AI Transformation Is a Portfolio Problem, Not a Use-Case Problem
A growing list of AI use cases does not resolve the choices that shape transformation. Leaders need to decide which business outcomes deserve investment, how shared capabilities change the economics, and when evidence warrants scaling, redirecting, or stopping work.
Consider an illustrative leadership team reviewing its AI agenda. The list includes a knowledge assistant, automated document processing, a forecasting tool, a customer-service copilot, and several internal productivity experiments. Each has an enthusiastic sponsor. Several have produced convincing demonstrations. All need some combination of engineering capacity, access to enterprise systems, expert attention, and organizational change.
The list describes possibilities. It does not explain which commitments the organization should make.
For enterprises already generating more ideas than they can deliver, the strategic constraint has shifted. Leadership must decide which outcomes matter enough to pursue, which initiatives belong together, what capabilities should be shared, and what will lose resources when something more valuable emerges.
Normatic's view is that these are portfolio decisions. Individual use cases remain useful for identifying where AI could help. Transformation requires an investment logic that connects them to business outcomes and manages the relationships among them.
The task is to allocate scarce attention, capital, and delivery capacity across a coherent set of transformation bets—and to change those commitments as evidence develops.
A collection of successful pilots can still underperform
Use cases are usually easier to sponsor than transformations. A team can demonstrate that AI extracts information, drafts an answer, or prepares a forecast without settling how the wider business will use it. That makes a use case a useful starting point. It also makes it an incomplete unit of investment.
Local success can conceal an unchanged business constraint. Faster document review may have little effect on a service cycle if the delay lies in obtaining approvals. Better recommendations may not change decisions when the responsible manager lacks authority to act. A capable assistant can remain marginal if employees must duplicate their work in another system.
Fragmentation adds a second cost. Separate teams may procure similar tools, build overlapping connectors, maintain competing knowledge sources, and evaluate performance differently. Some variation is justified by the work. Unexamined duplication consumes capacity and creates systems whose support and governance costs become visible only after the pilot.
Recent enterprise evidence illustrates the gap between individual benefits and organizational outcomes. McKinsey's August 2026 survey reports that 80 percent of respondents perceive improved individual productivity from AI, while 37 percent report positive enterprise-level EBIT impact. These are different self-reported measures, not a conversion rate or proof that portfolio design causes the gap. They do show why local productivity cannot stand in for an enterprise investment case. The state of AI in 2026.
Pilot accumulation can then become a resource-allocation problem. Initiatives stay active because nobody has explicitly decided whether they deserve production investment, should be combined with other work, or should stop. A politically prominent demonstration may continue receiving specialists while a less visible operational opportunity waits.
The cost is more than the pilot budget. Every continuing initiative makes a claim on future integration, business ownership, adoption, support, and managerial attention. A portfolio that records only current project expenditure understates those commitments.
Define the transformation bet before funding the solution
A transformation bet is a reasoned commitment to change a material business outcome through one or more workflows. It brings together the intended result, accountable business ownership, required capabilities, dependencies, investment, and the evidence needed to justify further commitment.
The term acknowledges uncertainty. It does not imply that every idea needs a precise return forecast before exploration can begin. An early investment may buy evidence about feasibility or demand. It should still identify the question being tested and the decision that the evidence will inform.
Consider an illustrative effort to reduce the time required to resolve disputed invoices. Information extraction, contract retrieval, case summarization, and response drafting could each appear as a separate use case. The transformation bet connects them to a business outcome: resolving eligible disputes faster while preserving accuracy, authority, and customer relationships.
That framing changes the investment discussion. The organization needs a business owner able to coordinate the relevant functions, a baseline or a plan to establish one, access to current records, and a credible route into operational use. It must understand where AI helps, where conventional software is sufficient, and where people must exercise judgment. Funding covers the work required to change the outcome, including integration and adoption.
Several capabilities may contribute to that bet without needing to launch together. A first release can test a bounded part of the workflow. The broader outcome remains the reason for investing, rather than an assumed benefit attached to a list of technical features.
The boundary should be large enough to capture material dependencies and small enough for ownership and evidence to remain clear. Bundling every related idea into a single transformation program can obscure weak investments as effectively as treating everything separately.
Compare opportunities without reducing judgment to a score
Leadership needs a common basis for comparison, but a weighted ranking can imply more certainty than the evidence supports. An attractive total can conceal an unresolved control requirement, an unavailable data source, or an owner who cannot commit the people needed for adoption.
Value concerns the business outcome that could change and the assumptions behind that expectation. Feasibility concerns whether an effective solution can be delivered. Readiness concerns whether the organization's data, systems, process owners, and users can support it now. An opportunity may be technically feasible and operationally unready.
Risk concerns the consequences of failure and the conditions required to keep those consequences acceptable. Certain requirements are boundaries, not factors to trade against a higher benefit estimate. Strategic leverage concerns whether the work creates a capability that other credible opportunities can reuse.
Time to evidence adds another distinction. A modest investment that quickly tests a critical assumption may deserve attention before a larger opportunity whose next useful evidence is distant. Speed alone is insufficient: a quick demonstration on convenient examples may reveal little about operational performance.
Dependencies connect these judgments. Two attractive initiatives may compete for the same experts or depend on the same unresolved permission model. Others may become feasible together because one supplies a capability the second needs.
Comparison therefore requires explicit trade-offs. What is the next commitment buying? Which uncertainty will it reduce? What else must be true? What will the organization postpone to make room? The answer is a leadership judgment informed by evidence, not a ranking produced by arithmetic alone.
Shared capabilities change the investment case
An isolated use case often carries the cost of building its own foundations. Knowledge access, identity, permissions, connectors, evaluation, deployment, and monitoring may all be necessary before the business can use the result.
A portfolio view allows leaders to distinguish common capability from workflow-specific work. An evaluation service can provide testing infrastructure while business owners define different acceptance criteria. A connector can expose an enterprise system while each workflow retains its own action permissions. A shared knowledge service can support several applications without granting every user access to the same material.
These distinctions matter economically. The first investment may create a capability that lowers the incremental effort required by subsequent transformations. Later work still needs adaptation, evaluation, integration, and adoption; reuse does not make those costs disappear. Savings depend on actual reuse and must be assessed against the continuing cost of the common service.
Research on the productivity J-curve offers a relevant foundation. Brynjolfsson, Rock, and Syverson explain how general-purpose technologies require complementary intangible investments whose costs and benefits can appear at different times. Their 2021 work concerns productivity measurement and technological complements, not the returns of a particular AI portfolio. It helps explain why the cost of building organizational capability should not be mistaken for evidence that the technology has no value. It does not excuse indefinite spending without evidence. The Productivity J-Curve.
Normatic's interpretation is that common capabilities should have a visible relationship to credible demand. Leaders should be able to name the transformations expected to use them, explain what those users require, and revisit the investment if demand changes. Reuse should be demonstrated through operating consumers, not counted from a list of hypothetical future applications.
The economics must also be reconciled at portfolio level. Shared costs should be counted once in the enterprise total and allocated transparently for individual decisions. Benefits claimed by several initiatives should not be added when they represent the same reduction in work or the same improvement in revenue. Double-counting can make a fragmented portfolio appear more valuable than it is.
Sequence the work around what it enables
The highest-value opportunity is not necessarily the right first investment. A smaller, more tractable workflow may create an integration pattern, establish usable evaluation data, or reveal an access-control issue that determines whether a larger bet can proceed.
In the disputed-invoice example, leadership might first fund a bounded version of case preparation using permission-controlled records. That work could establish document access and evaluation capabilities useful to a later customer-resolution workflow. The second bet would still require its own outcome evidence and authorization; sharing foundations would not constitute approval to scale it.
Exhibit 1. A first transformation can create foundations for a second
The alternative extremes are expensive. Rebuilding everything for each use case duplicates effort. Building a comprehensive platform before testing real demand can delay learning and lock the organization into assumptions that the first workflows would have challenged.
A more disciplined sequence develops the necessary foundation alongside an initial business outcome, then extends it where subsequent demand justifies the investment. Some work can proceed in parallel. Other work should wait because its critical dependency remains unresolved.
Shared foundations introduce concentration as well as leverage. Several bets may rely on the same data source, provider, or small group of specialists. A failure or capacity constraint can then affect much of the portfolio at once. Leadership should examine these common exposures, rather than assume that many initiatives imply diversification.
The portfolio must also respect the business's capacity to absorb change. Launching several transformations into the same team can overwhelm process owners and users even when engineering capacity is available. Sequencing is partly a decision about whose attention the organization can realistically commit.
Different investments require different evidence
A credible portfolio may combine immediate productivity improvements, broader workflow transformation, shared foundations, experiments, and strategic options. Their purposes differ, so the evidence expected from them should differ too.
Exhibit 2. Investment purpose determines what the next commitment must establish
| Investment purpose | What it funds | Evidence relevant to further commitment |
|---|---|---|
| Near-term productivity | A bounded improvement in existing work. | Sustained use, acceptable quality, and a credible use of the capacity released. |
| Workflow transformation | Changes across tasks, systems, and responsibilities to improve a business outcome. | End-to-end performance, adoption, operating ownership, and the cost of sustaining the result. |
| Shared foundation | Capabilities required by several identified transformations. | Fit for the first consumers, demonstrated reuse, service reliability, and credible subsequent demand. |
| Experimentation | A bounded test of an uncertain technical or business assumption. | A result that materially changes the decision to invest, revise, defer, or stop. |
| Strategic option | Limited investment to preserve access to a potentially important future capability. | Continued strategic relevance, a clear exercise trigger, and a bounded cost of keeping the option open. |
These purposes can overlap. They describe investment logic, not separate organizational departments or prescribed budget shares. Source: Normatic analysis.
Established research on organizational learning distinguishes exploring new possibilities from improving established activities. James March's foundational 1991 analysis explains why their different timing and feedback can bias organizations toward what is already familiar. It is a conceptual basis for balancing investment, not a formula for allocating an AI budget. Exploration and Exploitation in Organizational Learning.
The practical implication is to avoid demanding the same kind of proof from every investment. An experiment need not claim a production return before its central uncertainty has been tested. Conversely, a mature deployment should not continue to justify itself with the learning rationale that supported its initial pilot.
Balance also needs limits. Calling an initiative a strategic option does not justify leaving it active indefinitely. Its continuing cost, relevance, and trigger for further investment should remain visible. The right mix depends on the organization's priorities, constraints, and uncertainty; fixed percentages would hide those differences.
A portfolio learns when commitments change
Evidence matters only if it can alter resource allocation. An organization that records disappointing results but keeps every initiative funded has improved its reporting, not its portfolio discipline.
Useful evidence moves beyond demonstrations to representative workflow performance, actual use, business outcomes, integration effort, and continuing operating cost. Normatic's working principles make the business owner, intended outcome, baseline, and path into operational use explicit. They also require human judgment and accountability where consequences warrant them. Portfolio decisions should apply those commitments as investment conditions develop.
Research on disciplined experimentation supports the value of revising commitments. Camuffo and colleagues' 2024 study of 759 firms across four randomized trials found that training in a scientific approach to entrepreneurial decision-making increased idea termination. This is evidence from entrepreneurship, not enterprise AI portfolios. Its relevance is the role of explicit hypotheses and tests in deciding what to stop pursuing. A scientific approach to entrepreneurial decision-making.
An AI experiment can succeed as an investment in learning even when the proposed solution does not proceed. It may show that the business problem is smaller than expected, that conventional automation is adequate, or that the required information cannot support the intended decision. The useful outcome is a better allocation of the next commitment.
Exhibit 3. Evidence should lead to a different resource decision
| Portfolio decision | What the evidence indicates | What changes |
|---|---|---|
| Scale | The outcome and operating requirements hold under conditions relevant to expansion. | Commit additional delivery and operating capacity to a defined expansion. |
| Continue or revise | The thesis remains plausible, but a consequential uncertainty or correctable weakness remains. | Fund a bounded next test or redesign; make the next decision explicit. |
| Merge | Initiatives overlap materially in outcomes or capabilities. | Combine the relevant work, resolve ownership, and remove duplicate claims on resources and benefits. |
| Defer | A necessary dependency, owner, or adoption capacity is unavailable. | Release active delivery capacity and specify what would justify reconsideration. |
| Stop | The thesis is invalidated, requirements cannot be met economically, or better opportunities outweigh further investment. | End the commitment, retain useful learning, and redirect capacity. |
Illustrative decision logic. No single metric determines the outcome; relevant control requirements remain binding. Source: Normatic analysis.
These choices should apply to implementations as well as business ideas. A valuable workflow may warrant a different model, provider, or architecture. An unsuccessful implementation does not necessarily invalidate the outcome being pursued, just as an impressive model result does not establish a valuable business case.
Stopping also requires judgment about sunk costs. Past spending cannot be recovered by continuing weak work. The decision concerns the expected value and cost of the remaining commitment, compared with alternatives, including the consequences of withdrawal. Useful components and lessons can be retained without keeping the original initiative alive.
Portfolio authority and business ownership are distinct
Portfolio ownership concerns where the enterprise commits resources and how competing demands are resolved. Initiative ownership concerns the business outcome of a particular transformation. Shared capability ownership concerns the service on which several initiatives rely. Combining these responsibilities implicitly can leave the organization with a central team accountable for benefits it lacks authority to deliver.
As we argued in The AI Operating Model Is Not an AI Department, specialist coordination must connect to business authority. An AI or transformation office can assemble the portfolio view, expose dependencies, challenge evidence, and coordinate delivery. Executive investment authority still needs to decide which priorities prevail. Business owners need to commit the people, workflow changes, adoption effort, and operating resources required for outcomes.
This does not require central approval for every experiment. Teams can explore within clear limits while making material commitments and dependencies visible. Portfolio leadership should concentrate on choices that cross boundaries: scarce expertise, shared foundations, conflicting priorities, consequential risk, and the allocation of capacity among competing bets.
Reviews should therefore change more than a status report. They should confirm what deserves further investment, identify what no longer does, and make the resulting resource transfers real. Otherwise, portfolio governance adds another layer of coordination without improving allocation.
AI strategy becomes an ongoing investment discipline. Leaders select outcomes, fund the capabilities needed to pursue them, and adjust the portfolio as operational evidence replaces assumptions. The deeper advantage lies in making those connected decisions well—not in maintaining the longest list of use cases.
Sources
Established research on investment, innovation, and learning
- James G. March, Exploration and Exploitation in Organizational Learning, Organization Science, 1991. Foundational conceptual research on allocating resources between learning about new possibilities and improving established activities; not AI-specific evidence.
- Erik Brynjolfsson, Daniel Rock, and Chad Syverson, The Productivity J-Curve: How Intangibles Complement General Purpose Technologies, American Economic Journal: Macroeconomics, 2021. Research on complementary intangible investment and productivity measurement; not an estimate of returns from shared AI infrastructure.
- Arnaldo Camuffo, Alfonso Gambardella, Dario Messinese, Elena Novelli, Enrico Paolucci, and Chiara Spina, A scientific approach to entrepreneurial decision-making: Large-scale replication and extension, Strategic Management Journal, 2024. Experimental evidence from entrepreneurship; used for its relevance to hypothesis testing and termination decisions, with limits on transfer to enterprise AI.
Emerging enterprise AI evidence
- McKinsey, The state of AI in 2026: On the road to ROI, August 25, 2026. Survey of 1,719 respondents in 97 nations, fielded May–June 2026, with country responses weighted by contribution to global GDP. The cited productivity and EBIT measures are self-reported and address different levels of impact; they do not establish causal effects of portfolio management.
Normatic interpretation
The transformation-bet framing, investment implications, illustrative enterprise examples, and exhibits are Normatic analysis. They offer a basis for leadership judgment without prescribing a scoring formula, budget allocation, or universal sequence of investments.