The executive committee approves enterprise AI licences, funds an academy and sees thousands of employees complete training. Three months later, the operations director is still asking why customer cases take as long to close, why analysts are copying material between systems, and why managers are reviewing more drafts than before.
That is the adoption gap in its most familiar form. The organisation has made people more capable of using a tool without changing the work that gives the tool a purpose. Training has value. It teaches limits, safe handling of information and enough fluency to judge an output. It cannot remove an unnecessary approval, resolve competing data definitions, put evidence into the system where the next person works, or give someone authority to act on a recommendation.
OpenAI’s December 2025 enterprise report describes rising use of structured features such as Projects and Custom GPTs rather than casual querying alone. It is vendor-reported evidence from its own customer base, not a universal benchmark. Even so, the direction matters: repeated work with situated context is more likely to endure than a broad invitation to use AI whenever it seems helpful. OpenAI
The leadership task is therefore more exacting than raising AI literacy. Pick a business workflow, decide what must work differently, and redesign the conditions around the AI step until the change is visible in the operating result.
Training is an entry ticket, not an operating model
A course answers whether an employee can use a tool. Adoption asks whether that use changes the organisation’s performance. The distinction becomes clear when the work crosses a team boundary.
Take supplier-risk assessment. An analyst can learn to turn a supplier dossier into a coherent first draft in minutes. That does not establish which supplier record is authoritative, whether contractual and sanctions information is available, which risk findings require legal review, who may accept a residual risk, or how the decision is recorded. If the analyst still pastes the text into a spreadsheet, email and procurement system before waiting for the weekly committee, the organisation has improved prose rather than flow. It may also have created a larger queue for legal and procurement to inspect.
The UK Government’s 2025 guidance on scaling generative AI puts sustained, high-quality use at the centre of the task. It calls for attention to user journeys, support, risk management and the way a tool becomes embedded in routines and team processes. Course-completion rates cannot answer those questions. UK Government
A CIO should be wary of any adoption dashboard that shows licences provisioned, monthly active users and prompts submitted, but cannot show which decision, hand-off or service outcome has improved. Activity can be a useful diagnostic. It is not proof of business change.
Start where work actually slows down
Before choosing a model, prompt library or training cohort, map the current path of work. Leave the polished process diagram on the wall for a moment. Follow a real case through inboxes, chat messages, spreadsheets, systems of record, informal checks and escalations. The productive friction is usually there.
A compact discovery session should establish five facts:
- the event that starts the work and the outcome that closes it;
- where people wait, rekey information or chase missing context;
- which judgments are repeatable and which require named accountability;
- the source of truth for each important fact; and
- the evidence needed to defend the outcome to a customer, auditor, regulator or manager.
This is a control exercise as much as a design exercise. It will often show that AI is the wrong answer to the immediate problem. A missing interface may call for an API integration. A recurring decision may need a rule. A confusing policy may need rewriting. There is nothing disappointing about that result. Removing waste before adding a probabilistic system is sound management.
Where AI does fit, state its job in operational terms. It may retrieve relevant evidence, classify an incoming case, prepare a draft, recommend a next action or execute a bounded action. “Assist the team” is not a job description. It hides the questions that later surface as rework: what data is used, who relies on the result, what authority is delegated and where uncertainty goes.
Make every AI hand-off a contract
A prompt is an interaction between a person and a model. A production workflow needs a contract between roles and systems.
For each AI-supported step, define the input, output, recipient, decision right, assurance rule and exception path. If an AI-generated case summary feeds a service decision, it should carry links to source material, disclose incomplete context, land in the case record, and tell the reviewer whether to approve, amend, reject or escalate. The design should also say what happens if the reviewer finds a material error.
That last point separates learning from anecdote. A correction needs a destination. Was the failure caused by poor source data, an obsolete knowledge article, an ambiguous policy, a flawed prompt or an unsuitable task? If nobody captures the answer, the programme gathers colourful opinions—“it is excellent” or “it cannot be trusted”—while the same avoidable defects return every week.
NIST’s Generative AI Profile frames risk management across the lifecycle rather than as a final compliance gate. Its emphasis on governance, oversight, appropriate human-AI roles and evaluation proportionate to risk has a direct operational consequence: the work design and the controls must be designed together. NIST AI 600-1
A human reviewer is not automatically a control. Reviewers need enough time, evidence and authority to challenge an output. A person who is expected to rubber-stamp hundreds of recommendations before a service-level deadline is merely a liability transfer mechanism. A person who receives well-defined exceptions, sees the evidence and can change the decision is providing assurance.
Put accountability with the leader who can change the process
Central AI teams should own shared platforms, model standards, guardrails and reusable patterns. They rarely own the conditions that determine adoption. A learning-and-development function can lift competence; it cannot alter the target operating model. Someone close to the business result must carry the accountability.
For customer service, that is commonly the operations leader accountable for resolution quality, repeat contacts and escalations. In finance, it may be the controller responsible for the speed and integrity of close. In engineering, it is the leader accountable for delivery performance and reliability. The product or platform team owns the service; the operational leader owns the outcome. Both names should appear in the decision record.
This is where polished pilots often stall. An innovation team can demonstrate a useful assistant. It cannot settle a local dispute over data ownership, retire an old approval, reserve frontline time for experimentation or accept the risk of changing a customer decision. Those are leadership calls, and the owner needs the authority to make them.
Agree measures before rollout. Begin with a baseline, choose a target outcome and schedule a review. For service work, use resolution quality, repeat contact and escalation accuracy alongside cycle time. For engineering, watch review latency, defect escape and recovery work as well as throughput. For risk operations, evidence completeness, exception rate and time to decision matter more than documents generated. The aim is to reveal whether the workflow is healthier, rather than simply more active.
Change the conditions around the tool
Employees adopt a new path when it makes the right action easier than the old one. That requires unglamorous engineering and management work.
Approved knowledge must be available at the point of use and ownership must be clear when it is wrong or outdated. Recurring tasks need templates that reduce routine effort without prescribing away professional judgement. Sensitive information and high-impact decisions need explicit boundaries.
Do not make people choose between delivery and learning. If the only sanctioned space for AI experimentation is an innovation week or unpaid time, adoption will concentrate among the already confident and bypass the teams with the most demanding operational problems. Build a limited trial into the value stream: run the new path against a baseline, inspect the outcomes, improve the design and decide whether it deserves a wider rollout.
The UK guidance makes a related warning: human-in-the-loop arrangements and terms and conditions do not suffice when users lack the expertise, time or authority to challenge an output. Leaders should treat that as a design constraint, not a legal footnote. A review step that cannot alter the result is theatre.
Use a thin slice to expose the truth
Enterprise-wide licence announcements are easy to launch and hard to learn from. A thin, observable workflow slice is the better starting point.
Choose work with a stable trigger, accessible context, a known user group and a measurable outcome. It should be contained enough to instrument, yet important enough for failure to teach something useful. Preparing a first response for a defined class of customer query, assembling evidence for a low-risk compliance check, or turning a standard change request into a testable implementation plan can meet that test. Autonomous credit approvals and high-stakes employment decisions do not belong in the first wave.
Run the changed path beside the current one for a defined period. Sample outputs with the people accountable for the work. Compare quality, cycle time, escalation, rework, user effort and control exceptions. Then take a hard decision: scale, redesign or stop. A programme that cannot stop weak use cases will accumulate a costly estate of demonstrations.
DORA’s 2025 research makes a comparable point for software delivery: AI amplifies the strengths and weaknesses already present in an organisation, and the strongest returns come from attention to the underlying system rather than the tool alone. The principle travels well. AI exposes the quality of operational design. DORA
The leadership test is simple. On Monday morning, can the owner point to a named workflow and explain what now happens differently: which hand-off disappeared, what evidence is available, who can decide, where exceptions go and which outcome will move? If the answer is a list of courses, licences and prompt tips, the organisation has prepared people for AI without preparing the work.
Sources
- OpenAI, The state of enterprise AI, 8 December 2025.
- UK Government, The People Factor: a human-centred approach to scaling AI tools, 4 June 2025.
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1), 26 July 2024.
- DORA, State of AI-assisted Software Development 2025, 2025.