%20(1).jpg)
As Labor Day approaches, we are into the last throes of summer. AI-related content over the last week has been amazing. The MIT NANDA study that sent AI stocks tumbling on Friday revealed the data that is making nearly every healthcare, financial services, and HR technology executive lean forward.
MIT reported that 95% of AI pilots fail. After interviewing 150 executives, surveying 350 employees, and analyzing hundreds of individual corporate AI projects, MIT researchers discovered the problem isn't the technology—it refers instead to the last mile. There is a "learning gap” and a context gap. A lack of co-located data and contextual workflow design leads to limitations for deterministic AI, especially when the cost of an edge-case is significant.
While MIT’s are not necessarily outlier findings (e.g. Capgemini found in 2023 that 88% of AI pilots failed to reach production, while S&P Global reported that 42% of generative AI pilots were outright abandoned), the study speaks plainly as to the “why.” Companies are struggling not with AI capability, but with verticalized AI implementation.
Following the MIT study, we decided to share with our 3 hot takes in AI for 2026, 2027, and 2028:
In employee benefits, AI is both a minefield and panacea. We expect that over the next 3 years that well north of 95% of initiatives will fail. Generic large language models might excel at writing marketing copy, but they're hopeless at understanding why Sarah in accounting needs a different health plan than Michael in engineering (context: Sarah goes to a frequent doctor that is not in the secondary network). For all the “AI sizzle,” benefits technology companies are missing the context that Bob is not putting more money into his 401(k) because he recently took out $50,000 in graduate school loans. Then, you need workflow maturity in order to drive actions. Actions when a 1st degree burn has a different outcome than a 2nd degree burn.
At Nayya, we've been quietly building what we call "last-mile AI"—specialized models that don't just process benefits data, but understand the intricate relationships between employee demographics, plan designs, medical claims, carrier networks, and individual financial information. While some stitch together the latest foundation model and a new website, we've focused on the unglamorous but crucial janitorial data work: creating AI that actually knows benefits.
The MIT study's most striking finding might have been that corporations that bought specialized AI solutions succeeded 67% of the time, while those building specialized solutions internally succeeded only 33% of the time. The researchers noted that "[developing] AI models or systems from scratch requires a level of expertise many companies don't have." Ironically, generalized AI companies have that deep technological capabilities but not the last-mile expertise. In benefits technology, this expertise gap is a chasm. The largest context chasm of the HR suite.
Internal builds typically rely on open source or open weight language models, which, as the MIT report notes, "still lag their proprietary rivals. And when it comes to using AI in actual business cases, a 5% difference in reasoning abilities or hallucination rates can result in a substantial difference in outcomes." In benefits, where a wrong recommendation can trigger ERISA violations or cost an employee thousands in out-of-pocket expenses. The edge case isn't just significant—it's potentially catastrophic.
Consider the alternative: an internal IT team spending eighteen months trying to build a benefits actions engine, only to discover it can't distinguish between in-network and out-of-network providers, or worse, inadvertently creates compliance issues with personalized absence management advice. These generalized models cost your organization millions of dollars.
This is where co-located data becomes transformative. When AI has simultaneous access to plan documents, bank account information, claims data, payroll information, and individual employee profiles, it doesn't just provide answers, it provides context at the “last mile” that drives completely transformational experiences.
The companies that recognize the Benefits AI as a specialized discipline—not a generic technology play—will own the next decade. Embracing AI in benefits will be highly specialized, and highly verticalized: due to (1) the collision of a lack of co-located data (401K versus medical versus prescriptions) and (2) the dearth of contextual understanding, not just between employee benefits and other HR functions, but within employee benefits itself - where one program manager may own financial wellness, another may own health plans, etc.
The MIT study showed us that 95% of AI pilots fail because organizations don't understand how to implement the technology. In benefits, the stakes are higher: a 1% failure rate impacts not only employee financial wellness but also organizational risk.
Smart money bets on specialization here - both more broadly and especially within employee benefits. And there is a ton of money to be made. We look forward to sharing more with you over the coming months.
