From AI Experimentation to Enterprise Advantage

Artificial intelligence has moved from an emerging technology into a mainstream enterprise capability. The strategic question for leadership is no longer whether to adopt AI—it is how to convert widespread access and experimentation into measurable, responsible and sustainable enterprise value.

Audience: board, executive sponsors, CoE leaders and business-unit owners
Document type: thought-leadership and operating-model paper
Core recommendation: maximise responsible human-and-AI performance—not AI usage volume

---

Executive summary​

Stanford’s 2026 AI Index reports that 88% of surveyed organisations used AI in at least one business function during 2025, while 70% used generative AI in at least one function. However, the deployment of more autonomous AI agents remained at an early stage across most business functions.

The challenge for leadership is therefore no longer whether the firm should adopt AI. The strategic question is:

How can the firm convert widespread AI access and experimentation into measurable, responsible and sustainable enterprise value?

That distinction matters. Deloitte’s 2026 enterprise research reports that employee access to sanctioned AI tools rose significantly during 2025, yet access did not automatically translate into daily workflow integration or scaled production deployment.

The evidence indicates that AI can create meaningful productivity and quality improvements when it is applied to appropriate tasks. A large field study of customer-support employees found that AI assistance increased productivity by approximately 14%, with gains of around 35% among less experienced and lower-performing workers.

A field experiment involving management consultants found that AI improved speed, output quality and task completion when used within the technology’s capability frontier. However, when participants applied AI to a task outside that frontier, their performance deteriorated. This demonstrates that AI value depends not only on adoption, but also on task selection, human judgement and appropriate validation.

The firm should therefore avoid defining success as the number of licences purchased, employees logged in, prompts submitted or pilots launched. These measures may indicate activity, but they do not demonstrate value.

A successful enterprise AI strategy should instead achieve five outcomes:

1. Economic value: increased capacity, productivity, revenue, margin and speed.
2. Client and customer value: improved quality, personalisation, responsiveness and experience.
3. People value: stronger skills, more meaningful work and enhanced employee capability.
4. Risk value: better control, traceability, consistency and early identification of issues.
5. Strategic value: new services, differentiated propositions and faster organisational learning.

This paper recommends that the firm establish a federated AI operating model built around:

• Clear board and executive accountability.
• An enterprise AI Centre of Excellence.
• Distributed ownership within business units.
• A structured AI champion network.
• Secure and reusable technology platforms.
• Role-specific learning and workflow redesign.
• Proportionate governance and assurance.
• Value-based incentives.
• Fair and graduated accountability.
• A portfolio approach to investment and benefits realisation.

The central recommendation is:

The firm should not attempt to maximise AI usage. It should maximise responsible human-and-AI performance.

---

1. The strategic case for action​

1.1 AI adoption is accelerating, but value remains uneven​

The rapid increase in AI usage creates an impression that enterprise transformation is already well advanced. In practice, many organisations remain between experimentation and meaningful scale.

Stanford’s 2026 AI Index shows that organisational adoption has become widespread. Deloitte’s 2026 research similarly indicates rapidly expanding workforce access and high expectations for moving more AI initiatives into production. However, evidence from enterprise surveys continues to show a gap between providing access and achieving transformational outcomes.

This gap arises because AI is frequently approached as:

• A technology-purchasing programme.
• A collection of isolated pilots.
• A personal productivity tool.
• A prompt-engineering training initiative.
• A cost-reduction exercise.
• A responsibility delegated entirely to IT.

Each of these may contribute to adoption, but none is sufficient on its own.

AI is a general-purpose capability that affects operating processes, decisions, products, employee roles, client interactions and organisational knowledge. Its adoption therefore requires coordinated changes across business strategy, technology, data, people, governance and performance management.

1.2 AI is not uniformly effective​

The capabilities of generative and agentic AI are uneven. A model can perform extremely well on one task and poorly on another task that appears similar.

This is the “jagged technological frontier” identified in experimental research. Inside the frontier, AI can materially improve performance. Outside it, AI can make employees more confidently wrong.

For firms, this means that the correct unit of AI adoption is rarely the entire job.

The more useful unit is:

• A task.
• A decision.
• A workflow stage.
• A customer interaction.
• A knowledge process.
• A control activity.

For example, AI may be suitable for:

• Producing an initial document structure.
• Summarising a meeting.
• Identifying relevant policies.
• Comparing standard contract provisions.
• Drafting test cases.
• Extracting data from documents.
• Preparing a research brief.

The same AI may be unsuitable for independently:

• Making a high-impact employment decision.
• Approving a regulated financial treatment.
• Providing final legal advice.
• Determining an audit conclusion.
• Making an irreversible customer decision.
• Issuing an externally relied-upon professional opinion.

The firm must therefore design adoption around task suitability, decision consequence and required human judgement.

1.3 The economic opportunity is significant​

big 4 firm’s 2025 Global AI Jobs Barometer found that industries more exposed to AI were experiencing stronger productivity growth and that workers with AI-related skills attracted a substantial wage premium. These findings are observational rather than proof that AI alone caused the difference, but they indicate that the market increasingly values the ability to combine domain expertise with AI capability.

At an enterprise level, AI can create economic value through:

Capacity creation​

Employees can complete appropriate activities faster, allowing them to manage greater volumes or redirect time toward higher-value work.

Quality improvement​

AI can increase consistency, identify missing information, support structured review and make specialist knowledge more accessible.

Revenue growth​

The firm can create new products, improve conversion, personalise services, accelerate proposals and provide new AI-enabled client offerings.

Margin improvement​

AI can reduce rework, administrative effort, delivery time and duplicated analysis.

Risk reduction​

AI can support control monitoring, policy checks, document review, anomaly detection and traceable decision support.

Knowledge leverage​

The firm can make institutional knowledge easier to find, apply and reuse across teams.

The scale of these benefits will vary significantly by use case. The board should therefore require quantified business cases rather than relying on general claims about AI productivity.

---

2. The principal insight: AI adoption is an organisational transformation​

2.1 Access is not adoption​

A firm can deploy thousands of licences without materially changing performance.

A useful adoption chain is:

Access → awareness → capability → repeated use → workflow integration → measurable outcomes → enterprise value

Each stage requires a different intervention.

Stage	Leadership question
Access	Do employees have secure, reliable and approved tools?
Awareness	Do employees understand what the tools can and cannot do?
Capability	Can employees use and validate AI appropriately?
Repeated use	Is AI solving recurring, meaningful problems?
Workflow integration	Has the end-to-end process changed?
Outcomes	Have quality, speed, cost or experience improved?
Enterprise value	Has the benefit affected revenue, margin, risk or strategic capability?

Login and prompt statistics can help identify adoption patterns, but they should not be treated as evidence of business value.

2.2 Individual time savings are not automatically enterprise benefits​

Suppose an employee saves two hours preparing a document. The firm does not necessarily receive two hours of economic value.

The benefit may be lost when:

• The employee performs additional low-value work.
• A manager repeats the analysis.
• The output requires substantial correction.
• The next workflow stage remains constrained.
• The employee does not know how to use the saved capacity.
• The firm cannot reduce cost, increase volume or improve quality.
• The time-saving claim is not validated.

Every priority AI use case therefore needs a benefit-conversion mechanism.

That mechanism should answer:

1. What time, cost or quality improvement is expected?
2. How will it be measured?
3. Who owns the benefit?
4. What will happen to the capacity created?
5. Can the firm increase throughput, improve service or reduce cost?
6. What additional review or control effort is required?
7. Is the benefit repeatable at scale?

Without these decisions, AI may make individuals faster without making the firm materially more productive.

2.3 Workflow redesign is the critical value lever​

Weak AI adoption adds a chatbot to an existing process.

Strong AI adoption redesigns the process.

Consider a proposal-development workflow.

Weak implementation​

• Employees receive access to an AI assistant.
• They are encouraged to draft proposals with it.
• Usage is measured.
• No common content library is established.
• No workflow ownership changes.
• Claims are reviewed manually at the end.
• No reliable benefits baseline exists.

Strong implementation​

1. Approved account, market and sector information is retrieved.
2. AI structures the opportunity and identifies missing information.
3. The account team validates client priorities.
4. Approved credentials and propositions are matched to the opportunity.
5. AI creates a controlled first draft.
6. Commercial, legal and risk checks are embedded.
7. Subject-matter experts validate material claims.
8. Reusable content is captured.
9. Proposal effort, quality, margin and success rate are measured.
10. Lessons are used to improve the process.

The second model creates an organisational capability rather than merely improving document drafting.

---

3. Evidence on what works​

3.1 Select tasks with a measurable performance model​

Early use cases should normally have:

• A clear and recurring problem.
• Sufficient task volume.
• Accessible information.
• A recognisable definition of quality.
• A measurable baseline.
• Reviewable outputs.
• Manageable consequences if the AI is wrong.
• A committed process owner.
• A realistic route to scale.

Suitable early opportunities often include:

• Knowledge search.
• Document summarisation.
• Customer-service assistance.
• Meeting preparation.
• Proposal support.
• Software-development assistance.
• Document classification.
• Research synthesis.
• Internal policy guidance.
• Administrative workflow support.

This does not mean such use cases are automatically low-risk. The actual risk depends on the data, users, decisions and context.

3.2 Use role-specific and hands-on learning​

The OECD’s 2025 research on AI adoption found that firms valued practical, industry-specific training more than generic AI instruction. Interviewees particularly emphasised hands-on training using real-world projects, tools and relevant datasets.

The firm should establish a four-level learning model.

Level 1: Enterprise AI literacy​

Required for the broader workforce:

• Core AI concepts.
• Capabilities and limitations.
• Approved tools.
• Data-handling rules.
• Confidentiality.
• Intellectual property.
• Hallucination and error risks.
• Human accountability.
• Escalation and incident reporting.

Level 2: Role-based application​

Tailored learning for:

• Finance.
• Audit and assurance.
• Tax.
• Legal.
• Sales.
• Consulting.
• Software engineering.
• Human resources.
• Procurement.
• Customer service.
• Operations.

Employees should learn how AI changes actual tasks in their role.

Level 3: Workflow transformation​

Managers, process owners and champions should learn:

• Process mapping.
• Use-case selection.
• Benefits assessment.
• Human-control design.
• Evaluation.
• Adoption planning.
• Change leadership.
• Operational monitoring.

Level 4: Specialist capability​

Technical, product and risk practitioners require deeper skills in:

• Model selection.
• Retrieval-augmented generation.
• AI agents.
• Evaluation design.
• Data engineering.
• Security testing.
• Observability.
• Responsible AI.
• Model risk.
• AI FinOps.
• Production operations.

3.3 Consult employees and involve them in redesign​

OECD workplace research found that training and employee consultation were associated with better worker outcomes and stronger trust in workplace AI.

Employee involvement is particularly important because frontline teams understand:

• Informal process steps.
• Common exceptions.
• Client expectations.
• Workarounds.
• Quality risks.
• Data limitations.
• Control failures.
• Sources of unnecessary effort.

Consultation should not be treated as an unrestricted veto over change. It should be used to improve implementation quality, identify legitimate risks and build informed commitment.

3.4 Provide clear leadership direction​

Leaders should communicate:

• Why the firm is investing in AI.
• Which outcomes matter.
• How employees will benefit.
• What responsible use looks like.
• What will happen to productivity gains.
• Where human judgement remains essential.
• Which practices are prohibited.
• How employees can raise concerns.
• How roles and careers may evolve.

The adoption narrative should avoid two extremes:

Excessive optimism: “AI will transform everything immediately.”

Fear-based transformation: “Use AI or become irrelevant.”

A credible message is:

AI will alter how work is performed. The firm will invest in tools, skills, workflow redesign and responsible governance. Employees will be expected to develop relevant capability, while leaders remain accountable for providing the conditions for successful adoption.

3.5 Create rapid learning loops​

Every scaled AI capability should generate evidence about:

• User behaviour.
• Output quality.
• Errors.
• Exceptions.
• Model performance.
• Human overrides.
• Customer outcomes.
• Cost.
• Security events.
• Adoption barriers.
• Benefits achieved.

The operating model should enable the organisation to:

Test → measure → learn → improve → scale or stop

A failed experiment that produces reliable learning may be more valuable than an apparently successful pilot whose results cannot be reproduced.

3.6 Maintain psychological safety​

Employees must be able to:

• Report AI mistakes.
• Admit uncertainty.
• Ask for help.
• Challenge unsafe outputs.
• Question unrealistic benefits.
• Escalate data concerns.
• Recommend stopping a deployment.

Research on psychological safety found a strong relationship between psychological safety and learning behaviour in teams. This matters for AI because responsible adoption requires experimentation, error reporting and continuous adjustment.

A firm that punishes people for reporting AI problems will not eliminate the problems. It will reduce management’s visibility of them.

---

4. What does not work​

4.1 Technology-first adoption​

Starting with a model or vendor and searching for problems to justify it often produces weak use cases.

The correct sequence is:

Business outcome → workflow problem → user need → data requirement → risk assessment → technology choice

4.2 Disconnected pilot proliferation​

Pilots frequently fail to scale because they lack:

• Business ownership.
• Production funding.
• Data access.
• Integration plans.
• Security approval.
• Support models.
• User adoption.
• Benefits evidence.
• Operational accountability.

The firm should establish explicit decision gates:

1. Explore.
2. Prove desirability.
3. Prove feasibility.
4. Prove responsible performance.
5. Prove economic value.
6. Scale.
7. Operate.
8. Retire when appropriate.

4.3 Generic training​

A single introductory course may increase awareness but is unlikely to alter daily work.

Training without real workflows creates employees who understand AI terminology but cannot apply AI safely or productively.

4.4 Measuring visible activity rather than impact​

Measures such as prompts, users and licences can be useful leading indicators. They become harmful when converted into performance targets without context.

Employees may then:

• Submit unnecessary prompts.
• Use AI for unsuitable tasks.
• Divide one request into several interactions.
• Exaggerate time savings.
• Conceal quality failures.
• Avoid complex but valuable work.

The firm should not reward metric production. It should reward business and professional outcomes.

4.5 Forcing AI into every task​

A universal AI-first requirement may increase risk and inefficiency.

The research on the jagged capability frontier indicates that AI can improve some activities while worsening others.

The policy should therefore distinguish between:

• AI-required workflows: evidence supports AI as the approved process.
• AI-recommended workflows: AI is generally beneficial but judgement remains necessary.
• AI-optional workflows: individuals may decide based on context.
• AI-restricted workflows: use requires specialist approval or controls.
• AI-prohibited workflows: the risk is unacceptable.

4.6 Using fear and punishment as the main adoption mechanism​

Fear may create short-term compliance, but it is unlikely to create high-quality learning.

Employees who fear consequences may:

• Create artificial usage.
• Hide mistakes.
• Use unapproved tools privately.
• Avoid raising concerns.
• Comply visibly while resisting informally.
• Accept unsafe outputs rather than challenge management.

Accountability remains necessary, but it should follow enablement and should focus on responsible performance rather than arbitrary usage.

---

5. Recommended enterprise AI strategy​

5.1 Strategic ambition​

A suitable ambition statement could be:

The firm will use trusted AI to increase client and customer value, improve the quality and economics of delivery, strengthen professional judgement and create differentiated AI-enabled services. AI will augment human capability while accountability remains clearly assigned to people.

The board should agree a limited number of strategic outcomes, such as:

• Increase capacity in knowledge-intensive workflows.
• Improve speed and consistency of client delivery.
• Reduce administrative burden.
• Strengthen risk identification and control.
• Create new AI-enabled propositions.
• Build an AI-capable workforce.
• Improve reuse of institutional knowledge.

5.2 Portfolio structure​

The firm should manage AI investment across three horizons.

Horizon 1: Personal and team productivity​

Examples:

• Drafting.
• Summarisation.
• Search.
• Meeting preparation.
• Coding assistance.
• Analysis support.

Primary objective: capability, familiarity and immediate efficiency.

Horizon 2: End-to-end workflow transformation​

Examples:

• Proposal development.
• Contract review.
• Customer-service resolution.
• Financial close.
• Client onboarding.
• Software delivery.
• Regulatory reporting.
• Knowledge management.

Primary objective: measurable operational and economic value.

Horizon 3: AI-enabled services and business-model innovation​

Examples:

• Client-facing AI platforms.
• AI-enabled managed services.
• Intelligent monitoring.
• Decision-support products.
• Agentic workflow services.
• AI assurance.
• AI security and governance services.

Primary objective: growth and strategic differentiation.

A balanced portfolio prevents the firm from focusing only on quick productivity wins or pursuing ambitious innovation without near-term evidence.

5.3 Use-case prioritisation framework​

Each use case should be assessed across four dimensions.

Value​

• Revenue potential.
• Cost reduction.
• Time saving.
• Margin improvement.
• Customer impact.
• Quality improvement.
• Risk reduction.
• Strategic differentiation.

Feasibility​

• Data readiness.
• Process stability.
• Technical complexity.
• Integration effort.
• Model capability.
• User readiness.
• Operating cost.
• Vendor dependency.

Risk​

• Decision impact.
• Data sensitivity.
• Privacy.
• Security.
• Bias.
• Explainability.
• Regulatory exposure.
• Intellectual property.
• Professional accountability.
• Reputational impact.

Adoption readiness​

• Executive sponsor.
• Process owner.
• User need.
• Baseline.
• Change capacity.
• Training plan.
• Champion support.
• Benefits owner.
• Route to production.

A use case should not receive significant investment unless it has both a business owner and a benefit owner.

---

6. Enterprise AI operating model​

6.1 Recommended model: federated control with distributed delivery​

The most effective design for a large firm is generally a hub-and-spoke model.

The central function should provide:

• Strategy.
• Platforms.
• Standards.
• Governance.
• Reusable capabilities.
• Specialist expertise.
• Portfolio visibility.

Business units should own:

• Problems.
• Workflows.
• Adoption.
• Outcomes.
• Operational accountability.
• Benefit realisation.

The objective is to avoid:

Excessive centralisation​

This creates:

• Long approval queues.
• Weak local ownership.
• A detached technical function.
• Slow experimentation.
• Limited domain relevance.

Excessive decentralisation​

This creates:

• Duplicated investment.
• Unapproved tools.
• Inconsistent controls.
• Fragmented data.
• Weak reuse.
• Unmanaged vendor and security risk.

6.2 Governance structure​

Board​

The board should oversee:

• Strategic alignment.
• Material investment.
• Risk appetite.
• High-impact use cases.
• Workforce implications.
• Regulatory exposure.
• Enterprise value.
• Management accountability.

Executive AI Steering Committee​

Recommended membership:

• Chief executive or executive sponsor.
• Chief technology or information officer.
• Chief data and AI leader.
• Business-unit leaders.
• Chief risk officer.
• Legal and privacy leaders.
• Security leader.
• Human resources leader.
• Finance leader.
• AI CoE leader.

Responsibilities:

• Approve strategy.
• Prioritise investment.
• Resolve cross-functional barriers.
• Review benefits.
• Review material risks.
• Stop underperforming initiatives.
• Sponsor enterprise adoption.

AI Centre of Excellence​

The CoE should own shared capability and enablement.

Business AI councils​

Each business area should maintain a local portfolio, adoption plan and benefits view.

Product and use-case squads​

Cross-functional teams should build and operate specific AI capabilities.

Independent assurance​

Risk, compliance, legal, security, privacy and internal audit should provide appropriate challenge and independent assurance.

---

7. The AI Centre of Excellence​

7.1 Purpose​

The CoE should exist to make responsible AI delivery:

• Faster.
• Safer.
• More consistent.
• More reusable.
• More measurable.
• Less expensive to scale.

It should not become an isolated research team or a central approval bureaucracy.

7.2 Core responsibilities​

Strategy and portfolio​

• Maintain enterprise AI strategy.
• Operate use-case intake.
• Assess opportunities.
• Maintain the portfolio.
• Support business cases.
• Track enterprise outcomes.
• Recommend investment decisions.

Platforms and architecture​

• Provide approved AI environments.
• Maintain model gateways.
• Establish integration standards.
• Support identity and access.
• Enable observability.
• Provide reusable components.
• Manage model and vendor options.

Responsible AI and governance​

• Maintain AI policies.
• Define risk classification.
• Establish assessment requirements.
• Define human oversight.
• Support impact assessment.
• Maintain evidence standards.
• Coordinate incident management.

NIST’s AI Risk Management Framework organises risk activity through the functions Govern, Map, Measure and Manage. ISO/IEC 42001 provides a management-system approach for establishing, maintaining and continually improving organisational AI governance. These frameworks can provide complementary foundations for the CoE.

Evaluation and assurance​

• Define quality metrics.
• Test reliability.
• Evaluate hallucination and groundedness.
• Test bias and harmful behaviour.
• Evaluate security.
• Monitor drift.
• Establish production acceptance criteria.

Adoption and learning​

• Develop role-based curricula.
• Manage the champion network.
• Run workflow laboratories.
• Provide office hours.
• Publish patterns and guidance.
• Support managers.

AI economics​

• Track model and platform cost.
• Measure benefits.
• Manage unit economics.
• Compare build, buy and partner options.
• Optimise model selection.
• Reduce duplicated investment.

Knowledge and reuse​

• Maintain a use-case library.
• Publish approved patterns.
• Store evaluation templates.
• Share reference architectures.
• Capture lessons.
• Maintain prompt and workflow assets where appropriate.

7.3 CoE service catalogue​

The CoE should publish clear services, owners and expected turnaround times.

Service	Typical output
Use-case discovery	Prioritised opportunity statement
Value assessment	Benefits hypothesis and baseline
Risk triage	Risk category and control route
Architecture support	Approved solution design
Data assessment	Data readiness and control requirements
Evaluation design	Quality and safety test plan
Production review	Go-live recommendation
Adoption support	Change, learning and champion plan
Benefits tracking	Value dashboard
Incident support	Triage, containment and remediation
Vendor assessment	Technical, commercial and risk view
Cost optimisation	AI unit-economics analysis

Governance should have service levels. An undefined review process can become a barrier that encourages teams to bypass the CoE.

7.4 Funding model​

A blended funding approach is recommended.

Centrally funded​

• Enterprise platforms.
• Governance.
• Shared architecture.
• Core specialist capability.
• Champion network.
• Mandatory learning.
• Reusable assets.

Business-funded​

• Function-specific use cases.
• Workflow integration.
• Local adoption.
• Operational support.
• Benefits realisation.

Enterprise innovation fund​

• Cross-functional opportunities.
• Strategic experiments.
• New services.
• High-reuse capabilities.
• Emerging technologies.

Investment should move through evidence-based stages rather than receiving full funding at concept stage.

---

8. Designing the AI champion network​

8.1 Purpose of the network​

AI champions should connect central strategy with local work.

Their purpose is to:

• Translate policy into practical behaviour.
• Help employees use approved tools.
• Identify local opportunities.
• Support workflow redesign.
• Gather feedback.
• Spread successful practices.
• Escalate risks.
• Increase trust.
• Reduce the distance between the CoE and employees.

A 2024 systematic review of technology champions found that champion activities commonly involved promoting technology and helping colleagues adopt it in daily practice. It also found that champion responsibilities varied significantly by setting, reinforcing the need to define the role locally rather than apply a single generic model.

8.2 Champion roles​

Champions should:

• Demonstrate approved use cases.
• Provide first-line guidance.
• Organise local sessions.
• Support employees during initial adoption.
• Identify workflow opportunities.
• Capture user concerns.
• Share effective practices.
• Connect users to specialists.
• Monitor local adoption barriers.
• Support benefits measurement.
• Encourage responsible experimentation.
• Identify unsafe or inappropriate practices.

Champions should not:

• Give legal or risk approval.
• Operate as unpaid help-desk staff.
• Build every local solution.
• Police colleagues.
• Promote AI regardless of suitability.
• Own risks beyond their authority.
• Conceal failures to protect adoption statistics.

8.3 Champion selection​

Selection should consider:

• Peer credibility.
• Domain knowledge.
• Curiosity.
• Communication skills.
• Responsible judgement.
• Teaching ability.
• Influence without formal authority.
• Willingness to challenge.
• Capacity to perform the role.

The best technical employee is not automatically the best champion.

A balanced network should include:

• Technical employees.
• Business specialists.
• Managers.
• Operational employees.
• Junior employees.
• Experienced practitioners.
• Representatives from different locations and functions.

The network should include both:

• Emergent champions: employees already helping colleagues.
• Appointed champions: employees selected to ensure coverage.

8.4 Champion enablement​

Champions require:

• A formal role description.
• Protected capacity.
• Manager support.
• Priority training.
• Access to specialists.
• Reusable materials.
• A champion community.
• Recognition.
• Career development.
• Clear escalation routes.

Research reviewing champion preparation found that champion programmes often underinvest in adult-learning methods, organisational-change capability and sustained post-training support. The review recommends practical learning, personalisation, competency assessment and continuing mentor support.

8.5 Champion development pathway​

AI advocate​

• Understands policy.
• Demonstrates approved tools.
• Shares resources.
• Directs colleagues to support.

AI champion​

• Facilitates local learning.
• Supports workflow experiments.
• Collects feedback.
• Communicates successful practices.

AI practitioner​

• Designs advanced workflows.
• Supports evaluation.
• Contributes reusable assets.
• Coaches other champions.

AI transformation lead​

• Leads cross-functional redesign.
• Builds business cases.
• Coordinates business, technology and risk.
• Owns adoption and benefits.

This pathway turns the champion network into a talent and leadership pipeline.

8.6 Champion community of practice​

The CoE should provide:

• Monthly champion forums.
• Demonstrations.
• Office hours.
• Function-specific communities.
• Mentoring.
• Scenario-based training.
• An online knowledge hub.
• Use-case templates.
• Recognition events.
• Executive engagement.
• Access to pilots.
• Direct input into the CoE backlog.

The scaling cycle should be:

Local need → controlled experiment → measured evidence → reusable pattern → enterprise scale

---

9. Incentives for responsible adoption​

9.1 The design principle​

The firm should incentivise the outcomes it wants, not simply the visible activity it can count.

Desired behaviours include:

• Responsible experimentation.
• Measurable improvement.
• Collaboration.
• Knowledge sharing.
• Workflow redesign.
• Quality enhancement.
• Risk identification.
• Reuse.
• Employee coaching.
• Honest reporting.
• Stopping weak initiatives.

9.2 Financial incentives​

An evidence review by the Chartered Institute of Personnel and Development concluded that financial incentives can have a moderate-to-large positive effect on motivation and performance, although effectiveness varies by task and context. The review also found that incentives tend to work particularly well for less inherently interesting tasks, while effects can be weaker or negative for more intrinsically interesting work.

This implies that financial incentives should be:

• Closely connected to meaningful outcomes.
• Fairly distributed.
• Balanced between individual and team performance.
• Designed to avoid metric gaming.
• Supported by non-financial recognition.
• Reviewed for unintended consequences.

9.3 Recommended incentive portfolio​

Capability incentives​

• Funded learning.
• Credentials.
• Advanced assignments.
• Conference participation.
• Mentoring.
• Career pathways.

Recognition incentives​

• Leadership recognition.
• Internal awards.
• Champion status.
• Published case studies.
• Opportunities to present to executives.

Performance incentives​

Include relevant AI contributions in performance discussions, such as:

• Measurable workflow improvement.
• Quality improvement.
• Reusable asset creation.
• Successful coaching.
• Responsible risk identification.
• Verified client or customer impact.

Team incentives​

Reward cross-functional teams for:

• Scaled benefits.
• Sustainable adoption.
• Reuse across functions.
• Strong control performance.
• Positive user outcomes.

Innovation incentives​

• Protected experimentation time.
• Small innovation grants.
• Access to specialist support.
• Rapid approval for controlled trials.

9.4 What should not be incentivised​

Avoid rewards based primarily on:

• Prompt volume.
• AI login days.
• Number of tools used.
• Number of ideas submitted.
• Number of generated documents.
• Unverified hours saved.
• Pilot launches without adoption.
• Automation percentage without quality measures.

9.5 Reward responsible challenge​

The firm should recognise employees who:

• Detect material errors.
• Identify bias.
• Prevent data exposure.
• Challenge unsafe use.
• Improve evaluation.
• Recommend stopping an unviable pilot.
• Identify a cheaper or more reliable alternative.
• Report an incident promptly.

This demonstrates that responsible AI is not a barrier to innovation. It is part of professional quality.

---

10. Penalisation, consequences and fair accountability​

10.1 Reframe penalisation as consequence management​

Punishment should not be the primary mechanism for AI adoption.

Low AI usage may reflect:

• Poor tools.
• Inappropriate use cases.
• Lack of access.
• Insufficient training.
• Weak management.
• Accessibility needs.
• Legitimate professional judgement.
• Unclear policies.
• Lack of time.
• Fear about job security.

Applying penalties without diagnosing these factors is likely to create superficial compliance.

10.2 Accountability ladder​

Stage 1: Diagnose​

Determine the reason for low adoption or poor performance.

Stage 2: Enable​

Provide:

• Access.
• Training.
• Practice.
• Peer support.
• Workflow guidance.
• Manager coaching.
• Technical support.

Stage 3: Clarify expectations​

Define:

• Which workflow is expected.
• Why it is required.
• What good performance looks like.
• What validation remains necessary.
• When progress will be reviewed.

Stage 4: Formal performance management​

Where an employee persistently refuses to follow a reasonable, safe and clearly defined process after sufficient support, the issue may be managed through normal performance procedures.

The issue should be failure to meet a legitimate role requirement—not failure to generate a predetermined number of prompts.

Any formal process must be consistent with local employment law, contractual arrangements, consultation obligations, accessibility requirements and established HR policy.

10.3 Stronger consequences for deliberate misconduct​

Formal disciplinary consequences may be appropriate for:

• Uploading confidential information to prohibited tools.
• Bypassing required human review.
• Falsifying validation evidence.
• Concealing known harmful outputs.
• Deploying unauthorised systems.
• Deliberately bypassing security.
• Using AI for harassment, fraud or discrimination.
• Misrepresenting AI-generated evidence.
• Repeatedly breaching policy after training.

Consequences should be:

• Proportionate.
• Consistent.
• Evidence-based.
• Properly investigated.
• Connected to the risk created.
• Distinguished from good-faith mistakes.

10.4 Protect good-faith reporting​

Employees should not be penalised for:

• Reporting hallucinations.
• Raising ethical concerns.
• Admitting uncertainty.
• Seeking help.
• Stopping an unsafe activity.
• Reporting their own mistake promptly.
• Questioning an unsupported AI claim.

This protection is essential to preserve the learning behaviour required for safe adoption.

---

11. Measuring adoption, value and impact​

11.1 Balanced scorecard​

Access and foundations​

• Approved-tool availability.
• Provisioning time.
• Platform reliability.
• Data readiness.
• Support response.
• Champion coverage.

Capability​

• AI-literacy completion.
• Practical assessment.
• Role-based competence.
• Manager capability.
• Validation competence.
• Employee confidence.

Adoption​

• Relevant active users.
• Repeat workflow usage.
• Use-case penetration.
• Local adoption differences.
• Champion engagement.
• User satisfaction.

Operational outcomes​

• Cycle-time reduction.
• Output quality.
• Rework.
• Error rate.
• Cost per transaction.
• Throughput.
• Customer satisfaction.
• Employee experience.

Economic impact​

• Capacity created.
• Revenue generated.
• Margin improvement.
• Cost avoided.
• Client retention.
• Proposal conversion.
• Platform cost.
• Return on investment.

Risk and trust​

• Policy breaches.
• Security events.
• Validation failures.
• Human override rates.
• Complaints.
• Bias findings.
• Incident-resolution time.
• Audit findings.
• Model-performance degradation.

11.2 Benefits methodology​

Each use case should have:

1. A baseline.
2. A target outcome.
3. A benefits owner.
4. A measurement method.
5. A quality adjustment.
6. A cost view.
7. A capacity-conversion plan.
8. A review date.

A useful value equation is:

Net AI value = realised business benefit − technology cost − implementation cost − operating cost − control cost − rework and risk cost

This prevents gross time-saving estimates from being reported as realised financial value.

---

12. Impact the operating model can create​

12.1 Financial impact​

Potential outcomes include:

• Increased employee capacity.
• Lower administrative cost.
• Improved delivery margins.
• Reduced rework.
• Faster proposal production.
• Higher transaction throughput.
• Better use of specialist time.
• Lower duplication of technology investment.

12.2 Client and customer impact​

Potential outcomes include:

• Faster response.
• More consistent service.
• Greater personalisation.
• Improved access to expertise.
• Better issue resolution.
• Improved quality.
• More proactive insight.
• New AI-enabled services.

12.3 Workforce impact​

Potential outcomes include:

• Reduced repetitive work.
• Faster access to knowledge.
• Improved employee capability.
• Better support for less experienced staff.
• New career pathways.
• Increased internal mobility.
• Greater participation in innovation.
• Stronger collaboration between business and technology teams.

The evidence from customer-support work indicates that AI may be particularly valuable in helping less experienced employees apply knowledge more effectively.

12.4 Risk impact​

Potential outcomes include:

• Improved documentation.
• Better control consistency.
• Faster issue identification.
• More traceable decisions.
• Stronger policy checks.
• Better portfolio visibility.
• Reduced shadow-AI exposure.
• More consistent vendor review.

12.5 Strategic impact​

Potential outcomes include:

• Differentiated market propositions.
• New revenue models.
• Faster experimentation.
• Greater reuse of intellectual property.
• Stronger technology alliances.
• Improved attraction of AI-skilled talent.
• Enhanced organisational learning.
• Increased resilience as technology evolves.

---

13. Implementation roadmap​

Phase 1: Mobilise — first 90 days​

Leadership​

• Appoint executive sponsor.
• Establish AI Steering Committee.
• Agree strategic ambition.
• Define risk appetite.
• Communicate workforce narrative.

Portfolio​

• Inventory existing tools and pilots.
• Identify shadow-AI usage.
• Select priority workflows.
• Establish baselines.
• Stop duplicated initiatives.

Foundations​

• Approve initial platforms.
• Define interim policy.
• Establish secure experimentation.
• Create incident route.
• Define use-case intake.

People​

• Launch foundational AI literacy.
• Identify early champions.
• Survey employee concerns.
• Train initial managers.

Phase 2: Prove — months 3 to 6​

• Launch controlled workflow experiments.
• Establish the CoE service catalogue.
• Train the first champion cohort.
• Build reusable evaluation methods.
• Measure quality, time, cost and experience.
• Define benefits ownership.
• Review lessons monthly.
• Stop low-value initiatives.

Phase 3: Scale — months 6 to 12​

• Integrate successful use cases.
• Expand role-based learning.
• Grow champion coverage.
• Formalise portfolio governance.
• Establish AI FinOps.
• Publish reusable patterns.
• Introduce performance recognition.
• Conduct independent assurance.
• Report benefits to the executive committee.

Phase 4: Institutionalise — months 12 to 24​

• Embed AI capability in role frameworks.
• Incorporate relevant expectations into performance management.
• Mature the CoE.
• Expand external AI-enabled offerings.
• Establish ongoing model and vendor management.
• Align the management system with recognised governance frameworks.
• Publish an annual AI value, workforce and risk report.
• Continue retiring tools and use cases that do not create value.

---

14. Decisions required from the board​

The board should be asked to approve:

1. The enterprise AI ambition.
2. The federated operating model.
3. Executive accountability.
4. Initial CoE funding.
5. The enterprise AI platform direction.
6. Risk appetite and prohibited-use principles.
7. The workforce learning commitment.
8. The champion-network model.
9. The benefits-measurement framework.
10. The approach to incentives and accountability.
11. Priority use-case portfolios.
12. Quarterly reporting requirements.

The board should also request answers to the following questions:

• Which business outcomes will AI materially improve?
• Which workflows are being redesigned?
• What evidence supports the expected benefits?
• Who owns benefit realisation?
• What capabilities must the workforce develop?
• What risks are outside appetite?
• How are employees being consulted and supported?
• How will the firm distinguish genuine adoption from activity?
• Which initiatives should be stopped?
• How will the firm protect professional judgement and trust?

---

Conclusion​

AI adoption will not succeed through technology access alone.

The firms that create sustained advantage will be those that combine:

• Strategic clarity.
• Appropriate task selection.
• Workflow redesign.
• Trusted platforms.
• Role-specific capability.
• Employee participation.
• Local champions.
• Central enablement.
• Proportionate governance.
• Measurable benefits.
• Balanced incentives.
• Fair accountability.

The central leadership challenge is to avoid both uncontrolled experimentation and excessive central control.

The recommended model creates a disciplined system in which:

• The board sets direction.
• Executives own transformation.
• The CoE provides shared capabilities.
• Business units own outcomes.
• Champions support local adoption.
• Risk functions provide proportionate challenge.
• Employees build practical capability.
• Benefits are measured and realised.
• Deliberate misconduct has consequences.
• Good-faith learning and issue reporting are protected.

The ultimate measure of success is not whether employees are using AI more frequently.

It is whether the firm is producing better outcomes, developing stronger people, serving clients more effectively, managing risk more intelligently and building a sustainable competitive advantage.

Enterprise AI adoption should therefore be managed as a business transformation, a workforce transformation and a trust transformation—not simply as a technology rollout.