Why 95 Percent of AI Pilots Are Failing (And How to Avoid the Trap in Insurance)

Discover why, according to the MIT report, 95 percent of generative AI pilot projects are failing and learn practical strategies to avoid common pitfalls. Explore insights about buy vs. build decisions, vendor risk, and using cloud-based AI services for faster, more reliable results for the insurance industry.

In brief:

• According to the MIT report, 95 percent of generative AI pilot projects in insurance are failing.
• Most insurers struggle because technologies that should offer quick returns are not delivering as expected, making fast, iterative learning essential.
• Buying AI solutions from established vendors succeeds twice as often as building internally, but relying on small startups introduces significant vendor risk.
• Using AI services from major cloud providers like Microsoft, Salesforce and Amazon offers advanced capabilities with less risk and smoother integration, especially since most insurers already have relationships with these vendors.

I wasn’t surprised when the new MIT report confirmed that 95 percent of generative AI pilots are stalling — but it should worry every insurance executive reading this. Insurance research firm Datos Insights found similar failure rates in the insurance industry earlier this year.

What’s mind-blowing isn’t these statistics by themselves. These technologies should deliver immediate, certain payback, yet many insurers are struggling.

We all know how fast AI technology is advancing. The cost of getting it wrong is higher than with any technology that has come before. If you are going to fail, failing faster and getting another “at bat” is more critical than ever.

The Real Culprits Behind AI Implementation Failures

Building Instead of Leveraging

MIT’s research reveals that purchasing AI tools from specialized vendors succeeds 67 percent of the time, while internal builds succeed only one-third as often.

But the “buy vs. build” framing misses the real opportunity for insurers. Buying off-the-shelf insurtech solutions might get you there quicker, but it introduces vendor risk. Many AI startups are heavily invested in older-generation technology or may not survive the next funding cycle.

True custom building isn’t realistic for most insurers, despite the availability of model deployment tools like Copilot Studio that make fine-tuning seem accessible. In reality, fine-tuning existing models is unnecessary for most insurance use cases. Document extraction, workflow routing, and content analysis work exceptionally well with standard base models.

The sweet spot for many is using what your cloud vendors already offer. Microsoft, Salesforce, and Amazon aren’t going anywhere, and their AI services are sufficiently advanced to keep you at the bleeding edge without the vendor risk of smaller players. Most insurers already have enterprise relationships with these providers, along with the security, compliance, and integration frameworks needed for fast production deployment.

Aiming Too High, Ignoring Low-Hanging Fruit

More than half of AI budgets are going to sales and marketing tools, according to the report, while the biggest immediate ROI I see in insurance sits in back-office automation. Insurers fund chatbots and lead scoring while their underwriters manually extract data from PDFs, and their claims teams route submissions by hand.

Companies are ignoring the high-volume, repetitive processes that AI automates best in favor of customer-facing initiatives that sound good in board presentations but are challenging to get over the finish line. External-facing applications require extensive validation, yet most insurers still lack experience with non-deterministic AI testing (where outputs can vary with the same input), making projects lengthier and riskier.

Carriers are attempting to automate complex underwriting decisions before they’ve automated basic document routing. They’re building sophisticated models to replace senior underwriters while junior staff still manually categorize submissions. Start with the mundane, high-volume tasks you can supervise and deliver immediate value, not the strategic decisions that require years to validate. Short feedback cycles are key to fast value delivery.

Unrealistic Expectations and Job Anxiety

There’s another factor the MIT report didn’t address: We’re expecting more from AI than we do from humans. When a new hire makes mistakes for six months, we call it a learning curve. When AI produces imperfect results on day one, we declare it a failure.

Without question, some of this resistance stems from a deeper anxiety. If AI is going to replace knowledge workers before they retire, then it better be perfect immediately — otherwise, why should they help build their own replacement?

This creates an impossible standard where AI must not only match human performance but exceed it flawlessly from day one.

Search Paralysis (When You Already Have the Capabilities)

Between Request for Proposals (RFPs), demos, pilots, and decision committee ruminations, vendor evaluation can easily consume a full year and feel like progress. Meanwhile, the manual processes that prompted the search continue unchanged.

The most successful AI implementations use existing enterprise platforms rather than introducing new vendor relationships. Most insurers already own Microsoft enterprise licenses that include Azure Document Services, OpenAI integration, and Power Automate capabilities.

Many also have Copilot deployments they don’t use to the fullest (do these count as AI failures?) because employees weren’t trained on effective prompting techniques, or context size constraints for insurance-specific workflows limit the solutions.

These tools can automate document processing, workflow routing, and decision support today. Instead, carriers evaluate standalone AI vendors and extend procurement cycles by months.

What Success Actually Looks Like

The report goes on to say that five percent of organizations achieving revenue acceleration share common characteristics that insurance executives should note. They focus on single pain points rather than enterprise-wide transformation.

They start with low-hanging fruit — document processing, data extraction, simple routing decisions — rather than attempting to hit a home run and automate complex judgment calls. They execute quickly rather than endlessly planning. And they partner with proven technology providers rather than building everything internally.

In insurance terms, this means automating the boring stuff first. Document ingestion before underwriting analysis. Email routing before customer service interactions. Data extraction before autonomous agents. These mundane tasks offer immediate, measurable value while building organizational confidence in AI capabilities. And it introduces AI benefits in a non-threatening way. It’s hard to get defensive around document intake or work routing.

Most importantly, successful AI implementations set reasonable expectations. AI doesn’t need to be perfect to be valuable — it needs to be better than the manual alternative, which often sets a surprisingly low bar.

The Path Forward

The insurance industry has a choice: Continue the pattern of overengineering solutions and underdelivering results, or learn from the organizations that are actually succeeding with AI. The technology is ready, the platforms exist, and the use cases are obvious.

What’s missing is the discipline to execute simply and the courage to start small. Stop building what you can buy. Stop planning what you can test. Stop debating what you can prove.

The carriers that figure this out in 2025 will have operational advantages that compound over time. Those that don’t will be explaining to boards why their AI investments look impressive on paper but haven’t moved the business metrics that matter.

The 95 percent failure rate isn’t inevitable — it’s a choice. Choose differently.

Based on two decades of insurance technology consulting and research from Datos Insights and MIT’s recent AI implementation study. For insurers ready to avoid the common pitfalls and deliver immediate AI value, the path is more straightforward than most realize.

This blog was originally published on Medium.com.

Are you struggling with insurtech overload? Our Insurance team is uniquely positioned to provide you with a combination of pragmatic solutions and services focusing on achieving your objectives. We’ll put it all together so you can avoid the common pitfalls, such as disjointed, unconnected solutions. Contact us