Where AI Actually Delivers ROI (and Where It Doesn't)

Not all AI use cases have the same return profile. Some categories deliver measurable, recurring value within months. Others have consistently disappointed regardless of implementation quality. Knowing which is which before you invest changes the math.

The failure to distinguish between use cases with strong and weak return profiles is one of the most expensive mistakes in enterprise AI portfolio management. Organizations that selected use cases based on enthusiasm, vendor recommendations, or peer benchmarks rather than return profile analysis have portfolios heavy with investments that cannot produce the returns they were approved on.

Use case selection is not permanent. Organizations that recognize they are invested in low-return use cases can reallocate. But the cost of being in the wrong use cases for two to three years, including the opportunity cost of not being in better ones, is significant. Getting use case selection right in the first instance matters more than it might appear.

The High-Return Categories

Several categories have produced measurable, recurring AI ROI across multiple enterprise contexts. The common thread is not the specific AI technology used. It is a set of structural characteristics that make value both large and measurable.

The Low-Return Categories

The use cases with the highest AI ROI share a common structure: high volume of repetitive work, an objective quality metric, a measurable baseline, and a human review layer that catches failures before they propagate to customers or decision-makers.

The Domain Factor

Domain context changes the return profile of every use case category. The two most significant domain factors in enterprise AI are regulatory complexity and data sensitivity.

Healthcare: Regulatory requirements under HIPAA, FDA guidance on Software as a Medical Device, and clinical accountability standards all add compliance cost to every AI use case. The MEDFIT-LLM study (Rao, Jaggi, Sonam Naidu, IEEE RMKMATE 2025, DOI: 10.1109/RMKMATE64574.2025.11042816) demonstrated that fine-tuning small language models on domain-specific healthcare data using LoRA produces meaningfully better performance for clinical applications than using general-purpose models. This finding has a direct ROI implication: domain-specific fine-tuning reduces the error rate in clinical AI applications, which reduces the rework cost and the compliance risk. The investment in domain-specific model preparation is a cost that does not appear in generic AI ROI frameworks but that significantly affects the actual return in healthcare contexts.

Legal and financial services: The EU AI Act (Regulation 2024/1689) classifies AI systems used in certain legal and financial contexts as high-risk, requiring conformity assessments, technical documentation, and human oversight requirements that add cost to every deployment. The compliance overhead changes the minimum volume threshold required for any use case to produce positive ROI.

The Volume Threshold

AI ROI is inherently a compounding return story. The value of reducing cost per unit by a fixed percentage scales linearly with volume. The cost of the harness infrastructure required to make AI reliable in enterprise contexts is largely fixed. This means that AI ROI requires sufficient volume for the fixed harness cost to be amortized against a large enough base of unit cost savings.

The practical implication is that low-volume use cases rarely justify the full enterprise AI investment. A process that handles 200 transactions per month may produce demonstrable AI productivity gains per transaction while never recovering the engineering, evaluation, and maintenance costs of a reliable enterprise deployment. The same process at 20,000 transactions per month has a fundamentally different math.

A Use Case Scoring Framework

Before committing to any AI use case, scoring it on four dimensions identifies whether it belongs in the high-priority portfolio:

Applying the Framework: A Portfolio Audit Approach

For organizations with existing AI deployments, the use case scoring framework described above can serve as a portfolio audit tool. Walking through each active use case and scoring it on volume, measurability, baseline quality, and regulatory complexity often reveals a portfolio that is overweighted in low-return use cases and underweighted in the categories that reliably produce value.

The audit process itself is valuable beyond the scores it produces. It forces a conversation about each use case's value hypothesis that many organizations have not had since the initial deployment decision. Use cases that cannot generate a clear answer to "what specific outcome are we expecting AI to change, and how would we know if it is changing?" are candidates for redesign or exit regardless of their scores on the four dimensions.

McKinsey's State of AI in 2024 found that AI high performers allocate their AI investments more selectively than lower performers, concentrating on use cases where both business value and technical feasibility are high. The use case scoring framework described here is one approach to operationalizing that selective allocation discipline. Organizations that apply it before committing to new use cases, rather than after, build portfolios with better return profiles from the outset.

The WEF Future of Jobs Report 2025 highlighted that the highest AI productivity gains have appeared in tasks characterized by high repetition, clear quality criteria, and high baseline time investment. This finding from the workforce analysis perspective is consistent with the use case scoring framework from the business investment perspective: they independently arrive at the same conclusion about which types of work respond best to AI augmentation.

What the Research Supports About Use Case Selection

The research base on enterprise AI ROI is still developing, but the patterns visible in the available evidence are consistent with the use case scoring framework described above. Noy and Zhang (Science, 2023) demonstrated that AI productivity gains were largest for specific tasks where worker performance was currently constrained by skill or speed rather than by information access or judgment quality. That finding directly supports the scoring framework dimension on baseline quality: use cases where current performance reflects skill constraints have more headroom for AI improvement than those where current performance is already expert-level.

The Anthropic Economic Index (2025) documented that AI use is concentrated in computer and mathematical occupations, with document and information management as the most common use patterns. The concentration reflects revealed preference: workers who choose to use AI intensively are gravitating toward the task types where it demonstrably helps them. That revealed preference is consistent with the high-return use case categories identified above, since software development and document processing are both high-volume, high-measurability contexts that fit the scoring framework's strong profile.

The MEDFIT-LLM research (Rao, Jaggi, Sonam Naidu, IEEE RMKMATE 2025) demonstrates that domain context affects use case performance in ways that matter for portfolio management. The finding that domain-specific fine-tuning using LoRA materially improves performance for clinical AI applications has a direct analog in enterprise contexts: use cases in specialized domains with high accuracy requirements benefit from domain-specific investment that general-purpose AI does not provide. Portfolio managers who treat healthcare AI, legal AI, and general enterprise AI as interchangeable are making a category error. The return profiles differ because the performance requirements differ, and the investment required to meet those requirements differs accordingly.

The practical implication for enterprise AI portfolio management is that use case selection is not a one-time exercise. Portfolios should be reviewed at least annually against the scoring framework, with use cases that no longer meet minimum thresholds on volume or measurability considered candidates for exit. The budget freed by exiting underperforming use cases is better deployed in high-return categories where the framework's conditions are met. This active portfolio management approach is what distinguishes AI programs that improve over time from those that maintain the same profile of mixed performance indefinitely.