By Ganesh Padmanabhan
Two stories dominate the AI discourse right now. One side points to The Information’s tally of $18.5 billion in annualized revenue from “AI-native” apps as proof the AI economy is alive and growing. The other cites MIT NANDA research that 95% of enterprise GenAI pilots fail to deliver meaningful returns.
Both are correct — and both miss the point.
The question isn’t whether AI makes money for vendors or whether most experiments fizzle. The question that matters to operators is simpler: value for whom?
Revenue does not equal value
Rising revenue tells us only that enterprises are paying. It doesn’t tell us they are winning. In fact, several structural issues explain why vendor success doesn’t always translate into customer value.
The economics are fundamentally misaligned. Open-ended inference costs and orchestration overhead eat into ROI when usage is ungoverned or detached from measurable outcomes. Meanwhile, workflow mismatch creates theater instead of efficiency. Chat feels magical, but when it sits beside the work instead of inside it, you’re not actually solving anything.
What we’re seeing is concentrated capture, where a few vendors monetize enthusiasm while customers absorb integration cost, change management and compliance risk. Too many deployments target shallow problems because demos highlight what’s easy to show — general Q&A — not what’s worth fixing, like bottlenecks with dollars attached.
Brass tacks: Revenue is a lagging indicator of willingness to pay, not proof of realized value.
‘95% failure’ is a diagnosis, not a verdict
A high failure rate doesn’t prove the technology cannot deliver. It shows most deployments lack the conditions for impact. Teams start with a model, not a business leak — denials, rework, overtime, backlog. Without pre/post comparisons on turnaround time, first-pass yield, or error rates, success becomes largely unmeasurable.
Most critically, these failed pilots never embed AI into actual workflows. Agents outside the workflow remain optional and invisible. Add to that the complete absence of governance — no explainability, provenance, data traceability or audit hooks — and wins can’t be trusted or scaled.
The problem isn’t that AI doesn’t work. It’s that we keep applying it to the wrong problems with the wrong instrumentation.
Signal vs. noise: The scoreboard that matters
If you want to know whether AI creates value, measure it where value shows up:
Cost to serve per unit of work;
Cycle time from intake to decision;
First-pass yield (work completed without rework);
Error and exception rates;
Throughput per reviewer or agent;
Compliance outcomes (audit exceptions, avoidable penalties); and
Customer-level metrics in healthcare, quality measures, appeals/overturn rates and member experience
If these metrics improve at scale, the AI is working. If not, revenue is just noise.
A field guide for the 5% who win
Enterprises that succeed follow a different playbook entirely. They start with value pools, not models, and they define knife-edge metrics from day one: baseline turnaround, touches, first-pass yield, dollars per case.
Instead of building chat interfaces, they design for decisions and embed agents directly in the system of record.
The winners constrain scope ruthlessly and instrument everything. They keep human-in-the-loop by design and own their context layer completely — policies, SOPs and historical decisions become competitive assets.
Pricing gets tied to outcomes, not tokens, and they A/B test at the workflow level, not by user.
Perhaps most importantly, they plan for the boring parts — identity, access, PHI handling, retention policies. Only when all the infrastructure works do they think about scaling, and only when metrics hold steady across expanded use cases.
This isn’t sexy work. But it’s how the 5% of those who succeed compound advantage.
What real value looks like
At one major health insurer, a prior-authorization agent now reads charts, applies clinical criteria, cites sources and drafts determinations inside the same system nurses already use. Review times dropped from 35 minutes to under 15. The breakthrough wasn’t that “it talks.” It was that it shortened the path to a defensible decision, while keeping humans in the loop.
At a provider group, replacing free-form Q&A with an agent that assembles cite-and-explain chart summaries raised first-pass yield for audits. The win wasn’t a bigger model. It was a tighter, trustworthy loop.
The takeaway
The revenue boom proves enterprises are willing to pay. The failure statistics reveal that most are solving the wrong problems. Leaders who succeed aren’t chasing demos or headlines. They’re embedding AI where money is won or lost, instrumenting outcomes and scaling only when the numbers hold.
Debates about totals and failure rates make good copy. But value is local, operational and measurable. If your AI program can’t show impact on cost, speed, quality and compliance inside the workflow, you don’t have an AI strategy. You have a press release.
Ganesh Padmanabhan is the CEO and co-founder of Autonomize AI, a pioneering healthtech company transforming how healthcare organizations unlock value from unstructured data. Autonomize AI uses specialized AI agents to simplify healthcare’s most complex processes, transforming them into streamlined, AI-native operations. The company’s solutions reduce administrative bottlenecks across utilization management, care management, contracts, claims, and payments—blending automation with human expertise to deliver faster decisions, lower costs, and better patient experiences.
Related Crunchbase query:
Illustration: Dom Guzman
Stay up to date with recent funding rounds, acquisitions, and more with the
Crunchbase Daily.