Beyond AI: Why the Operating Model Is the Hardest Part of Agentic Transformation
Models can be deployed in weeks. Redesigning authority, workflows, incentives, controls, and capital allocation is the real enterprise challenge.
Enterprises can buy access to advanced AI models almost instantly. They can launch copilots in a quarter and build a compelling agent demonstration in weeks.
What they cannot change nearly as fast is the enterprise around the technology: decision rights, incentives, workflows, risk ownership, workforce roles, data boundaries, and investment governance.
That is why the defining question has changed. It is no longer, “Should we use AI?” It is:
How do we redesign the enterprise when AI can participate in work, make recommendations, call tools, and take action?
This is not primarily a model-selection problem. It is an operating-model problem, and it helps explain the widening gap between AI activity and enterprise value.
McKinsey’s 2026 State of AI survey found that 44% of respondents report scaling AI across their enterprises, up from 38% the year before. Eighty percent say AI improves individual productivity. Yet only 37% report a positive contribution to EBIT, and just 6% qualify as AI high performers.
The technology is advancing. Adoption is deepening. Financial impact remains concentrated.
Closing that gap will require more than deploying more agents. It will require an operating model that determines where agents belong, what they may decide, how their performance is measured, and who remains accountable when something goes wrong.
Why the operating model is harder than the technology
Technology changes capability. An operating model changes behavior.
It determines how strategy becomes execution: who owns outcomes, who controls resources, how decisions are made, how risk is accepted, how performance is rewarded, and how the organization learns. Agentic AI touches every one of those mechanisms.
An agent does not merely produce content. It may interpret an objective, form a plan, retrieve data, call software tools, coordinate with other agents, initiate a transaction, and escalate an exception. As autonomy rises, the boundary between software execution and organizational decision-making begins to blur.
That exposes tensions a technology pilot can avoid:
- Business units want speed, while risk teams need evidence.
- Central platforms want consistency, while domain teams need flexibility.
- Finance wants predictable economics, while agent costs vary with usage and complexity.
- Employees want augmentation, while leaders may be tempted to treat automation as immediate capacity reduction.
- Vendors provide technology, but the enterprise remains accountable for outcomes.
The hardest transition is not from one model to another. It is from a hierarchy designed around human coordination to a system in which humans and agents share work under explicit authority.
That transition rests on four pillars: AI mindset, AI mandate, AI decision authority, and AI portfolio strategy.
Pillar 1: AI Mindset
Redesign the work, not just the task
Most AI programs begin with a productivity question:
How can AI help us perform this task faster?
That question can produce useful gains, but it often preserves the existing process. The same steps, queues, approvals, and handoffs remain. AI becomes another tool layered onto yesterday’s operating model.
An agentic mindset begins elsewhere:
If intelligence and execution were available from the start, how would we design this workflow differently?
That question shifts the focus from task automation to system redesign.
A customer-service agent, for example, might do more than draft a response. It could identify intent, retrieve the authoritative policy, authenticate the customer, propose a resolution, execute low-risk actions, and route exceptions to a person with the evidence needed to decide.
A software agent might do more than generate code. It could interpret a backlog item, inspect a repository, write tests, propose a change, run validation, and prepare a pull request, while remaining unable to merge or deploy without the required approvals.
The real question is not “humans or agents?” It is which work humans should stop doing, which work they should continue doing, and where human judgment becomes more important because machines can now execute more of the surrounding process.
This shift requires visible leadership. Leaders must use the tools, ask teams to challenge inherited workflows, reward evidence over enthusiasm, and make it safe to stop experiments that do not create value.
The evidence supports that emphasis. McKinsey reports that nearly three-quarters of AI high performers fundamentally redesign workflows, compared with only one-quarter of other respondents. High performers are not simply adopting faster. They are changing the work itself.
Mindset sets the direction. The next pillar turns that intent into an enterprise commitment.
Pillar 2: AI Mandate
Convert ambition into a bounded enterprise mission
“We will be AI-first” is not a mandate. It is a slogan.
A useful mandate tells the organization which outcomes justify AI investment, where the boundaries are, and how success will be measured. For a large enterprise, it should fit on one page and answer seven questions:
- Which three to five business outcomes matter most?
- Which workflows are strategic enough to redesign?
- What risk appetite applies to different classes of decisions?
- Which data, sovereignty, privacy, and security boundaries are non-negotiable?
- What is the workforce promise?
- What shared capabilities will the enterprise fund centrally?
- Which measures must move within 12 months?
The mandate prevents agent proliferation from masquerading as strategy.
Without it, functions build agents around local priorities. The enterprise accumulates overlapping tools, duplicated retrieval systems, inconsistent controls, and unclear ownership. Each project may be rational on its own, while the combined portfolio becomes expensive and ungovernable.
With a mandate, teams can innovate within a common frame. They know which outcomes matter, which controls are mandatory, and which reusable services they should consume rather than recreate.
The mandate must also include a workforce position. If employees believe every agent exists to enable immediate headcount reduction, they will withhold knowledge, resist redesign, or quietly route around the program. If leaders promise that no role will ever change, they avoid the reality of transformation.
A credible workforce promise is more balanced: the enterprise will redesign roles deliberately, invest in capability building, measure the human impact, and avoid converting speculative productivity estimates into automatic capacity targets.
A clear mandate establishes the destination and the boundaries. Decision authority determines how agents may operate inside them.
Pillar 3: AI Decision Authority
Make autonomy explicit before agents act
Most governance frameworks define who approves an AI system before launch. Agentic systems require a second layer: who authorizes each class of action during operation?
The relevant question is not simply whether a human is “in the loop.” Leaders must define:
- which human;
- at which stage;
- with what evidence;
- accountable for which outcome;
- able to stop or reverse which action; and
- responsible for reviewing exceptions and incidents.
An enterprise should establish an autonomy ladder:
- Assist: The agent drafts, retrieves, or summarizes.
- Recommend: The agent proposes; a human decides.
- Act with approval: The agent prepares an action; a human confirms it.
- Bounded action: The agent executes within pre-authorized limits.
- Orchestrate: The agent plans and coordinates multiple actions across a workflow.
Movement up this ladder should be earned through evidence. The decision should account for potential impact, reversibility, uncertainty, data sensitivity, affected parties, legal duty, observability, fallback options, and the quality of evaluation evidence.
This is where an AI control plane becomes essential. Policies cannot live only in documents or committee minutes. The enterprise needs runtime enforcement across identity, permissions, data access, tool calls, approvals, logging, cost, and rollback.
The NIST AI Risk Management Framework organizes AI risk management around continuous govern, map, measure, and manage functions. The UK Government AI Playbook similarly calls for meaningful human control and a chain of responsibility across the AI lifecycle.
For agents, those principles must become executable controls.
Authority establishes what an agent is allowed to do. Portfolio strategy determines where the enterprise should place its bets.
Pillar 4: AI Portfolio Strategy
Fund evidence, shared capability, and strategic options
Agentic AI portfolios often drift toward one of two extremes.
The first is pilot sprawl. Every team launches a proof of concept, but few projects have baselines, production controls, or scale criteria.
The second is narrow cost reduction. Labor-saving use cases dominate, while growth, customer experience, resilience, and new business models receive less attention.
A stronger portfolio balances three horizons:
- Horizon 1, reinvent work: Improve cycle time, service, quality, cost, and employee leverage.
- Horizon 2, grow: Create new offers, personalization, intelligent operations, and revenue.
- Horizon 3, create options: Explore new agentic capabilities, business models, and strategic control points.
Funding should advance through evidence gates:
- Explore: Is there a clear user, outcome, baseline, owner, and risk tier?
- Prove: Does the redesigned workflow create value and earn adoption under the required controls?
- Scale: Do unit economics, reliability, platform fit, and organizational readiness support expansion?
- Run or retire: Is the agent still creating enough value to improve, or should it be stopped safely?
Finance the next uncertainty, not the full vision.
That discipline matters because agent economics can be counterintuitive. McKinsey found that one in five respondents already says AI operating costs constrain usage, while 60% expect their organizations to increase AI investment. Token consumption, model routing, retries, retrieval, tool calls, human review, and exception handling all contribute to total cost.
The right metric is not cost per model call. It is cost per successful business outcome.
The portfolio should also recognize shared assets as sources of value. An evaluation harness, policy service, identity broker, knowledge layer, tool gateway, audit trail, or reusable workflow pattern may be more valuable than a single visible agent because it makes the next ten use cases faster and safer.
Together, the four pillars define the operating model. The next challenge is translating them into an organizational design that can scale.
How large enterprises should design an agentic operating model
The right design is neither fully centralized nor fully federated. Large enterprises should centralize what must remain consistent and federate what must remain close to value.
Centralize the control plane
The enterprise platform should provide:
- agent identity and ownership;
- model and tool registries;
- access controls and credential brokering;
- policy enforcement;
- common evaluation services;
- observability and trace capture;
- approval orchestration;
- cost attribution and budget controls;
- incident response and kill switches; and
- reusable data, knowledge, and retrieval patterns.
Federate workflow ownership
Business domains should own:
- outcome hypotheses;
- workflow redesign;
- domain-specific evaluation sets;
- user adoption;
- exception design;
- process and role changes; and
- realized value.
This division matters. A central AI team cannot understand every workflow deeply enough to redesign it. A business unit cannot independently build the full security, evaluation, observability, and governance stack every time it launches an agent.
The platform creates leverage. The domain creates value.
Pair product ownership with risk ownership
Each production agent needs a named outcome owner, technical owner, and risk owner. These roles should operate as a team, not as sequential approval gates.
The outcome owner defines success. The technical owner ensures reliability and integration. The risk owner ensures that authority, controls, and evidence match the impact.
Govern the workflow, not only the model
An agent’s behavior emerges from more than the model. It depends on prompts, memory, retrieved data, tools, permissions, policies, routing, and workflow state.
Model evaluation alone is therefore insufficient. The enterprise must test complete trajectories, including tool selection, data access, escalation, fallback, and recovery.
Build an operating cadence
At scale, leaders need a regular rhythm:
- weekly operational reviews for incidents, exceptions, reliability, and cost;
- monthly portfolio reviews for evidence, adoption, and investment decisions;
- quarterly mandate and risk-appetite reviews; and
- post-incident learning that updates policies, evaluations, and autonomy levels.
This turns governance from a launch event into a learning system.
Several enterprises are already demonstrating parts of this model, although none offers a complete blueprint.
What leading enterprises are doing differently
No company has completed the agentic transformation. The market is too early, and most public claims are self-reported. Still, several enterprises demonstrate useful operating-model patterns.
Morgan Stanley: evaluation before autonomy
Morgan Stanley’s wealth-management AI program is a strong example of decision discipline. According to an OpenAI case study, the firm built evaluation frameworks around real advisory use cases, used expert reviewers, tested retrieval and summarization quality, and kept advisors responsible for reviewing outputs. The company reports more than 98% adoption across advisor teams.
The lesson is not simply that Morgan Stanley deployed a capable assistant. It built trust through evaluation, narrow initial outcomes, domain expertise, and human accountability.
That is AI decision authority in practice.
Moderna: leadership, participation, and local invention
Moderna illustrates the mindset and mandate pillars. An OpenAI case study reports that the company’s earlier internal chatbot reached more than 80% employee adoption, followed by 750 custom GPTs within two months of adopting ChatGPT Enterprise. The program included executive sponsorship, structured change management, an internal community, and employee-led use-case creation.
The lesson is cultural. The platform mattered, but participation scaled because leaders created permission, capability, and visible commitment.
Walmart: purpose-built agents inside a shared architecture
Walmart describes a deliberate agentic strategy built around specific retail tasks, reusable capabilities, multi-agent orchestration, fallback handling, and explicit choices about where humans remain involved. Its corporate account of the strategy emphasizes purpose-built agents rather than general autonomy. Walmart also reports that its Trend-to-Product capability can shorten parts of the fashion production timeline by as much as 18 weeks, while its customer support and supply-chain systems increasingly route, recommend, and act.
The lesson is architectural. Agents become scalable when they share a platform but remain anchored to bounded workflows.
These examples are not directly comparable, and their performance claims have not all been independently audited. Their value lies in the common pattern: clear outcomes, deliberate controls, reusable foundations, and leadership ownership.
The same patterns become clearer when they are absent.
The perils when the operating model lags
The enterprise is accountable for the agent
The Air Canada chatbot case remains a compact warning. In Moffatt v. Air Canada, 2024 BCCRT 149, a customer relied on incorrect information from the airline’s chatbot about bereavement fares. The British Columbia Civil Resolution Tribunal found the airline liable for negligent misrepresentation.
The strategic lesson is larger than the damages awarded.
A company cannot place an automated system in a customer journey and treat its statements as separate from the enterprise. Customer-facing agents need authoritative knowledge, update ownership, traceability, escalation, and controls around policies, prices, eligibility, and commitments.
Accountability does not end at the model boundary.
Cost efficiency can hide experience degradation
Klarna shows why agentic success should not be reduced to automation volume. In its public filings, the company reported that its AI assistant handled 69% of service chats in the 12 months ending June 2025 and generated an estimated $39 million in 2024 cost savings. The same disclosures describe a dual-track approach that combines scalable AI with high-quality human support. Klarna’s SEC filing is a useful reminder that containment rate and cost savings are not the full outcome.
Customer trust, resolution quality, complex-case handling, recourse, and access to a capable human also matter.
The peril is not automation itself. It is optimizing one metric while the experience degrades somewhere the scorecard does not see.
Pilot success can create production risk
A demonstration runs in a controlled environment. Production introduces changing data, edge cases, ambiguous instructions, tool failures, permission drift, adversarial inputs, and cost variability.
If the organization scales agents without common observability, it may not know which agent acted, which data it used, why it chose a tool, or how to reverse the result.
Local success can become global fragility.
Premature workforce decisions destroy learning capacity
If leaders treat early productivity estimates as guaranteed capacity removal, they may eliminate the domain expertise needed to train, evaluate, supervise, and improve agents.
The human role often changes before it disappears. People move toward exception handling, judgment, customer empathy, policy ownership, and system supervision. Removing expertise before the new operating model is stable can reduce quality and slow improvement.
The answer is not to wait for perfect certainty. It is to build an operating cycle that creates evidence quickly and safely.
A practical 90-day starting point
Large enterprises do not need a perfect target-state organization before they begin. They need a repeatable operating cycle.
Days 0 to 30: Frame
- Publish the one-page AI mandate.
- Name executive, outcome, technical, and risk owners.
- Select three to five workflows with measurable baselines.
- Define autonomy tiers and non-negotiable controls.
- Inventory current agents, tools, platforms, and hidden usage.
Days 31 to 60: Prove
- Redesign each workflow with users.
- Instrument value, adoption, quality, cost, and risk.
- Test complete agent trajectories, not only model responses.
- Exercise approval, escalation, fallback, rollback, and incident paths.
- Identify reusable platform capabilities.
Days 61 to 90: Institutionalize
- Scale, reshape, or stop each initiative based on evidence.
- Publish delegated decision pathways.
- Establish the agent registry and production control plane.
- Rebalance the portfolio across efficiency, growth, and strategic options.
- Launch quarterly executive reviews of value, autonomy, risk, and capital.
The output of the first 90 days should not be a showcase agent. It should be a system that can repeatedly turn strategic intent into trusted execution.
The real competitive advantage
Models will improve. Agent frameworks will mature. The cost of intelligence will fall in some places and rise in others as organizations ask agents to do more complex work.
Those technologies will matter, but they will not be the enduring differentiator.
The durable advantage will be organizational:
- the ability to redesign work;
- the clarity to define a mandate;
- the discipline to delegate authority safely;
- the evidence to allocate capital well;
- the infrastructure to observe and control action; and
- the leadership to remain accountable for outcomes.
The winners will not be the companies with the most agents. They will be the companies with the strongest operating model for turning autonomy into trusted, compounding value.
That is what lies beyond AI.
Sources
- McKinsey & Company, The State of AI in 2026: On the Road to ROI
- NIST, AI Risk Management Framework Core
- UK Government, Artificial Intelligence Playbook
- OpenAI, Morgan Stanley Uses AI Evals to Shape the Future of Financial Services
- OpenAI, Moderna
- Walmart, Inside Walmart’s Strategy for Building an Agentic Future
- Moffatt v. Air Canada, 2024 BCCRT 149
- Klarna Group SEC filing