Why AI Requires Operating Model Reinvention

Local gains need a path to enterprise value

There is credible evidence that AI assistance can improve real work. In research published in 2025, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied a staggered rollout involving 5,172 customer-support agents serving one company. Access to assistance increased issues resolved per hour by 15% on average, with substantially different effects across workers. The rollout occurred mainly in late 2020 and early 2021. This is evidence from a particular operating setting, not a universal savings forecast. Generative AI at Work, The Quarterly Journal of Economics.

The operating question is where a gain travels next. If preparation becomes faster but approval remains the constraint, the customer may still wait. If more cases reach a specialist without better context, review effort may increase. If time is released in scattered intervals that cannot be reassigned, useful relief may never become additional service capacity.

These are different outcomes and should be valued honestly. Reduced effort can improve employees’ working experience. Usable capacity may help absorb demand. Cash savings require an actual change in expenditure. Leaders should identify which benefit they intend to realize and what has to happen after the AI task for that benefit to appear.

That connection determines the appropriate scope of redesign.

AI changes the assumptions beneath the work

An operating model describes how an organization delivers its purpose: how work moves, how decisions are made, who carries responsibility, and how resources and controls support the result.

AI can alter the assumptions behind that arrangement. In the customer-support study, less experienced workers benefited more, while the most skilled workers saw small speed gains and small quality declines. That finding gives leaders a reason to test different roles and review arrangements instead of assuming one deployment pattern suits everyone. Generative AI at Work.

There is also a different uncertainty to manage when generative AI produces the information used in a decision. NIST’s 2024 Generative AI Profile identifies risks from false outputs presented confidently and from people placing excessive trust in AI. A fluent recommendation can therefore require a different kind of checking from a completed rules-based transaction. NIST Generative AI Profile, sections 2.2 and 2.7.

The design implication is to examine the relationships around the capability. If an employee is accountable for a decision, can they see the supporting evidence, challenge the recommendation, and obtain help in time? If the answer depends on a policy, who establishes which version applies and resolves conflicting guidance?

Those choices are connected. More delegation increases the importance of reliable context and effective intervention. Retaining review preserves a human decision point but consumes attention and can create a bottleneck. The operating model must make that trade-off workable at the expected volume.

An illustration: redesigning a claims decision

Hypothetical example — proposed design, not a reported engagement or measured result.

Consider an insurer seeking to reduce the time from receiving a complete claim to communicating a well-supported decision. It introduces an assistant that prepares a case summary and assembles relevant information for an adjuster.

In the initial design, the assistant produces a useful summary, but the adjuster still has to locate the applicable policy, resolve missing evidence, and join the same review queue. Measuring summary preparation alone would miss much of the work determining the claimant’s experience.

A broader design could change the following relationships:

Operating element	AI added to the existing arrangement	Proposed redesign
Case preparation	The summary arrives separately from the evidence needed to decide.	The summary travels with source references, the applicable policy version, and clearly identified information gaps.
Decision and review	Every case follows the existing approval path, regardless of what the assistant has changed.	Authorized adjusters retain coverage and payment decisions. Review routes reflect case complexity, established authority, and unresolved evidence.
Exceptions and people	Employees absorb verification and correction work alongside their existing workload.	Review capacity, training, and specialist support are planned explicitly; employees can challenge outputs and escalate concerns.
Ownership and improvement	Each team measures its own activity; recurring corrections stay in individual case files.	A claims outcome owner coordinates the workflow, while named policy, quality, and technology owners resolve recurring problems.

Illustrative design comparison. Any improvement would need to be demonstrated in the insurer’s own operating environment.

The proposed benefit is a shorter path to a sound decision with less avoidable rework. Test that proposition against the current baseline using complete-case turnaround time, decision quality, reopened cases, and total handling effort, including verification and exceptions. Keep case complexity comparable so a shift toward easier cases does not masquerade as improvement.

If preparation becomes faster while correction work or decision errors increase, the design needs revision. If knowledge cannot be made reliable enough, narrow the assistant’s role. The operating model creates conditions for success; the capability must still prove it can perform within them.

Give the outcome owner authority across the workflow

The claims illustration exposes a leadership issue: the team introducing AI may be unable to change the conditions that determine its value.

Operations controls staffing. A policy owner controls authoritative guidance. Risk and quality functions shape review requirements. Technology operates the capability. If each protects its own targets without a shared outcome, the necessary changes can remain unresolved between them.

I would give a named Outcome owner responsibility for the complete outcome, supported by explicit decisions about what that owner can change and what requires executive resolution. Specialists retain their responsibilities. The outcome owner needs a way to reconcile dependencies, secure resources, and escalate conflicting priorities.

Funding should cover those dependencies. Research on the productivity J-curve explains why general-purpose technologies depend on complementary investments, including changes to processes and human capability. It supports treating organizational work as a substantive part of technology investment; it does not establish a guaranteed return or timetable for an individual AI initiative. Brynjolfsson, Rock, and Syverson, The Productivity J-Curve, January 2021.

The workforce commitment also needs to be explicit. Removing routine work can leave a role concentrated on difficult cases. Leaders should determine how people will develop that capability, how workloads will change, and how released time will be used. If staffing changes are contemplated, communicate them honestly. A promise of more meaningful work needs an actual role, adequate support, and credible expectations behind it.

Match the redesign to the outcome

Reinvention can start with one defined service or workflow. It need not begin with a company-wide reorganization.

A low-risk internal drafting assistant may fit within existing responsibilities once appropriate use and review are established. A service that changes who decides, which approvals remain, or how work crosses departments requires broader operating design. The scope should include the dependencies capable of preventing the intended result.

Some constraints should remain. A necessary independent review is a design requirement. Leaders can improve the evidence reaching that reviewer or adjust capacity without eliminating the safeguard. Where the cost of reliable operation exceeds the expected benefit, keeping the capability in a narrower assistive role may be the better decision.

As the work changes, use performance evidence and employee feedback to revisit the arrangement. A recurring correction may call for better knowledge, a different routing rule, additional training, or less delegation. The response should follow the cause.

This is the design question within Phase 2, Design the Operating Model, of the Transformation Operating Map. The Digital Operating Model Framework provides the reusable structure for examining these connected choices.