Cuban is half right. The services that survive AI are the ones AI cannot do without

Mark Cuban says old services are finished and new ones are the opportunity. The part rarely quoted: consistency is unsolved, and human judgment is becoming more valuable. That distinction splits the entire services market in two.

Mark Cuban has been making two claims across spring 2026: at the Dallas Regional Chamber's Convergence AI event, on X, and on the Big Technology Podcast. First: old-model services built on labor arbitrage and document processing are finished. Second: the biggest opportunity is helping companies rebuild around AI. Both got wide coverage. But Cuban also made a third observation that received far less attention. He called consistency the fundamental unsolved problem in enterprise AI (the same question asked twice produces different answers) and argued that human judgment and domain expertise are becoming more valuable, not less.

That third observation is the one that matters most, and it splits the entire services market into two categories that have very different futures.

Substitutive services shrink. Complementary services compound.

The services Cuban says are finished are substitutive, tasks where AI replaces human effort directly. Routing a call. Extracting a form field. Flagging a transaction against a static rule. When the model improves, the human becomes overhead. Cuban is right that this category is compressing fast.

But Cuban's consistency observation points toward a second category he did not name explicitly. Complementary services are the work AI cannot do without: expert evaluation of model outputs, domain reasoning over edge cases, judgment calls that require understanding consequences the model does not weigh. These do not shrink as AI scales. They compound. Every model deployed into a regulated workflow creates new demand for the human expertise that determines whether its output is safe to act on.

The distinction is not theoretical. It is already playing out across healthcare and banking, and it is producing consequences that executive teams are not fully tracking.

What consistency failure looks like when the stakes are clinical

Consider what happens when a probabilistic model runs inside a clinical workflow without domain-expert evaluation between the model and the patient. The Epic Sepsis Model was deployed across hundreds of hospitals. When Wong et al. validated it externally against 38,455 hospitalizations at the University of Michigan, the model's sensitivity was 33%: it missed two out of three sepsis cases. The AUC was 0.63, barely above chance for a production clinical tool.

This is not a story about a bad model. It is a story about what happens when a model moves into production without the complementary layer, domain experts who evaluate whether performance on one patient population transfers to another, who understand that sepsis presents differently in post-surgical patients than in medical admissions. That expertise was not embedded in the deployment pipeline. The model ran. Clinicians experienced alert fatigue. The gap between the benchmark and the bedside went unmonitored.

The same mechanism, different regulated industry

In a health plan, the mechanism produces a different exposure. A 2024 U.S. Senate Permanent Subcommittee on Investigations report examined what happened after UnitedHealthcare automated post-acute care authorization using NaviHealth's nH Predict algorithm. The denial rate surged from 10.9% in 2020 to 22.7% in 2022. For skilled nursing admissions specifically, denials rose from 1.4% in 2019 to 12.6% by 2022.

The algorithm estimated days of care needed based on patterns from similar patients, a substitutive function. What it did not do was weigh the clinical particulars of an individual patient's recovery, the kind of judgment a care-coordination specialist with geriatric expertise would apply. The complementary layer, expert evaluation that catches where the pattern does not fit the patient, was not structurally embedded in the workflow.

In banking, the consistency problem takes a different form. Pindrop's 2025 Voice Intelligence and Security Report documented a 1,300% surge in deepfake fraud attempts, from roughly one per month to seven per day. Synthetic voice attacks against banking contact centers increased 149% year over year. The substitutive layer, automated voice authentication, increasingly cannot distinguish real from synthetic. The complementary layer is fraud-domain expertise that trains detection models on adversarial samples, evaluates false-positive rates against loss data, and updates decision boundaries as the threat evolves.

The data layer is the canonical complementary market

Cuban's observation about consistency is ultimately an observation about data. Generative AI systems are probabilistic: the same input produces variable outputs because the model samples from a distribution rather than following a deterministic path. That variability is a feature in creative applications and a liability in regulated ones. The only thing that constrains it in production is the data layer beneath the model: training data, evaluation data, preference data that shapes how the model ranks its own outputs.

A 2026 arXiv study examining behavioral variance across 50 identical runs of Claude 4.5 Sonnet, GPT-5, and Llama-3.1-70B found that even the most consistent model failed through consistent wrong interpretation 71% of the time: the same incorrect assumption, every run. Consistency without correctness. The researchers concluded that interpretation of accuracy matters more than execution consistency for production deployment.

If the bottleneck is interpretation (correctly understanding what a clinical note means, what a claims adjudication requires), then the constraint is the data that teaches the model to interpret correctly. Building that data requires domain experts who understand the consequences Cuban says AI does not weigh: a nurse who knows what a post-surgical recovery trajectory looks like.

Gartner projects that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. In a separate 2025 Gartner survey of 782 I&O leaders, only 28% of AI use cases fully met ROI expectations.

The part that is not inside the building

Here is where Cuban's framing leaves a gap. He describes the opportunity to help companies rebuild around AI. True. But the specific capability of production AI needs most (domain-expert data infrastructure for training, evaluation, and preference alignment) is rarely inside the organization deploying the model. A health system has clinicians. It does not have nurse-annotators trained to evaluate sepsis-model outputs against clinical ground truth across six patient populations.

That expertise is the complementary services market, the one that grows every time a new model enters production. Every deployment creates evaluation demand. Every edge case creates annotation demand. Every regulatory requirement creates domain-reasoning demand. The substitutive layer compresses. The complementary layer compounds. Cuban identified the mechanism. The data layer is where it lives.