Scale —
From Pilot to Enterprise

The protected pilot is a familiar failure mode in enterprise technology. The pilot succeeds because it is protected: the use case is carefully chosen, the users are volunteers who wanted to participate, the data is clean, the facilitator is expert, and the steering group is engaged. When the pilot ends and the results are presented, leadership approves the rollout. And then the rollout begins to fail, slowly and quietly, in ways that no one wants to report.

The failure is not technical. The conditions that made the pilot succeed — controlled, voluntary, supported, clean data — do not replicate across the organisation. The mandatory user in Wave 3 who never wanted to try AI and whose line manager is sceptical is not the same person as the volunteer in Wave 1 who was excited to be selected. The messy SharePoint of the Operations team is not the tidy SharePoint of the Finance team that was the pilot. The support infrastructure that was highly responsive for 20 users becomes stretched for 2,000.

Gartner's finding that only one in three AI projects reaches production is partly a story about this gap. The organisations that close it are not the ones that ran better pilots — they are the ones that invested as much effort in the conditions for Wave 3 success as they did in Wave 1 selection. That investment includes governance infrastructure, manager enablement, change management capability, and — crucially — a Centre of Excellence that exists not to manage the pilot but to sustain the programme.

An AI Centre of Excellence (CoE) is not a committee. A committee is a group of people who meet to approve things. A CoE is a practice — a small, cross-functional team with clear ownership, active responsibilities, and the authority to make decisions. The distinction matters because committees produce delays, ambiguity, and decisions-by-default. A CoE produces momentum.

The core membership covers five roles. The AI Lead owns the strategy, the programme, and the relationship with IT and Legal on governance. The Data Governance owner ensures that the organisation's data classification and acceptable use framework stays current and enforced. The Change Lead owns the adoption programme, the champions network, and the communication strategy. Business Unit Champions represent the frontline perspective and carry intelligence from deployment back to the centre. The L&D Partner owns the learning programme, the personal learning plans, and the connection to ESCO competence frameworks.

The CoE should meet fortnightly at minimum in the first year, moving to monthly once the programme is established. Its output is not minutes and action logs — it is decisions made and communicated, problems identified and resolved, blockers escalated and cleared. Measure the CoE by its velocity, not its process.

The McKinsey Global Institute's June 2023 report on generative AI estimated $2.6 to $4.4 trillion in annual value potential across the global economy. That number is real but useless for measuring your deployment. The relevant question is not what AI is worth in aggregate but what it is delivering for your organisation, in your context, measured against the investment you have made.

There are three legitimate ways to measure AI value, and they require different measurement approaches. Time savings — the most commonly cited metric — is quantifiable if you measure it properly. Not self-reported estimates ("Copilot saves me about an hour a week") but structured measurement: before/after time logs for specific task types, collected systematically across a sample of users, with sufficient volume to be statistically meaningful. The Microsoft Work Trend Index found that AI power users save an average of 30 minutes per day — but that is an average across many use cases and many user profiles, not a number you can simply apply to your context.

Quality improvements require a different measurement approach — one that most AI ROI frameworks skip because it is harder. Quality improvement means: the output of this task was better after AI deployment than before it. For some task types this is measurable: error rates in document production, customer satisfaction scores for AI-assisted responses, revision cycles on first drafts. For others it requires qualitative assessment by subject matter experts. Both are valid. The key is to define what "better" means for each task type before you measure it.

New capability — the value of being able to do things you could not do before — is the hardest to measure and potentially the largest. The Finance team that can now run a scenario model in three hours that previously took three days has not just saved time — it has fundamentally changed what decisions it can make and how fast it can make them. This is strategic value, and it belongs in the ROI picture even though it resists precise quantification.

The governance framework that works for 20 Wave 1 users needs fundamental redesign before it reaches 2,000. Not because the principles change — the same data classification, the same acceptable use rules, the same ethics commitments — but because the enforcement mechanisms, the exception handling processes, and the incident response infrastructure all need to operate at a different scale.

The key structural insight is that governance tiers need to match the zone framework. Zone 2 Copilot use — an employee using Copilot to draft an email that they review before sending — requires governance that is lightweight and enabling. The acceptable use policy, a data classification label on the email, and a brief user training is sufficient. Zone 4 automation — an agent that processes invoices end-to-end, matches them to purchase orders, and updates the ERP system without human review of individual transactions — requires a fundamentally different governance tier: a formal risk assessment, tested exception handling, audit trail, periodic human review of aggregate output, and a named process owner who can suspend the automation if something goes wrong.

Exception handling is where most governance frameworks fail at scale. The framework is clear about the standard cases; nobody has thought through the edge cases. What happens when a Zone 2 user accidentally processes confidential data through a Zone 2 tool? What is the incident response path? Who needs to be notified? What is the remediation process? These questions need answers before you have incidents, not after. Build the playbook in advance.

The ethics board — or ethics advisory function, in smaller organisations — does not need to be a large formal body. It needs to be a named group of people, meeting at least quarterly, who review the AI programme against the organisation's stated ethics commitments and surface concerns. Its value is in the discipline of asking the questions, not in the formality of the structure.

A generic AI adoption programme delivered uniformly across every business unit is a missed opportunity. The McKinsey analysis of where generative AI creates most value identified four sectors — customer operations, marketing, software engineering, and R&D — but within any enterprise, the task profiles and therefore the AI opportunity profiles are radically different between functions. The path to value in Marketing looks nothing like the path in Finance, which looks nothing like Operations.

Business unit playbooks, built from O*NET and ESCO task maps, give each function a deployment plan that speaks to their actual work. The worked example below illustrates how the four-zone framework applies differently across three representative functions. The purpose is not to create separate programmes but to create a common framework applied with specific, credible detail to each context.

Each playbook should be developed with a subject matter expert from the relevant BU, validated against O*NET task data for the primary occupations in that function, and reviewed against ESCO competences to identify the learning resources that support Zone 2 and Zone 3 adoption. The playbook is a living document — it should be updated at each wave boundary as the frontier becomes clearer.