Does the following sound familiar? You ask AI to analyze a 40-page strategy document. Within seconds, it identifies patterns, compares competing arguments, and produces a remarkably useful synthesis. Then you ask it to make one seemingly minor change to a table. It removes a row. Or changes two numbers. Or confidently tells you the change has been made when nothing has changed at all. You try again. It makes the same mistake. That is AI jaggedness: the peculiar reality that AI can outperform humans in some areas while struggling, sometimes headscratchingly, in others.
The strange part isn't that AI makes mistakes. Humans do, too. It is where AI excels and where it fails—and how difficult that can be to anticipate. The same technology that performs work we would consider highly sophisticated can stumble over something that seems comparatively straightforward.
And as AI becomes more agentic, jaggedness becomes more consequential: an error can travel through a sequence of actions before a human even notices it.
For an individual user, jaggedness can be frustrating. Embed the same unevenness into a workflow involving multiple people, systems, decisions, customers, and accountability, and it becomes something else entirely: a work design problem.
AI jaggedness is not simply a matter of AI being better at easy tasks and humans remaining better at difficult ones. The boundary is far less predictable.
Dell'Acqua and colleagues (2026) demonstrated this in their work with 758 consultants at Boston Consulting Group. When consultants used GPT-4 for tasks that fell within its capabilities, they completed more tasks, worked faster, and produced higher-quality results. Yet on a task deliberately selected outside that capability frontier, consultants using AI were less likely to arrive at the correct answer.
The dividing line did not follow a simple progression from easy to difficult. Tasks that appear to require similar levels of intelligence or expertise can fall on different sides of what the authors call the jagged technological frontier. That is precisely what makes the phenomenon so challenging: extraordinary capability here, surprising weakness there, without an obvious line telling us where one ends and the other begins.
There is another distinction that matters for work design: jagged does not mean dynamic. Jagged means uneven.
AI happens to be both. Its capabilities are uneven at any given point in time, and the frontier itself continues to move as the technology advances.
That leaves organizations with two challenges at once. They need to design work around capabilities whose boundaries can be difficult to determine—and they need to be prepared to redesign that work as those boundaries change.
The stakes change considerably when that jagged AI capability becomes part of how an organization gets work done.
Morgan Stanley offers an interesting example. Its financial advisors increasingly work with AI throughout a workflow that includes preparing for client conversations, accessing the firm's knowledge, conducting meetings, capturing and interpreting information, determining next steps, documenting interactions, communicating with clients, and following through.
AI does not simply take over one end of that process.
The AI @ Morgan Stanley Assistant can retrieve and synthesize information from the firm's extensive knowledge base. AI @ Morgan Stanley Debrief can capture client meetings, create notes and summaries, identify action items, draft follow-up communications, and create records for the firm's systems. Advisors review AI-generated outputs, interpret them within the client's broader situation, exercise professional judgment, and determine how to proceed.
AI enters the workflow, humans interact with its output, and AI may enter again further downstream.
The human-AI interface isn't a line. It runs through the workflow.
That makes the question of who does what considerably more complex than simply separating “AI tasks” from “human tasks.” Where is AI sufficiently reliable to act? Where does a human need to review its output? Where is contextual knowledge indispensable? Where does professional judgment enter? When should a human intervene? And who ultimately owns the decision?
Morgan Stanley's approach offers another important clue. The firm has evaluated its AI against real-world advisor use cases, incorporated expert assessment of its outputs, and continued to expand those evaluations as the technology and its applications evolved.
That matters because the jagged frontier cannot simply be assumed. Organizations have to learn where it lies in their own work.
And once they do, the challenge becomes one of design: creating workflows that make clear where AI adds capability, where human expertise matters, where the two need to interact, and where safeguards, handoffs, and decision rights belong. In other words: AI is jagged. Work design shouldn't be.
If AI capability is uneven, inserting AI into an existing process is not work design. Neither is compiling a list of tasks that AI can perform.
Start with the work itself.
Map how value is actually created from beginning to end. Where does information enter? Where is it transformed? Where are judgments made? Where do decisions happen? Where does work move from one person, team, or system to another? Where do delays, duplication, or recurring problems already exist?
Then add AI to that picture.
At each point in the workflow, the question is not simply whether AI can perform an activity. It is whether AI can perform it reliably enough, in this context, with these inputs and these consequences. Where the answer is uncertain, the workflow needs to make that uncertainty visible rather than quietly passing it downstream.
That may mean human verification at one point, an automated check at another, a clear escalation trigger somewhere else, or deliberately keeping a decision with a person.
With agentic AI, those intervention points become even more important. If an agent can move through several actions autonomously, organizations need to determine not only what it is allowed to do, but where it must stop, check, escalate, or hand the work back.
The objective is not to put a human into every loop. That would defeat much of the value AI can create. Nor is it to remove humans from as many loops as possible.
The objective is to design the right loops.
There is a second part of the equation that is easily missed. When AI changes one part of a workflow, the human work around it changes as well.
If AI produces the first analysis, human value may shift toward questioning assumptions, recognizing context, comparing alternatives, or making the decision. If AI captures and structures information automatically, people may have more capacity for the conversation from which that information comes. If an agent executes several steps independently, a role may require less execution and more monitoring, exception handling, judgment, or intervention.
That changes more than individual tasks. It can change what people need to know, how roles are defined, which capabilities need to be developed, who collaborates with whom, and where decision rights belong.
AI can also create new and more work.
Verification takes time. Monitoring takes attention. Exceptions need somewhere to go. Someone needs to understand enough about the underlying work to recognize when an AI output that looks perfectly plausible is wrong. And as AI performs more of the work that once required human expertise, organizations need to consider how that expertise will be built and maintained.
This is why AI-enabled work design cannot stop at productivity.
Every change in what AI does creates a question about what humans need to do differently—and what the organization around them needs to enable.
The result may be different roles, different collaboration patterns, different information flows, different spans of responsibility, or different decision rights. At some point, what begins as workflow design becomes organization design.
There is one more complication: the workflow we design today may no longer be the right one tomorrow.
AI learns and advances. Capabilities that sit outside the frontier today may move inside it. Tasks that require verification now may become highly reliable. New capabilities may emerge that change not just one activity, but several steps around it.
Humans change, too.
People learn how to work with AI. They become better at asking questions, evaluating outputs, recognizing weaknesses, and using the technology in ways we may not have anticipated when a workflow was first designed. At the same time, capabilities can erode when people no longer perform the work through which expertise was once developed.
The human-AI interface is therefore not something organizations can design once and consider finished.
This is where adaptive work design becomes essential.
A well-designed workflow should provide clarity about where AI acts, where humans intervene, how work moves between them, and where decisions and accountability sit. But those boundaries must be revisited as capabilities on both sides change.
Clear does not have to mean static.
In fact, the challenge is to create work designs that are clear enough to perform reliably today and adaptive enough to be reconfigured tomorrow.
AI jaggedness will not disappear simply because the technology becomes more powerful. The frontier will move, but the challenge of understanding where AI performs reliably, where human capability matters, and how the two should interact will continue to evolve with it.
The organizations that benefit most from AI may therefore not be those that automate the greatest number of tasks or deploy the greatest number of agents.
They may be those that become exceptionally good at seeing the human-AI interface, designing it deliberately, learning from what happens there, and redesigning it as human-AI capabilities evolve.
That’s why AI’s jaggedness may be unavoidable. Jagged work design isn’t.
References
Dell’Acqua, F., McFowland, E., III, Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423.
Morgan Stanley. (2024, June 26). Morgan Stanley Wealth Management announces latest game-changing addition to suite of GenAI tools.