Why Most Healthcare AI Projects Fail Before They Scale

Most healthcare AI projects never make it past the pilot stage because they launch without product-market fit, the wrong organizational timing, or weak governance, not because the technology fails. That is the diagnosis from Irena Kaplan, a data scientist with a decade of experience building and scaling AI inside some of the largest health insurance companies, now consulting startup founders on the same discipline. Her framework lines up closely with the gap Engini was built to close: the execution layer between a sound AI strategy and a compliant, working system.

RAND Corporation research has found that more than 80% of AI projects fail, roughly twice the failure rate of non-AI IT projects. Kaplan's explanation for that gap, drawn directly from her work inside healthcare and insurance, is less about model accuracy and almost entirely about discipline: research before building, timing the rollout to the organization, and governance strong enough that compliance and operations teams actually trust the system.

Why Do Most AI Projects Never Make It Past the Pilot Stage?

Most AI projects stall after the pilot because teams build the tool before confirming anyone actually needs it solved that way, a pattern Kaplan calls the tool-first approach. The fix is not a better model. It is research before the build starts.

"Our AI strategy should not be for the sake of AI, it should be for the sake of our customers. ... When you're doing this tool-first approach, you sometimes go to the market and you realize that you've built a great product, but there is nobody who's willing to pay for this." - Irena Kaplan

This is the same principle behind Engini's approach to enterprise automation. The platform is scoped against documented operational bottlenecks that already exist inside a finance or compliance workflow, rather than introduced as a generic capability that teams then have to find a use for.

What Does Real Product-Market Fit Look Like for an Enterprise AI Tool?

Real product-market fit means the team actually using the tool finds it useful, even inside a massive enterprise where the end customer is internal, not the model's technical accuracy. Kaplan is direct that this applies even when an AI product is being sold to the largest companies in an industry.

"Even in very large organizations, there still needs to be product-market fit. And in this case, your customers might be your ops people. If your ops people do not find the product useful, they're not going to use it, completely regardless of what you've built. ... A good AI product, a good model, is a model that's useful. It doesn't have to be 110% correct, it doesn't have to be perfect, but it needs to be useful." - Irena Kaplan

For an AI execution layer like Engini, that translates directly into how workflows get scoped: against the exceptions and bottlenecks an operations team already deals with, like claims or transactions that keep getting kicked back for manual review, with usefulness tested against the real workflow before an agent goes into production.

When Is the Right Time to Launch an AI Project Inside a Large Organization?

The right time to launch is when there is organizational willingness to adopt and train, funding in place, and the team is not already overwhelmed with other priorities, not whenever the technology happens to be ready. Kaplan describes this as a cadence problem more than a technology problem.

"You need to come up with the right product at the right time when there is organizational willingness to adopt and train, and then when there are resources, so your implementation can be funded. ... If you launch it at the wrong time, once everybody's super busy, they can't drop everything they're doing and do AI now, because this is essentially not their core role." - Irena Kaplan

This is also why Engini is built as a non-disruptive overlay rather than a rip-and-replace project. It reads and writes through the systems a finance or operations team already uses, so a rollout does not compete with the team's existing workload for attention at exactly the wrong moment.

What Should a HIPAA and SOC 2-Compliant AI Governance Framework Actually Include?

A compliant AI governance framework starts with legal non-negotiables like HIPAA and SOC 2 Type II, then cascades those requirements down into internal policy that gets implemented directly inside the AI system, not bolted on afterward. Kaplan is explicit that this is not a paperwork exercise.

"You need to be HIPAA compliant, you need to be SOC 2 Type II compliant. Those are non-negotiables. And the reason they're non-negotiables is not even the fact that it's legislation. It's because when you're not HIPAA compliant, my data and your data is not safe. ... We took state and federal regulations and interpreted them with our compliance and legal teams into internal policies that then were implemented onto AI." - Irena Kaplan

She also points to a practical example that most teams overlook: confirming whether an AI note-taker or enterprise tool is even allowed to record a HIPAA-covered conversation, and disabling it if it is not. This is precisely the layer Engini adds on top of any AI workflow. Every action an agent takes is logged, scoped to defined permissions, and reviewable, producing the audit trail a compliance review actually requires instead of one assembled after the fact.

Does AI Replace Healthcare Jobs, or Just the Busywork?

AI in healthcare and insurance has consistently absorbed the busywork inside a job, especially around claims and payment integrity, rather than replacing the people doing it. Kaplan calls this the most rewarding part of her work.

"We never replaced anybody. What we had done was providing them the tools that made them much more efficient, much more compliant, and also made them much more accurate. ... A concern in the insurance industry in general is payment integrity, how do you pay claims accurately in a timely manner. What AI did was automate a lot of the busy work, so you were able to prioritize your work better." - Irena Kaplan

That is the exact category of work Engini's AI agents are built to absorb across finance and operations teams: repetitive exception handling that consumes hours without requiring real judgment, leaving the judgment calls for the people trained to make them.

Why Do So Many Enterprise AI Vendor Pitches Fail to Land?

Most AI vendors underestimate how many internal stakeholders, finance, the end-user team, and legal or compliance, all have to independently sign off before a large enterprise will buy, and they expect a sale that moves far faster than that consensus can actually form.

"There are the people who are actually in charge of the finance. There are the people who are your customers, that might be the ops team. And then there are the people who are authorizing the purchase, which might be legal and compliance, because nothing goes in unless it's compliant. The question is how do you coordinate all of them. ... A lot of vendors really expect a very quick sale, but the truth is you need to create internal consensus that's robust enough to get buy-in." - Irena Kaplan

Engini's overlay model is built around this reality rather than against it: no core system replacement, no months-long migration, and an audit trail that gives legal and compliance teams something concrete to sign off on rather than asking them to trust the rollout on faith.

Is Shadow AI a Real Risk Inside Large Enterprises?

Yes. Employees downloading or using unauthorized AI tools without IT or compliance approval is a genuine risk, not a hypothetical one, especially anywhere protected health information or other sensitive data is involved. Kaplan's rule is blunt.

"The idea in large enterprises is can does not mean should. If your IT department had missed a particular app and you can download it, does not mean that you should, especially not without consulting IT and compliance. ... We do handle incredibly sensitive information that has to do with real people. If anything leaks, it's your data, it's your neighbor's data, it's your family's data out there." - Irena Kaplan

This is why governance has to be built into the AI layer itself rather than enforced through policy memos alone. Engini's permission-scoped access model means an agent can only touch the systems and data it was explicitly authorized for, so the shadow AI risk Kaplan describes is closed at the architecture level, not left to individual judgment calls.

Key Takeaways & FAQ

• AI projects fail most often from missing product-market fit, bad timing, and weak governance, not weak technology. RAND Corporation puts the overall AI project failure rate above 80%.
• Product-market fit in an enterprise means the internal team using the tool, not just the buyer, finds it genuinely useful.
• AI rollouts succeed when they launch during calmer periods, with organizational willingness, budget, and team bandwidth in place.
• HIPAA and SOC 2 compliant governance requires cascading legal requirements into internal policy that is actually implemented inside the AI system.
• AI in healthcare and insurance has consistently automated busywork like claims processing, not eliminated the roles around it.
• Enterprise AI sales stall when vendors underestimate how many stakeholders, finance, end users, and compliance, must independently agree before a deal closes.
• Shadow AI, unauthorized tools used without IT or compliance sign-off, is a real and avoidable risk, not a hypothetical one.

Why do most AI projects fail to scale past a pilot?

Most fail because they launch without confirmed product-market fit, on the wrong organizational timeline, or without governance that gives compliance and operations teams confidence to use the tool. According to Irena Kaplan and RAND Corporation research showing an over 80% AI project failure rate, the technology itself is rarely the actual point of failure.

What does product-market fit mean for an enterprise AI tool?

It means the internal team actually using the tool, such as operations or claims staff, finds it useful, even when the buyer is a different stakeholder entirely. As Kaplan puts it, the model does not need to be perfect, it needs to be useful.

What does a HIPAA and SOC 2 compliant AI governance framework require?

It requires cascading state and federal legal requirements into internal policy that is implemented directly inside the AI system, plus an audit trail documenting what the AI did, when, and under what permission, so a compliance review does not require reconstruction after the fact.

Does AI eliminate jobs in healthcare and insurance operations?

No. According to Kaplan's direct experience, AI absorbed the busywork inside claims processing and payment integrity, making staff more efficient and accurate, without replacing the roles themselves.

Why do enterprise AI sales take so much longer than vendors expect?

Because finance, the end-user team, and legal or compliance all have to independently sign off, and building that internal consensus takes time regardless of how strong the product demo looks.

How does Engini relate to the strategy Irena Kaplan describes?

Kaplan's framework defines what sound AI strategy looks like: validated product-market fit, the right timing, and governance strong enough to earn trust. Engini is built as the execution layer underneath that strategy, an AI Financial Execution Layer that automates exception handling, enforces permission-scoped governance, and produces the audit trail compliance teams need, without requiring a rip-and-replace of the systems already in place.

For the full conversation with Irena Kaplan, watch the episode on YouTube or visit The Engini Room podcast page for the full archive of episodes. For more on avoiding the most common pitfalls in enterprise AI rollouts, see our guide to the most expensive AI implementation mistakes, and review the official HIPAA guidelines from HHS before scoping any healthcare AI workflow.