Matching capabilities to applications: a strategy framework

21/3/2026 ☼ strategy ☼ organisation design ☼ AI ☼ resource allocation ☼ uncertainty ☼ decision-making ☼ frameworks

tl;dr: When you’re matching capabilities to applications — AI tools to roles, research teams to industry sectors, platforms to use cases — the relationship has two properties: how tightly each capability connects to each application, and how many relevant connections exist. These two dimensions produce four quadrants, each requiring a qualitatively different strategy. The people involved (the person overseeing the system, the capability holder, and the application holder) have different jobs depending on which quadrant they’re in. Applying a single mandate uniformly across all quadrants mistakes a slogan for a strategy.

A large technology company decides to go “AI-first.” Leadership is confident: the technology exists, the talent exists, everyone just needs to lean in. From the top, the message is clear.

On the ground, the experience is different. Some roles have obvious automation candidates, such as process-heavy work in accounting or financial operations. Other roles are genuinely confused: FP&A analysts making judgement calls about which variances matter, controllers flagging unusual transactions for review. Years of restructuring and downsizing — and the recent Block layoffs in particular — have made people reluctant to say “I don’t think AI is the right tool for this part of my job.” “I’ll get AI to format my slides” is not a compelling answer to “How will you be using AI in your role?”

This is a diagnostic problem, not a communication problem. The capability (AI tooling) couples differently to different applications (role types and tasks). Leadership assumes the coupling is uniform, but it isn’t.

The logic of coupling between sets of capabilities and sets of applications is a general one. Whether it is research institutes matched to industry sectors, technology platforms matched to use cases, or teams matched to projects, the question is always the same: how tightly does each capability connect to each application, and how many relevant connections exist?

Two dimensions

Two properties define the relationship between any pool of capabilities and any pool of applications.

Coupling tightness describes how specifically a capability connects to an application. Tight coupling: the capability is highly specific to the application and loses most of its value elsewhere. Equivalently, the application can only be served by a narrow set of capabilities and substitution is difficult. Loose coupling: the connection is shallow. The capability does not need heavy adaptation to serve a given application, and the application does not depend critically on that specific capability. Either side could be substituted without major loss. Herbert Simon called this property “near-decomposability”: systems where interactions within subsystems are strong but interactions between them are weak.

Multiplicity of relevancies describes how many capability-application pairings exist. This is bidirectional: an application may need many capabilities to deliver results, and a capability may be relevant to many applications. These are distinct and both matter. Few relevancies: a small number of clear matches. Many relevancies: a large number of potential matches.

These dimensions are independent. A general-purpose capability may have few known applications (loose coupling, few relevancies). A highly specialised capability may be in demand across many domains (tight coupling, many relevancies). Coupling describes the depth of each connection; multiplicity describes the breadth of the field. Karl Weick argued that loosely coupled elements remain responsive to each other while retaining their own identity and separateness. Coupling is not a binary; it is a continuous property, and the interesting cases are those where responsiveness and distinctiveness coexist.

Together they produce four quadrants, each with qualitatively distinct strategic implications. The implications differ depending on where you sit: as the network manager (the person overseeing both capability holders and application holders), the capability holder (the entity that holds resources or capabilities), or the application holder (the entity that represents an application area).

Tight coupling, few relevancies: the lead model

This is the simplest quadrant. A capability connects tightly to one or a small number of applications. The match is clear. The stakes of getting it right are high.

For the network manager, the job is straightforward assignment: put the right capability with the right application and get out of the way. The coordination cost is low but the selection cost is high, because a mismatch here is expensive when both sides are committed. Milgrom and Roberts called this property “complementarity”: activities where doing more of one raises the return to doing more of another, but which tend to be adopted together or not at all.

For the capability holder, nearly all strategic attention goes to maintaining and deepening the specific capability that the matched application requires. Diversification is a distraction.

For the application holder, the priority is securing reliable access to the tightly coupled capability. The application cannot function without it: securing access is existential, not optional. There is no need to demonstrate application value — it is self-evident which capabilities are relevant.

The coupling is tight enough and the relevancies few enough that the capability holder and the application holder are often best served by being the same entity. One individual or team is the natural counterparty to both roles. The coordination cost of separating them exceeds the benefit. If the match is right, this quadrant runs itself.

A verticalised AI startup incubator for medium-sized businesses in traditional industries sits here. Each startup builds an AI application tailored to a specific industry’s existing business processes: a construction company tool that takes updated schedules and site-manager status reports and generates draft materials delivery orders and cashflow forecasts for project managers. The capability couples tightly to the application. Each product must be customised to industry-specific regulation and business-specific legacy systems; it would require major overhaul or complete rebuilding to serve a different industry. The incubator’s job as network manager is matching the right startup teams to the right industry verticals. Get the match wrong and the startup fails. Get it right and the intended outcome is not a unicorn but a small-headcount, stable-cashflow technology company: a “technology Mittelstand.”

Tight coupling, many relevancies: the choice burden

A capability connects tightly to applications. Specificity matters. But there are many potential applications to choose from. This is the quadrant of difficult decisions.

For the network manager, the challenge is prioritising which matches to lock in, accepting that choosing one forecloses others. Many matches are possible, but each demands commitment because the coupling is tight. The network manager must structure processes for capability holders and application holders to evaluate matches together, because neither side has enough information alone. The people with domain knowledge are often not the people with resource authority, and the two groups rarely share a vocabulary for making allocation decisions. Without a shared language, the case for any given match is opaque to those who must approve resources but lack deep domain expertise.

For the capability holder, the strategic question is: which applications to commit limited capabilities to? Every commitment is significant because tight coupling means the capability’s value is concentrated in the match.

For the application holder, the challenge is competing for access to a capability that many others also want. The application holder’s strategic lever is domain knowledge: demonstrating that this particular application will reveal more of the capability’s value than the alternatives.

TSMC sits squarely in this quadrant. Its advanced chip fabrication capability — currently the dominant source of leading-edge fabrication below 5nm — couples tightly to each customer’s chip design: every engagement requires months of co-optimisation between TSMC’s process engineers and the customer’s design team. But the number of customers who need those leading-edge nodes is large and growing: Apple, Nvidia, AMD, Qualcomm, MediaTek, and others all compete for allocation. TSMC cannot serve them all at full capacity simultaneously. Each commitment absorbs fab time, engineering attention, and wafer starts that become unavailable to other customers. When TSMC decided that Apple would get first access to its N3 process node, that decision rested on reported criteria: revenue commitment, design readiness, volume guarantees. Making the selection logic explicit gives all parties a shared basis for evaluation — and makes the reasoning auditable rather than implicit. Without explicit criteria, allocation decisions dissolve into political manoeuvring for fab slots, and customers who lose out have no basis for understanding why or what would change the outcome.

Making selection logic explicit reliably generates pushback. The pushback is evidence of value, not evidence that the exercise is unwelcome. Disagreements that stay implicit do not disappear; they surface as friction, delay, and political manoeuvring downstream. Bringing them into the open while choices are still being made is cheaper than discovering them after commitments are locked in.

Loose coupling, many relevancies: strategic complementarity

A capability connects loosely to many applications. Any individual match is low-stakes: the capability doesn’t lose its value if one application disappears. But the number of potential connections creates a different kind of strategic opportunity.

For the network manager, the priority shifts from choosing the right match to choosing matches that are strategically complementary to each other. No single connection matters much, but the pattern of connections matters enormously. The question is not “which application should this capability serve?” but “what portfolio of connections creates the most value across the network?”

For the capability holder, the risk is spreading too thin: investing lightly in too many applications and building depth in none. The strategic work is maintaining optionality across a broad set of connections while developing enough presence in each to learn which connections are worth deepening.

For the application holder, the risk is accumulating capabilities without combining them. The value in this quadrant comes from synthesis: drawing on multiple loosely coupled capabilities and integrating them in ways that none of the individual capability holders would have conceived alone.

This is, loosely speaking, the Moderna story. When the vast majority of pharmaceutical companies were invested in conventional drug development (small molecules and proteins), Moderna made the largest bet on RNA therapeutics as a platform — following CureVac and BioNTech, but with far more aggressive funding — a platform that coupled loosely to a very large number of potential disease targets, including entirely new framings such as individually tailored therapeutics. RNA was not tightly specific to any single application; it was a general approach that could potentially address many targets. The strategic value came not from any single RNA-disease match but from the breadth of potential matches. Moderna’s competitors’ discomfort with the uncertainty of a loosely coupled, high-multiplicity platform was itself part of what made the position defensible.

Loose coupling, few relevancies: the open frontier

A capability connects loosely to a small number of applications. This quadrant looks strategically undesirable for anyone focused on immediate returns: the connections are few and not especially deep. But this is where future tight couplings and future multiplicities originate. Robert Burgelman distinguished between “induced” strategy (concentrating capabilities on known applications) and “autonomous” strategy (exploring new combinations of capabilities outside the current strategic intent). The loose/few quadrant is where autonomous strategy lives. A capability-application field is not static: positions shift as applications crystallise and connections deepen.

For the network manager, the temptation is to defund or deprioritise this quadrant. That temptation should be resisted. The network manager’s job here is to protect the conditions for exploration without demanding premature commitment.

For the capability holder, the work is exploratory: maintaining the capability and searching for applications, even when the connections are speculative.

For the application holder, the connection is a bet, not a dependency. The challenge is investing attention in a loosely coupled, low-multiplicity capability on the chance that it becomes something that rewrites the field. No one should expect delivery now.

DeepMind’s early work on protein structure prediction illustrates this quadrant. Before AlphaFold produced its breakthrough results in 2020, the application of deep learning to protein folding was loosely coupled — the foundational techniques — attention mechanisms, neural networks — were general-purpose machine learning methods, though significant domain-specific engineering was required to apply them — and the relevancies were few: protein structure prediction was essentially the only serious application in play. Most pharmaceutical and biotech firms saw no reason to invest. DeepMind protected the exploration anyway, without demanding near-term returns. The capability sat in this quadrant for years before the results moved it decisively toward tight coupling and multiplying relevancies as drug discovery, enzyme engineering, and synthetic biology all began to connect.

The sharpest contrast

The two corners where operating logic diverges most are tight/few (upper left) and loose/many (lower right). The difference is not just in what each role does, but in whether the roles can coexist within a single entity at all.

In the tight/few corner, the capability holder and the application holder converge naturally. One entity can hold both roles because the match is self-evident and the coordination overhead of separating them exceeds any benefit. The network manager’s job reduces to selection and protection.

In the loose/many corner, collapsing any two of the three roles into one entity creates tension rather than efficiency. The network manager must think at the portfolio level, assembling complementary sets of connections. The capability holder must choose where to invest depth without overcommitting. The application holder must build coalitions across multiple loosely coupled capabilities. These are genuinely different functions pulling in different directions — portfolio design, investment discipline, and coalition-building cannot be optimised by the same decision-maker without structural conflict.

This is the framework’s core structural claim: the number of distinct roles a situation requires is not fixed. It is determined by the quadrant. Tight/few collapses three roles into one or two. Loose/many demands all three operating independently. A strategy that assumes a fixed number of decision-makers across all quadrants will under-resource coordination where it is most needed and over-resource it where the match runs itself.

Simple rules by quadrant

The three diagrams below show how the framework applies to each role, with distinct strategic imperatives in each quadrant.

Network manager 2x2

Capability holder 2x2

Application holder 2x2

The table below describes different strategic imperatives per quadrant.

	Network manager	Capability holder	Application holder
Tight / Few	Assign and protect the match	Deepen the specific capability	Secure access; it is existential
Tight / Many	Prioritise which matches to lock in	Choose commitments carefully	Demonstrate domain value to compete
Loose / Many	Design the portfolio for complementarity	Maintain optionality; avoid spreading too thin	Synthesise capabilities; integration is the value
Loose / Few	Protect conditions for exploration	Maintain the capability; search for applications	Invest attention as a bet, not a dependency

But the deeper point is not that each quadrant has a fixed rule. It is that the quadrant tells you how to think about what rules should look like when you’re operating in it. What’s needed is not a permanent assignment of control to any single role, but a decision procedure: who has authority over what depends on the nature of the match, the stage of the work, and the quadrant in which the relationship sits. The framework provides the basis for specifying these contingency rules — different quadrants imply different default assignments, which can then be adjusted as circumstances change.

Back to the financial services company. Their error was treating all role-AI pairings as if they sat in a single quadrant. Process-heavy accounting roles sit in tight/few: the AI application is clear and specific. FP&A judgement calls sit closer to loose/few: the coupling between AI and the analyst’s work is shallow and the relevant applications are still emerging. Different quadrants require different strategies. A single mandate applied across the board is not a strategy. It is a refusal to diagnose. And the discomfort people feel when told to “just use AI” is not resistance — it is a diagnostic signal. They sense the inadequacy of a one-size-fits-all model even when they cannot yet articulate why.

---

I’ve spent the last 15 years investigating how organisations can succeed in uncertain times. The Uncertainty Mindset is my book about how to design organisations that thrive in uncertainty and can clearly distinguish it from risk. Part 6 of my book contains detailed descriptions and case studies of how organisations can use desperation by design to build capacity for sustained innovation.

I’ve also been working on tools for learning how to be productively uncomfortable. idk is the first of these tools for productive discomfort.