TBM 432: Bundling & Unbundling Capabilities (and AI)

Early in my product career, I wore many hats. I conducted research, analyzed data, sketched interfaces and wireframes, wrote requirements, coordinated releases, answered support questions, and even, shudder, wrote some code. Some of that was empowering. Some of it was amateur hour. But it was all bundled into the loose idea of “product management.”

Over time, those capabilities separated into dedicated roles and disciplines.

Designers, analysts, product marketers, program managers, and product-operations teams developed deeper expertise than any one generalist could reasonably maintain. Some joined cross-functional teams and became part of the daily work. Others became centralized experts that teams accessed through requests, reviews, and shared services. Repeatable parts of their expertise were then encoded into analytics platforms, design systems, research repositories, templates, standards, and workflows.

A role is only a temporary bundle inside the system. Capabilities themselves are distributed across people, teams, routines, tools, platforms, institutions, and external partners. When technology changes work, it rarely swaps cleanly for one person or role. Instead, capabilities move through several different motions:

• Specialization: broad work splits into deeper expertise.

• Diffusion: expertise spreads to more people.

• Centralization: expertise concentrates somewhere else.

• Integration: separate capabilities are coordinated.

• Embedding: expertise moves into tools and routines.

• Rebundling: separated capabilities come back together.

• Externalization: capability moves outside the organization.

• Elimination: work genuinely disappears.

• Loss: work disappears before its underlying capability has been reproduced elsewhere.

The Headlines

Scan the recent headlines and work appears to be changing in every direction at once.

Product managers are becoming “AI builders.”

Designers are becoming “designer engineers.”

Software engineers are becoming managers of AI-generated code.

Data platforms are absorbing parts of data engineering, analysis, and model development.

Middle managers are becoming “player-coaches” and “org leads.”

Product titles are collapsing into a single category: “builder.”

Roles are simultaneously broadening, disappearing, becoming more specialized, and merging with adjacent roles. There are several different motions happening at once. You have all the patterns I described above.

• Technical and analytical skills are diffusing across role boundaries.

• Product, design, engineering, and management tasks are being rebundled into broader roles.

• Knowledge is being embedded in AI tools and externalized to platforms.

• New specialties are appearing around architecture, evaluation, orchestration, reliability, and governance.

• Some repetitive work is genuinely going away. And some organizations (though they may not realize it yet) are losing the apprenticeship, coordination, context, and judgment that the old work helped sustain.

The headlines are observing different capabilities moving in different directions and collapsing that story into catchy headlines.

Capability Moves Through Loops, Not Stages

It is easy to imagine the motions in the design example as a linear journey, or a single, repeating loop. In practice, there are many loops occurring simultaneously, and each motion is creating pressure and tension for other motions.

Each motion changes the conditions around a capability and creates pressure for another motion. Specialization produces depth but ALSO handoffs. Diffusion creates access can ALSO eventually produce inconsistency. Centralization produces consistency but pulls decision-making away from the people closest to the work (or context away from centralized capability). Moving expertise into tools so people don’t have to perform the work directly (embedding) can weaken the learning pathways through which they develop judgment.

These loops do not occur in a fixed sequence. Several may operate simultaneously, at different levels and over different periods of time. A capability might be:

1. Diffusing across teams while control over its infrastructure centralizes;

2. Embedded in technology while new specialist work emerges at its frontier;

3. Externally supplied while internal capability is rebuilt around direction and assessment;

4. Rebundled into broader roles while its underlying knowledge becomes increasingly specialized;

5. Made more productive while the resulting increase in demand creates more work overall.

Each loop describes a recurring tension that may produce useful adaptation, harmful instability, or both.

What Actually Changes With AI?

I often ask myself, “Does AI change the physics of this pattern?” That is difficult to answer because a system can continue to obey the same underlying principles while the variables change so dramatically that those principles appear in entirely new forms.

There is already evidence that AI is affecting specialization, diffusion, centralization, integration, embedding, externalization, convergence, and loss. The real questions are to what extent, under what conditions, and with what second- and third-order effects. For example:

Diffusion

To what extent will AI allow experts to distribute their expertise more easily, without being constrained by the rigid interfaces we rely on today?

AI can make codified expertise available to more people, but we do not yet know how well it transfers tacit, contextual, or embodied expertise. Polanyi (1966) put it simply: we know more than we can tell. If that remains true, diffusion through AI will have a ceiling that is difficult to see from outside the expertise.

Integration

Can it actually lower the cost of specialists working with generalists and with each other?

The translation, handoff, and queueing burden created by specialization is real, and AI can friction-bust. But it hasn’t eliminated the “social” work of coordination, negotiation, and shared judgment.

Convergence and Rebundling

There’s a version of the future where individuals and teams recover capabilities that previously required several dedicated roles. A new kind of higher-level generalist.

AI does let people perform tasks across more role boundaries. But producing more kinds of artifacts is not the same as possessing the underlying expertise.

Specialization

And where will the opposite happen? AI is already creating new areas of specialization, new forms of technical complexity, and new specialists required to build, evaluate, govern, and adapt these systems. It may reduce demand for some specialist tasks while creating entirely new specialties around context, evaluation, orchestration, governance, and exception handling.

Embedding and Extension

Expertise is increasingly being encoded directly in tools and workflows, allowing people to act competently without possessing the underlying skill themselves. This is not hypothetical. It’s happening now. The question is what happens to the role that remains: it shifts toward direction, assessment, and oversight, which are genuinely different skills.

Centralization

AI may broaden local access to a capability while concentrating control over models, infrastructure, standards, data, and defaults in a small number of firms. More people become locally capable. Organizations become collectively dependent on increasingly centralized systems. Both things are true at once.

Loss

What happens when tasks disappear before the judgment, context, relationships, and learning pathways they sustained have been reproduced elsewhere? An organization may preserve output while losing the mechanisms through which people become capable of evaluating, correcting, and extending that output. The MIT Work of the Future task force (Autor, Mindell, and Reynolds, 2020) warned of precisely this: cutting out middle tasks can destroy workers’ skills, networks, and career pathways even when the immediate output looks intact.

All of these dynamics are in effect. What remains unknown is how far each motion will go, which will dominate in different contexts, and where those capabilities will eventually settle.

Limitations

What’s getting in the way of having honest discussions about AI’s long-term effects? The hype is deafening. Science always has trouble with longer-term effects. We don’t understand each other’s work well enough. And time plays tricks on how we process what is happening.

Hype

The hype is so omnipresent that many organizations are spending more time copying other companies, listening to vendors, and signaling that they are “doing AI” than conducting disciplined experiments around specific problems.

That is one of the defining dynamics of a bubble: spending money and taking visible action become signals in themselves. Each company’s activity becomes evidence that other companies should act, regardless of how much value is actually being produced.

This makes the underlying phenomenon harder to study. We are observing the technology, vendor incentives, executive expectations, labor-market anxiety, competitive signaling, and FOMO interacting at the same time. The hype does not mean the technology lacks value. It means adoption is not clean evidence of value.

Research

The second limitation is not unique to AI. Researching any change that touches individuals, teams, technologies, incentives, workflows, institutions, and markets creates severe limits on the general claims we can make.

A study may show that AI improves performance on a particular task (and population, model, and context). That does not necessarily tell us how a role will change, how a team will reorganize, what new coordination costs will emerge, or whether the initial gains will persist.

Many of the outcomes that matter most are also difficult to observe in short-term studies. Do people continue developing expertise? Will teams become more or less capable? Will organizations become more resilient, or more dependent? Strong evidence about a bounded task can easily be stretched into a much larger claim about jobs, organizations, or society.

Consider something like a codebase. Understanding how AI changes quality, maintainability, and the team’s collective knowledge would require a long-term, deeply contextual study. Long-term ethnographic studies of a codebase are difficult.

Flawed Mental Models

Beyond the hype and the limits of research, we bump up against a common foe: ignorance.

What we do know is that flawed mental models will remain flawed mental models, and they will continue to produce flawed assumptions.

1. If you imagine that making products is akin to operating a factory, you will likely miss the mark when assessing AI’s impact on product development. You will be more likely to assume that producing requirements, designs, code, and releases faster necessarily means producing better products faster.

2. If you miss the fact that data scientists spend a great deal of time helping people frame useful questions, you may assume that natural-language querying eliminates the need for data expertise.

3. If you treat a codebase as a pile of code rather than the cumulative record of decisions, constraints, failures, compromises, and hard-won knowledge contributed by everyone who built it over time, you may assume that generating replacement code is equivalent to replacing the system.

The average patient cannot determine whether a doctor truly understands cardiology, so they rely on licensing, training, certification, hospital privileges, and peer review to make that judgment for them. Inside companies, specialist work is rarely protected by an equivalent system.

Instead, people encounter familiar-looking outputs: a screen, a chart, an interview, a roadmap, a status report. Because the artifact looks understandable, they assume the underlying work is understandable too. The question shifts from “Is this person qualified?” to “Doesn’t this look like something I could do?” AI makes that confusion even easier. It can reproduce the visible artifacts of expertise long before it can reliably reproduce the judgment, context, coordination, and exception handling that made those artifacts valuable.

Time

These limitations share a common theme:

We have not had enough time.

History helps us recognize broad patterns, but not how they will settle around AI. We may be late in the S-curve of the current model architecture, early in the curve of AI adoption, and barely beginning the curve of organizational adaptation. AI is not one curve, but a stack of technical, economic, and social curves unfolding at different speeds.

So asking whether AI is “early” or “late” is too simplistic. The timing, magnitude, and interaction of these curves will determine which capabilities diffuse, which concentrate, what new organizational forms emerge, and how much damage the surrounding bubble dynamics cause along the way.

Conclusion

How do you introduce these ideas safely into your organization?

Carefully.

Raise a measured concern about the erosion of skill and judgment when AI use veers more toward substitution and delegation than augmentation and learning. This concern is already fairly widespread and accepted, so you should not get into too much trouble.

You might ask:

Which routine activities are helping people build the judgment they will later need for difficult situations?

And then:

If those activities disappear, how will that judgment be developed and maintained?

From there, introduce the capability motions and feedback loops. Do not present them as a grand framework that everyone must adopt. Use them as language for describing what people are already seeing.

• Where is expertise diffusing?

• Where is it becoming embedded?

• What is centralizing?

• What is being externalized?

• What has been rebundled into broader roles?

• Where are handoffs shrinking, and where are new ones appearing?

• Which loops seem to be producing the outcomes you want?

The point is not to classify everything perfectly. It is to make more of the system visible. The discussion about AI magic has been especially fruitful for me and several colleagues. Once people stop treating the technology as an independent intelligence arriving to join the workforce, they can begin asking more grounded questions about the system around it:

• What context does it have?

• Whose knowledge has been encoded into it?

• What remains outside its reach?

• Who is capable of assessing its output?

• Who controls the infrastructure and defaults?

• What happens when it reaches an exception?

Good luck! Whatever we can do to encourage richer conversations, grounded in reality, not mired in the weird duality of “humans vs AI”, will help us adapt and learn.