AI transformations usually stall because the operating model stays the same while the technology changes. Decision rights are unclear, funding ends with the project, no team owns the workflow after launch, and governance meets too rarely to matter. Fix those four things around two or three real workflows first, and tool choices become easier and less risky.
What the evidence says about where value comes from
McKinsey's The state of AI in 2025, based on a survey of 1,993 participants fielded between June 25 and July 29, 2025, describes a gap between use and results. Only 39 percent of respondents reported EBIT impact at the enterprise level. McKinsey defines high performers as respondents who attribute more than 5 percent of EBIT to AI and report significant value, and they make up about 6 percent of the sample.
The most useful finding for an operating-model discussion is what separates them. McKinsey reports that high performers are nearly three times as likely as other respondents to have fundamentally redesigned individual workflows, and that this redesign is among the strongest contributors to business impact of the factors it tested. This is survey correlation, not proof of cause. It still points away from the tool and toward how work, ownership and decisions are arranged.
That fits what many executives see. A model dropped into an unchanged process produces a faster step inside the same queue, the same approvals and the same handoffs. The cost falls a little. The outcome does not change.
Define what an AI operating model is
An AI operating model is the set of arrangements that decide who may build and deploy AI, who is accountable for what it does, how money flows to it, how teams are organized around it, and how performance and risk are reviewed. It is not the technology architecture, and it is not a policy document. It is the working answer to five questions: who decides, who pays, who builds, who runs, and who checks.
The next sections take each in turn. The order matters, because decision rights shape every other choice.
Decision rights: who decides what
Most stalled programs have no shortage of committees. They have no clarity about which decisions each committee can actually make. A use case sits in review for months because nobody can say whether the business owner, the technology function, risk or legal has the final word.
List the recurring decisions and assign each one to a single accountable role.
- Use-case approval: is this workflow worth building, and who is the business owner?
- Risk acceptance: which residual risks are acceptable, and who signs for them?
- Data access: which data may the system use, and who grants it?
- Tool and model selection: which platforms are approved, and by whom?
- Release and rollback: who can put a system into production, and who can switch it off?
- Autonomy level: how much may an agent do without a person approving it?
NIST's AI Risk Management Framework 1.0, published in January 2023, says the same in governance terms. GOVERN 2.1 calls for documented roles, responsibilities and lines of communication for managing AI risk, clear throughout the organization. GOVERN 2.3 states that executive leadership takes responsibility for decisions about risks associated with AI development and deployment. In practice this means a named executive, not a committee, owns the risk appetite, and delegates specific decisions below it with limits.
Write the delegations as limits, not as principles. A business owner may approve a use case below a cost ceiling, in a risk tier the executive has pre-approved, using approved platforms. Anything outside those limits goes up. Teams then know when they can proceed without asking.
Funding: from projects to owned capabilities
Project funding assumes that work ends. AI systems do not end. Models drift, data changes, prompts need revision, usage costs scale with volume, and the process around the system keeps evolving. A budget that closes at go-live leaves nobody paying for the monitoring and improvement that decide whether value persists.
Fund the workflow, not the project. Give each AI-enabled workflow a standing budget with three lines: build and improvement, run costs such as compute, licenses and monitoring, and the people who review and handle exceptions. Release money in stages against evidence, and let the business owner, not the technology function, hold the target for the outcome.
Keep a small central pool for shared capabilities, such as evaluation tooling, data pipelines and approved platforms, so that each team does not rebuild them. Charge usage transparently so that owners see what their workflow costs. Where benefits depend on removing work or changing a process, name who gives up capacity and who confirms the saving. Tracking methods for banked value are covered in the benefits realization guide on this site, so that detail is left aside here.
Product teams: one team owns the workflow after launch
The usual pattern is a project team that builds a system and then hands it to an operations team that did not help design it. The handoff is where context is lost. The fix is a persistent product team that owns a workflow end to end, before and after launch.
A workable team has a business product owner who holds the outcome, engineers who build and maintain the system, someone responsible for data and evaluation, and an operations lead who runs the exception handling. A risk or compliance partner joins the team's rhythm rather than reviewing at the end. Where reviewers are people in the line of business, they are part of the team, not an external service.
Pair the product teams with a small platform group that provides shared services and standards. The platform group sets the paved road: approved models, logging, evaluation and security patterns. Product teams use it by default and justify any departure. This keeps speed in the teams and consistency in the platform.
Start with two or three workflows, not twenty. Pick ones with a clear owner, available data and an outcome measure you can check. The point is to prove the arrangement, then copy it.
Governance cadence: who checks, and how often
A policy cannot govern a system that changes weekly. Governance needs a cadence with different forums for different time scales.
- Weekly or biweekly delivery review: each product team shows outcomes, incidents, costs and open risks, and requests decisions that are within its limits.
- Monthly portfolio and risk review: a small group under the accountable executive reviews all AI workflows by tier, looks at incidents and drift, approves changes to autonomy and stops weak projects.
- Quarterly executive review: leadership revisits risk appetite, funding, capability gaps and what the portfolio has delivered.
Anchor the cadence in an inventory. NIST's GOVERN 1.6 asks for mechanisms to inventory AI systems, resourced according to risk priorities, and GOVERN 1.5 asks for ongoing monitoring and periodic review of the risk management process with clearly defined roles. You cannot review what you have not listed.
ISO/IEC 42001, published in December 2023, is a management-system standard. It specifies requirements for establishing, implementing, maintaining and continually improving an AI management system. Whether or not you seek certification, its logic is useful: leadership sets the direction, roles are assigned, risks and impacts are assessed, performance is evaluated and the system is improved. That is the loop your cadence should run.
If you deploy high-risk systems in the EU, some of this is also a legal duty. Article 26 of the AI Act requires deployers to monitor operation, to suspend use when they have reason to think a system poses a risk, and to keep automatically generated logs for at least six months where they control them. A monthly review with real logs behind it is how you show that you do it.
Skills and capacity are part of the model
An operating model also needs people who can work in it. The World Economic Forum's Future of Jobs Report 2025 found that 63% of employers surveyed see skill gaps as a major barrier to business transformation over 2025 to 2030, and 85% plan to prioritize upskilling their workforce. Include in the model who is trained to own, review and improve AI workflows, and how those skills are recognized in roles and pay.
What to change first
Sequence matters. Changing everything at once recreates the committee problem. A workable order follows.
- Decision rights. Name the accountable executive, list the recurring decisions and set delegation limits. This takes weeks, costs little and removes the largest source of delay.
- Workflow selection and ownership. Choose two or three workflows, name a business owner for each and set an outcome measure and baseline.
- Product team. Form a persistent team around each workflow, with the operations and risk roles in the team.
- Funding. Move those workflows to standing budgets with staged release and a shared platform pool.
- Cadence and inventory. Start the inventory and the monthly review, and add the quarterly executive view after two cycles.
Where to start
Begin with a one-page decision-rights map, since it exposes most of the problem in a single meeting. Ask who can approve, fund, deploy and stop a given AI workflow today, and count how many different answers you hear.
Then take your most advanced pilot and ask four questions. Who owns the outcome after launch? Who pays for it next year? Who reviews it, and how often? What would trigger a rollback? Gaps in any answer are the first items to fix. Redesign the workflow around the answers before buying further tools, because that is the factor McKinsey's high performers share.
References
- McKinsey: The state of AI in 2025, Agents, innovation, and transformation
- NIST: Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1
- ISO: ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system
- EU AI Act: Article 26, Obligations of deployers of high-risk AI systems
- World Economic Forum: The Future of Jobs Report 2025





