Insights
An operating model for enterprise AI
Assigning decision rights, ownership, funding, and operational responsibility across an enterprise AI portfolio.
In brief
Who owns an AI system’s value, changes, and failures?
- Assign a business owner while distinguishing shared platform and domain responsibilities.
- Fund evaluation, human review, and operations throughout the lifecycle.
- Make release decisions and portfolio reviews explicit, repeatable, and evidence-based.
An enterprise AI operating model determines who chooses the work, who owns its consequences, and how the organization sustains it after launch. Technology choices sit inside that structure. Without clear ownership, a promising prototype can remain dependent on a small group of enthusiasts while its users assume someone else is responsible for reliability.
The operating model should connect business decisions, research, engineering, security, and domain expertise. It should allow experimentation while defining the evidence required to move a capability into a consequential workflow. The objective is a repeatable way to turn technical capability into accountable institutional practice.
Assign a business owner to the outcome
Every application needs an owner who can define the workflow, approve changes to it, and measure whether the system helps. This person or team should have authority over the process affected by the AI. A sponsor who can fund a demonstration but cannot change the workflow may be unable to realize the intended value.
The owner should define the target users, eligible tasks, desired outcomes, and unacceptable failures. They also need to identify what happens when the system is unavailable. These decisions shape the architecture and evaluation program before model selection begins.
Outcome ownership continues after deployment. If users stop trusting the system, if review effort rises, or if the process changes, the business owner should participate in deciding whether to repair, narrow, or retire the application. Ownership is an ongoing responsibility rather than a signature at launch.
Distinguish shared capabilities from domain decisions
A central platform team can provide identity integration, model access, evaluation infrastructure, observability, deployment tooling, and reusable security controls. Domain teams contribute workflow knowledge, source interpretation, acceptance criteria, and user feedback. Neither group can substitute fully for the other.
The boundary should be explicit. A shared retrieval service can enforce access and provenance, while a domain team decides which source governs a particular business measure. A model gateway can manage provider access, while the application team decides whether a response meets the task’s quality threshold.
Excessive centralization can create a queue for every experiment. Excessive fragmentation can produce duplicated infrastructure and inconsistent controls. A practical model standardizes the boundaries that need consistency while allowing domain teams to adapt the parts that determine usefulness.
A synthetic portfolio allocation
Consider an organization evaluating three initiatives. One assists employees with internal policy questions, another prepares investment research drafts, and a third proposes actions in a production operations system. All use language models, but their operating requirements differ.
| Initiative | Primary owner | Critical acceptance evidence |
|---|---|---|
| Policy assistant | Internal service function | Current sources and access-aware answers |
| Research drafting | Research leadership | Attribution, numerical accuracy, reviewer effort |
| Operations agent | Operations leadership | Authorized actions, recovery, confirmed outcomes |
A single generic approval process can overlook these differences. Each initiative should inherit common controls and add requirements tied to its actual consequences. Funding should cover those requirements, including the people who review outputs and operate the system.
The portfolio review should also identify reusable investments. A common evidence store or evaluation service may support several initiatives, but its value should be assessed through the applications it enables and the duplication it removes.
Create decision rights for model and system changes
AI systems evolve through model updates, prompt changes, source changes, and new tools. The organization needs to know which changes a product team can make independently and which require additional review. The answer should depend on the effect of the change and the system’s operating boundary.
A wording adjustment with unchanged behavior may require routine testing. Adding a tool that can modify external records changes the authority surface and deserves a different review. A new model may require evaluation even if the application code remains identical.
The NIST AI Risk Management Framework offers a voluntary structure for incorporating trustworthiness considerations into AI design, development, use, and evaluation. An enterprise can use that structure to organize responsibilities while defining its own concrete operating procedures. NIST AI RMF.
Fund the complete lifecycle
The cost of an AI application extends beyond implementation and model usage. It includes source maintenance, evaluation updates, incident response, user support, access management, and the work required to respond when a provider or business process changes.
A funding model that pays only for the prototype creates an incentive to defer these responsibilities. The resulting system may launch with unclear ownership of recurring costs. A more useful investment case separates discovery, implementation, and ongoing operation, with explicit assumptions about volume and review effort.
Research on technical debt in machine-learning systems draws attention to dependencies and system-level maintenance burdens. The operating implication is to budget for the surrounding system and its evolution, alongside the model itself. Hidden Technical Debt in Machine Learning Systems.
Build a pathway from research to production
Research and production engineering have different immediate objectives. Research may investigate an uncertain capability, while production engineering must deliver predictable behavior within a defined service boundary. A useful operating model creates a clear transfer between those activities.
A research result should arrive with its task definition, data conditions, baselines, evaluation results, and known limitations. The production team then assesses deployment constraints, integration, security, cost, and recovery. A successful experiment can justify further engineering without being treated as a ready production component.
The reverse transfer also matters. Production failures can reveal research questions about reasoning, representation, calibration, or long-horizon behavior. A structured feedback process allows operational evidence to inform research priorities without turning every incident into an unbounded model-development project.
Make human review a designed service
Human review requires capacity, expertise, and an interface. Assigning every output to a reviewer without measuring the work can create a hidden bottleneck. The operating model should specify which outputs need review, who is qualified, and what evidence the reviewer receives.
Reviewers should be able to correct the system and record why a correction was necessary. Their decisions can improve evaluation cases and identify source or workflow defects. The organization should distinguish disagreement with a model from a change in policy or business preference.
Review quality also needs monitoring. A high approval rate may reflect accurate outputs, superficial review, or pressure to clear a queue. Sampling and adjudication can help determine which explanation applies. Human presence alone does not establish an effective control.
Establish a compact portfolio review
A recurring portfolio review should examine outcomes, adoption, operating cost, incidents, and unresolved dependencies. It should also decide where to expand, where to narrow scope, and where to stop. These decisions prevent the organization from accumulating applications whose original purpose has disappeared.
The review can use a small set of comparable measures while preserving domain-specific detail. Cost per accepted outcome, active usage, quality trends, and operational ownership are often useful. A single ranking based on model accuracy or number of users can obscure whether an application delivers meaningful value.
Retirement should be planned. Users need a transition path, records need an appropriate retention decision, and credentials or integrations should be removed when no longer required. An application that is rarely used can still create operational exposure if it remains connected to important systems.
Design the organization around accountable learning
An effective operating model makes learning continuous and responsibility clear. Teams can experiment within bounded environments, produce evidence for expansion, and revise decisions when conditions change. Shared infrastructure reduces repeated work, while domain ownership preserves the meaning of the task.
The result is an enterprise capability that can survive changes in models, vendors, and personnel. AI becomes part of the organization’s operating practice through clear decision rights, funded maintenance, and measurable outcomes. Those foundations allow technical ambition to translate into durable execution.