Normatic.Discuss your AI transformation
العربية
← All insights
Perspective

AI Transformation Happens at the Workflow Layer

AI changes organizational performance when its capabilities reshape how an outcome is produced. The workflow connects individual task improvements to business value, human judgment, and the operating model required to sustain change.

An organization can give every employee access to AI and still leave its underlying way of working largely intact. Reports may be written faster, documents summarized more easily, and analyses produced with less effort. Yet decisions can remain slow, information fragmented, and customers waiting for an outcome that no individual team fully owns.

The distinction matters because improving the production of information is only part of improving the performance of an organization. Performance also depends on what happens to that information: who trusts it, who acts on it, which systems must change, and how responsibility passes between people.

Normatic's view is that the workflow is the practical unit of AI transformation. It connects a business intention to an observable result through the interaction of information, decisions, people, and systems. Redesigning that interaction is how an organization turns a new technical capability into a different way of operating.

This does not diminish the importance of models. Their capabilities and limitations determine what is feasible. It places model capability within the larger system that must produce the outcome.

From access to organizational change

Access to useful AI capabilities has become substantially less expensive. Stanford's 2025 AI Index documented a more than 280-fold decline, between November 2022 and October 2024, in the inference cost of reaching roughly GPT-3.5-level performance on the MMLU benchmark. That is a comparison at a particular capability threshold, not a measure of the total cost of an enterprise deployment. It nevertheless illustrates how quickly access to a given level of model capability can spread. Stanford AI Index, 2025

As capable models become more widely available, access alone offers a less durable basis for differentiation. Our interpretation is that more of the advantage will depend on the organization's ability to combine those models with its knowledge, operating relationships, and capacity to change. Model selection remains consequential: reliability, cost, latency, and deployment constraints can make one solution viable and another unsuitable. But selecting a model does not settle how the business should work.

Four different kinds of change are often grouped together under AI adoption. Model adoption gives people access to a capability. Task augmentation improves an activity, such as drafting a response or extracting information. Workflow transformation redesigns the connected activities through which an outcome is produced, potentially removing tasks and handoffs altogether. Operating-model transformation changes the responsibilities, decision rights, capabilities, structures, and management mechanisms that support that work.

These distinctions describe the scope of change, rather than a mandatory maturity sequence. A focused task improvement can be worthwhile on its own. A redesigned workflow may also require decisions about ownership before any system is built. The strategic risk is treating success at one level as evidence that change has occurred at all four.

The gap between a faster task and a better outcome

There is credible evidence that AI assistance can improve performance without wholesale organizational redesign. In their study of 5,172 customer-support agents, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that access to a generative AI assistant increased issues resolved per hour by 15% on average, with substantial differences across workers. This was a deployment within an existing service setting, not simply a laboratory writing exercise. Generative AI at Work

That finding is an important boundary on the argument. Workflow transformation is not a prerequisite for every economic benefit. Where an improved task directly affects the outcome, augmentation can already be valuable. The study does not, however, establish an equivalent increase in total company productivity or prove that benefits will transfer unchanged to other settings.

The difficulty arises when the improved activity is only one element in a longer chain. Consider an illustrative internal approval process. A request that previously took an hour to prepare can now be drafted in minutes. But it still enters a queue, arrives with missing evidence, moves between uncertain owners, and waits for a decision. Faster preparation may reduce effort while barely changing the time to completion. If it increases the volume of submissions, it may even add pressure to the unresolved approval bottleneck.

A similar problem appears when AI makes it inexpensive to produce more analysis. The organization may gain useful insight, or it may create additional material that other people must verify and reconcile. The outcome depends partly on whether the capacity to judge and act develops alongside the capacity to generate.

Time saved is therefore an intermediate measure. It can become greater throughput, better service, more attention to difficult cases, or lower operating cost. Those outcomes require different organizational choices. Adding up minutes saved across employees does not show which benefit has actually occurred.

For leaders, the consequential question is whether local improvements change the performance of the whole flow of work. That requires looking beyond the activity where AI is most visible.

The workflow connects capability to consequence

A workflow has a beginning, a result, and the dependencies that connect them. A customer issue is resolved, a request is fulfilled, or an operational decision is made and enacted. Defining that boundary makes it possible to examine both the value created and the effort, delay, error, and risk incurred along the way.

Return to the approval example. A redesigned workflow could help a requester articulate the need, gather missing evidence from authorized sources, and assemble the context an approver requires. Defined rules could route the request to the appropriate person. The decision would be recorded, permitted actions carried out in connected systems, and the outcome returned to the requester. Ambiguity and policy exceptions would have explicit routes to human judgment.

The change is larger than a faster draft. Information is assembled earlier, responsibility becomes clearer, and the decision is connected to execution. Some coordination work may disappear because the redesigned system no longer requires people to reconstruct context at each handoff. This is an illustrative design, not a Normatic client case or a claim of achieved savings.

Its success would be visible in the completed workflow: time to fulfillment, decision quality, rework, exception backlogs, and the accuracy of executed actions. Model accuracy remains relevant, but it is one contributor to those results. The design may combine AI for interpreting varied requests with conventional software for validation and routing, and people for decisions that require authority or contextual judgment.

This is why integration is part of the transformation itself. A system that prepares a persuasive recommendation but cannot access the right records or carry an authorized decision into the system of record leaves essential work outside its boundary. Data access, permissions, and enterprise connections determine whether the proposed way of working can actually operate.

The same boundary makes adoption more meaningful to observe. Opening an assistant is evidence of use. Completing work through the redesigned process, handling exceptions, and sustaining quality are evidence that the operating practice has changed. If employees routinely revert to email or parallel spreadsheets, the formal design may still be missing something essential.

A workflow also gives accountability a concrete object. Someone can own its performance across handoffs, while specialists retain responsibility for their decisions and controls. Without that connection, each function can improve its own activity while the overall outcome remains disappointing.

Human judgment is part of the design

AI changes where judgment is needed and the information available to support it. It does not remove the need to determine who may act, on what evidence, and with what consequences.

Research also cautions against assuming that capability is uniform. In a preregistered experiment involving 758 Boston Consulting Group professionals, Fabrizio Dell'Acqua and colleagues found that AI assistance improved performance on tasks within the tested model's capabilities but reduced correctness on a complex managerial task outside that boundary. The experiment used GPT-4; it does not establish the limits of today's models. It demonstrates why success on one task cannot establish reliability across an entire workflow. Navigating the Jagged Technological Frontier

Normatic's implication is that human authority and accountability must be designed according to consequence, uncertainty, and demonstrated system reliability. Routine, reversible actions may be appropriate for automation within defined limits. Ambiguous judgments and consequential commitments may require a named person with the authority and information to decide.

Simply inserting a reviewer does not resolve the issue. A person asked to approve an output without its supporting evidence, enough time to examine it, or the ability to challenge it provides weak oversight. Equally, requiring a person to inspect every low-consequence action can consume the capacity needed for decisions where judgment matters most.

Effective oversight therefore depends on the structure of the work: what the system is permitted to do, which conditions trigger escalation, what the reviewer can see, and what happens when confidence or evidence is inadequate. Responsibility remains human even when individual actions are automated.

There is a longer-term capability question as well. If AI assumes much of the routine work through which employees previously learned, the organization must consider how people will develop the expertise needed to interpret exceptions and challenge recommendations. Sustaining human judgment requires deliberate opportunities to build it.

Redesigned workflows change the operating model

Once work changes across functional boundaries, questions about the operating model become unavoidable. Who can change the workflow? Who funds improvements whose costs sit in one function while benefits appear in another? Who maintains its knowledge sources, evaluates performance, and decides when a system's authority should expand or contract?

An organizational structure built around separate stages of work may need a stronger owner for the complete outcome. Specialist teams may spend less time processing routine cases and more time resolving exceptions, improving rules, and maintaining quality. Incentives may need to recognize completed outcomes rather than the volume of intermediate activity. These are possible consequences of redesign, not a prediction that every organization should adopt the same structure.

The economic logic has a precedent. Brynjolfsson, Daniel Rock, and Chad Syverson's work on the productivity J-curve explains how general-purpose technologies can require substantial complementary investment, much of it intangible, before their benefits are fully reflected in measured productivity. Their model concerns a broader economic pattern; it is not proof that any particular AI program will eventually pay off. The Productivity J-Curve

For AI transformation, we see workflow design, operating knowledge, evaluation practices, and organizational capability as important forms of that complementary investment. They must be built alongside the technology, and their value must still be tested against outcomes. Organizational change is part of the investment case, not a reason to suspend scrutiny of it.

This also changes the leadership agenda. AI literacy and access help people discover possibilities. Leadership must connect those discoveries to business priorities and resolve the dependencies that individuals cannot change alone. Funding an assistant while leaving fragmented ownership untouched can limit the return even when the technology performs well.

The resulting advantage can accumulate. An organization that learns from errors, exceptions, and actual use can improve its knowledge sources, controls, and workflow design. A competitor may be able to buy access to the same model. Reproducing the operating knowledge and coordinated practices around it may be harder. This is Normatic's strategic proposition, rather than a result established by the cited studies.

Where transformation becomes real

The workflow sits between the capabilities of a model and the ambitions of an enterprise. It is specific enough to build, evaluate, and measure, and broad enough to reveal the organizational decisions that isolated task improvements can leave unresolved.

For that reason, we expect durable AI advantage to depend increasingly on how well organizations redesign and govern the production of outcomes. The objective is not to automate every activity or remove human judgment. It is to create a more capable way of working, with explicit authority, connected systems, observable performance, and the ability to improve.

AI transformation becomes real when a new capability changes how an outcome is produced—and the organization can take responsibility for sustaining that change.

Sources