test: tests/infra/test_log_retention.py
  - id: C3
    commitment: Adoption figures are reported only for groups of ten or more
    control: dashboards/adoption/staff-assistant.yaml
    test: tests/dashboards/test_minimum_group_size.py
  - id: C4
    commitment: Use of the assistant stays voluntary for the first twelve months
    control: hr/role-descriptions/
    test: manual, verified at the joint annual review
  - id: C5
    commitment: Staff complete AI literacy training before access is granted
    control: iam/groups/staff-assistant-users.yaml
    test: tests/iam/test_training_precondition.py
review:
  joint_review: annually, first on 2027-04-22
  reopen_triggers: [new data source, new user group, change of model provider]

The file feeds the consultation_commitments_met metric in the adoption dashboard below. Commitment C4 has no automated test, and the file says so openly; the reopen_triggers turn consultation from a one-off event into a condition the pipeline can detect.

Communication and expectation management

The hardest communication task in an AI programme is explaining a probabilistic system to people accustomed to deterministic ones. A model that is usually right and sometimes confidently wrong fits neither of the categories staff know, right or broken.

Three disciplines help. The first is precision of language: do not say the system understands, decides or knows; say it classifies, recommends or generates, with what measured accuracy on what population, because anthropomorphic language inflates expectations and makes failures feel like betrayals. The second is publishing error rates alongside success rates, so that known failure modes cost less trust when they occur. The third is a visible route for reporting problems and a visible record of what happened to the reports, since people stop reporting when reporting changes nothing. Shneiderman (2022) framed the goal of human-centred AI as systems that are reliable, safe and trustworthy, achieved through high levels of both human control and automation rather than a trade between them (74, Shneiderman, 2022). That framing rejects the false choice staff fear, that either the machine decides or they do; the honest message is that the system does more of the routine work and the person retains the authority, and the architecture must make that message true.

Illustrative example. The announcement Nordvik Region sent when the staff assistant reached general availability on 1 June 2026 applied all three disciplines. It said the tool drafts text and finds passages in guidance, that the member of staff decides what to use, and that it makes no decisions on any case; it reported that on 1,200 real guidance questions the assistant linked to the correct section in 91 per cent of cases and to a wrong or outdated one in 9 per cent; it restated the consultation commitments in plain words; and it showed that the 64 pilot reports had led to 41 fixes and 15 corrected guidance documents. The pilot figures on reports matter most, because they are the evidence that reporting changes something.

Across the usage modes

The six usage modes of Chapter 2 change different things for staff, so a single adoption plan for "AI" fits none of them.

Usage mode	What changes for staff	Communication and support it needs
Mode 0 Ambient AI	AI summaries and notes simply appear in existing tools	Guidance on checking AI output; notice when a supplier switches on a feature
Mode 1 Conversational assistants	A new general tool; individual habits change	Rules on what may be entered, verification habits, champions
Mode 2 Embedded workplace assistants	Familiar tools surface anything the user can open	Advance notice, a permissions clean-up as part of the launch
Mode 3 Agentic work tools	Specialists delegate tasks and review results	Sandbox onboarding, team norms for reviewing agent work
Mode 4 Built AI applications	A defined process changes; decision paths may shift	Consultation on design, published error rates, oversight training
Mode 5 Autonomous production agents	Work moves from doing to supervising	Early consultation on role change, retraining paths

Mode 2 fails most quietly, when staff discover the assistant can retrieve a document they were never meant to see. Modes 4 and 5 carry the heaviest consultation duties, because high-risk uses under Annex III of the AI Act and changes to working conditions concentrate there.

Adoption nobody chose

Ambient and embedded features (Mode 0 and many Mode 2 functions) invert the adoption problem. Nobody has to persuade anyone to use a meeting summary that appears after a supplier update; the questions become whether the organisation knows the feature is live, has decided to permit it, and has consulted the representatives as it would for a built tool with the same effect.

The response is a standing process rather than a campaign. A supplier's announcement of a new AI feature opens a review in the portfolio register, as the supplier-feature entry of Chapter 16, with an owner and a deadline. Features enabled by default are switched off at tenant level, where the product allows it, until the review asks the questions any consultation would ask. A feature that attributes contributions in meetings or summarises an individual's activity is a technical system capable of monitoring employees under German co-determination law, so it goes to the works council before it is enabled for German staff. The useful measure is then not the share of staff using the feature but the share of live AI features that have passed review.

Psychological safety and generative culture

Chapter 5 introduced Westrum's typology and the evidence that a generative culture predicts delivery performance. For adoption the stakes are higher, because a model's failures are statistical and often invisible to its builders, and the only reliable detector of a model that has started to discriminate, hallucinate or drift is a user who notices and says so. A generative culture for AI therefore treats overrides as evaluation data rather than insubordination, acknowledges error reports within a defined time, never measures anyone on the acceptance rate of recommendations, because that metric rewards not looking, and thanks those who raise fundamental objections for finding the problem before the regulator did.

Psychological safety also has an architectural expression. A system that logs every override with the user's identity and shows the log to managers discourages overrides; one that logs overrides for model evaluation, aggregates them before reporting and separates evaluation from performance use encourages them. The same data, governed differently, produces opposite cultures.

Measuring adoption

Adoption metrics serve the change lead, who needs to know where the programme stands on the curve; the system owner, who needs to know whether human oversight is functioning; and the steering committee of Chapter 16, which needs evidence rather than anecdote. Define them once, in a dashboard definition under version control, and let the pipeline that deploys the system also deploy the dashboard.

# dashboards/adoption/benefit-triage-adoption.yaml
dashboard: adoption-benefit-eligibility-triage
system_ref: ALG-2026-014
refresh: daily
audiences: [change_lead, system_owner, steering_committee]
metrics:
  - id: active_users_ratio
    formula: active_users / eligible_users
    window: 28d
    target: ">= 0.70 by 2026-12-31"
  - id: override_rate
    formula: overridden / recommendations_shown
    window: 7d
    healthy_range: [0.03, 0.15]
    note: Below range suggests rubber-stamping; above suggests low quality
  - id: error_reports_open
    source: feedback_tracker
    alert_above: 20
  - id: literacy_training_completion
    formula: trained_users / eligible_users
    regulatory_ref: AI Act Article 4
  - id: consultation_commitments_met
    source: policy_test_results
    target: "== 1.0"
privacy:
  minimum_group_size: 10
  individual_level_access: none
  retention_days: 400

The healthy_range on the override rate encodes the insight that both extremes are failures: a rate near zero means oversight has become ceremony, and a high rate means the model is not good enough to deploy. The privacy block encodes the cultural commitment of the previous section, reporting only for groups of ten or more and with bounded retention; without it the dashboard becomes a surveillance tool, and the works council will be right to say so. Late in a programme, the open error reports and consultation commitments show whether the change has been anchored or merely declared.

Nordvik Region and Hansa Mutual

Nordvik Region's champions network, consultation commitments and staff announcement all concern the internal staff assistant (AIDR-0211), rolled out to about 2,400 and then about 3,200 staff. The Region consults under one national co-determination regime and chose to run its high-risk information procedure for a system that is not high-risk, because trust from unions and staff is the condition for adoption in a public body.

Hansa Mutual runs the same change in several labour regimes at once. Rolling the claims assistant out from the pilot unit of about 320 claims handlers to every market means a separate consultation in each country: co-determination with the German works council where the tool can monitor performance (130, Germany, 2001), works-council rights in the Netherlands, union consultation in Sweden and the arrangements of the other entities. The commitments agreed in each country become controls with an applies_in scope, so that a commitment made to Dutch employees does not silently lapse when a group release changes logging. Nordvik's adoption depends on trust in one workforce; Hansa's depends on agreements that differ by country but run on one platform.

Summary

Adoption is the binding constraint on enterprise AI, and it is an architectural concern because the system's design shapes how it is adopted and because adoption data is governance evidence. Rogers' attributes of innovation translate into design choices such as sandboxes, visible dashboards and integration that leaves the person in authority, and the chasm between early adopters and the early majority is crossed by changing the system, not only the message.

Kotter's steps hold for AI, but each has its own pitfall, from hype-driven urgency to demonstrations mistaken for wins. European labour law and Article 26(7) of the AI Act make employee consultation a legal step in deployment, best treated as a design review whose outcomes are enforced as controls. Precise language, published error rates and a visible feedback loop manage expectations about probabilistic systems, and a generative culture lets the organisation hear about failures while they are still cheap.

The adoption dashboard produces the evidence that the next chapter consumes. Chapter 16 therefore extends Management as Code to AI portfolios, showing how use-case registers, value hypotheses, funding models and steering committees can run on evidence generated by pipelines rather than on slides prepared the night before.

Sources

1. European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. (1, European Union, 2024)
2. Nordqvist, G. (n.d.). Architecture as Code. Companion volume, aac.geon.se. (66, Nordqvist, 2025)
3. Shneiderman, B. (2022). Human-Centered AI. Oxford University Press. (74, Shneiderman, 2022)
4. Kotter, J. P. (2012). Leading Change. Harvard Business Review Press. (75, Kotter, 2012)
5. Rogers, E. M. (2003). Diffusion of Innovations. 5th edition. Free Press. (76, Rogers, 2003)
6. Federal Republic of Germany (2001). Betriebsverfassungsgesetz (BetrVG) (Works Constitution Act), as last amended on 19 July 2024. Federal Ministry of Justice, Gesetze im Internet. (130, Germany, 2001)
7. European Parliament and Council of the European Union (2016). Regulation (EU) 2016/679 on the protection of natural persons with regard to the processing of personal data (General Data Protection Regulation). Official Journal of the European Union. (2, European Union, 2016)