Every org I talk to has an AI project in limbo. Not dead. Not deployed. Just... waiting. Sitting in a folder somewhere between The Next Big Thing and the thing that's actually trusted to run the business.
Ask five people at the same company whether that product is close to ready, and you'll get five different answers, none of them backed by a shared standard for judging it. That's the real bottleneck in this market right now. Not model quality, not ambition, not even budget. It's trust. Specifically, the total absence of any structure for building trust: a repeatable way to tell the difference between a great idea that needs more runway and a flop that needs to be killed.
Without that structure, both get the same treatment. They sit. Forever, if nobody intervenes.
.png)
This isn't a motivation issue, and Snowflake's The ROI of Gen AI and Agents 2026 research backs that up. 92% of early adopters report positive ROI on gen AI, and teams keep pushing forward: 32% already have agentic AI in production, and another 37% have definite plans. AI teams want to prove this stuff works. They're not the bottleneck.
What's actually blocking the move from dev/test into production comes down to three gaps, and they show up in the data over and over.
On average, only 20% of unstructured data assets are considered AI-ready — and worse, the share of orgs saying more than half of their unstructured data is AI-ready actually fell this year, from 11% to just 7%. That's not stagnation. That's regression, even as ambition and investment climbed. Garbage in, garbage out. You can't build trust in an output when the input is known to be — or even perceived to be — questionable.
56% of respondents report at least one governance or compliance-related challenge. That's not a fringe concern; it's a majority. And it's compounding: the share of respondents calling data governance tasks not just challenging but very challenging rose year-on-year across every category measured — breaking down silos, monitoring quality, enforcing policy, meeting infrastructure requirements. Nearly all (95%) also reported at least one cost overrun tied to their gen AI build, most often in the data and compute infrastructure meant to support exactly this kind of governance.
35% of respondents cite employee expertise or skill as a top challenge, and that number climbs for midmarket companies (43%) versus enterprises (34%) — the classic case of talent gravitating toward whoever can pay for it and offer the most interesting problems to solve. Without literacy, teams can't evaluate what they've built. They can use the model. They can't judge it.
Readiness tells you if the data can be trusted. Governance tells you if the process can be trusted. Literacy tells you if the people evaluating both actually know what they're looking at.
Miss any one of the three, and you don't get a failure — you get stagnation. Purgatory. Without conviction behind not just the AI solution but the data itself, your organization stays in limbo and falls behind those that have committed to building trust, transparency, and stable foundations.
Healthcare is the clearest example of this dynamic playing out at scale. It's simultaneously one of the more eager sectors on gen AI adoption — the second-highest for broad use-case adoption, at 47% — and one of the more conservative, structurally, by necessity. That combination should be a strength. Instead, the report shows it producing exactly the kind of stalled middle ground I'm describing: healthcare organizations report lower usage than their peers across nearly every functional team — customer service, ITOps, cybersecurity — and lower investment growth in AI security, privacy, and general AI literacy training than the cross-industry average.
That's an industry that wants to move, has to move carefully, and hasn't yet built the governance and literacy scaffolding that would let it move carefully and confidently at the same time. The result is R&D effort that goes in but doesn't come out the other end as production value. Not because the ideas were bad. Because nobody had a structure to confirm they were good.
This is where the mechanism actually has to change, and it's the piece most orgs skip: treating AI products as static assets that get a single risk review at launch and then get left alone. AI products aren't static. They drift, they get retrained, they get used in ways nobody scoped for. A one-time sign-off can't generate ongoing trust in something that keeps changing.
What does generate it is shared, continuous ownership — an AI engineer and a business user, paired for the life of the product, not just its launch. The technical side watches for the things a business owner can't see: drift, degraded inputs, model behavior creeping outside its original scope. The business side watches for the thing the technical side can't see: whether the output is actually still serving the goal it was built for. Neither perspective alone is enough to build trust. Together, applied continuously instead of once, they're what actually lets an org tell a great idea from a flop — and act on that difference instead of just guessing at it.
That's the throughline across everything I've been circling this year: infrastructure has become table stakes, but rollout stalls without the readiness, governance, and literacy to back it up, and without an ownership model built for assets that change over time instead of ones that don't. Get that right, and production purgatory stops being inevitable. It becomes a phase you pass through, not a place things go to disappear.
.avif)