You can't fine-tune your way out of a data problem

A team of researchers recently reviewed 44 studies of medical AI deployment in low-resource settings and asked a useful question: what actually gets in the way once AI meets a real health system? The answer was striking. The barriers reported most often weren’t about the sophistication of the model. They were data quality, infrastructure, fragmented records, workforce capacity and governance. The review’s authors put it more plainly than most consultants would dare to: successful deployment is less about advanced technology than about having the underlying systems in place.

What the 44 studies reported

The review sorted barriers into four domains and counted how many of the 44 studies reported each.

Read either list from the top. The things that break AI in these settings are, overwhelmingly, the things underneath the model: whether the data was recorded at all, whether it is clean and interoperable, whether power stays on, whether anyone will still run the system a year from now.

Now think about where the money usually goes. A ministry announces an AI pilot. A foundation earmarks a budget. An NGO writes “AI-enabled” into a proposal. The spend lands on a model license, a batch of compute, a tool with a recognizable name. The visible part. Rarely the part the evidence says is actually broken. That gap, between where the failures live and where the money goes, is the whole subject of this piece.

It isn’t only a low-resource-health problem

Deloitte’s State of AI in the Enterprise survey for 2026, built from 3,235 leaders across 24 countries, found a similar gap inside some of the best-funded organizations on earth. Forty-two percent considered their AI strategy highly prepared. But readiness for the operational machinery underneath was less reassuring: 40% for data management and just 20% for talent. More telling still, confidence in infrastructure, data and talent had declined from the previous year, even as confidence in strategy rose.

Deloitte has a name for what follows: the “proof-of-concept trap.” A pilot runs happily for a few months on clean data in a sealed environment. Then production arrives, demanding integration, security, monitoring and upkeep, and a use case scoped at three months stretches to eighteen. One healthcare AI leader in the report describes the predictable result as “pilot fatigue,” chasing the next tool with no plan for how any of it scales.

Two very different worlds, a rural clinic network and a multinational balance sheet, reach the same finding. The excitement pools at the top of the stack. The failures pool at the bottom.

AI is a stack, not a product

It is tempting to treat this as a preferences debate, data versus compute versus infrastructure, rank them however you like. That misses the point. These aren’t rival options on a menu. They’re a stack, and the stack has a direction.

Each upper layer feeds on the ones below it, and it swallows their weaknesses along with their strengths. The layer most people forget is integration and workflow. A model can sit on good infrastructure and good data and still fail because nobody decided where its output goes, which system consumes it, who is accountable for acting on it, or what happens when its recommendation collides with a clinician’s judgment. Deloitte’s proof-of-concept trap is largely a story about that layer.

This is also why compute applied to broken data can be worse than no AI at all. The danger isn’t merely a weaker output. Modern AI can turn incomplete, biased or badly structured data into an answer that is fluent, polished and wrong, then reproduce that error at machine speed. The Frontiers review makes the clinical version of the point directly: when a model learns from data that doesn’t represent the local population, it serves some groups worse than others. In a clinic that is not a cosmetic flaw.

In health systems especially, “the data problem” is rarely one neat problem. The HIV test result may live in one EMR, the lab result in another, the facility identifier may not match across the two, and the national platform may hold only the aggregate. A model sitting above that architecture cannot infer the missing lineage. Someone has to fix the system. That is the work the model can’t do for you, and the work the budget rarely names.

What the fundamentals actually buy you

None of this argues for freezing until the data is perfect. That would stall every program forever. It argues for sequence, and for what the lower layers unlock once they hold.

The same review that catalogs the failures also lists what works, and the enablers are pointedly unglamorous: resilient and offline-capable infrastructure, interoperable standards, shared data repositories, ongoing training for the people who run the systems, and governance wired into national health strategy rather than bolted onto a donor pilot. It also names the cost of skipping them. Projects funded as short-term pilots, cut off from national systems and steady budgets, tend to rot once the external money ends.

The upside is the mirror image. Once the layers beneath are solid, you can do the hard things on top. You can ground an answer in a verified figure someone can trace to its source. You can impose confidence thresholds, validation rules, provenance checks and human review, instead of asking a model to bluff its way through missing context. That discipline is available precisely because the fundamentals hold. It is the payoff for doing the boring layers first.

Five questions to ask before you sign

For the people approving budgets and signing contracts, the whole argument collapses into a better first question. Not “which model,” but “what is the system underneath it, and is it strong enough for the consequence of being wrong.”

1. What decision or workflow is this supposed to improve? What happens today, how well does it work, and what measurable improvement would justify deploying AI at all?
2. Does the required data actually exist, and does it represent the people and situations where this will be used? Check completeness, timeliness, granularity, lineage and interoperability before the demo, not after.
3. What must this integrate with to work in production? EMRs, APIs, identity, data warehouses, reporting systems, connectivity, security controls, and what happens when one of them is unavailable.
4. How will we know when the system is wrong? What is validated, logged, monitored and reviewable? Who can override it, and what happens when confidence is low or inputs are incomplete?
5. Who owns this after the pilot? Who pays for infrastructure, support, monitoring, retraining, licenses and staff capacity twelve months after the grant ends?

A vendor who can field those calmly is worth more than one with a slicker demo. A proposal that budgets for the layers under the model will usually outlast one that spends the whole line on the model itself.

The honest exception

Fundamentals first is a sequencing principle, not a prohibition on experimentation. A narrow, well-scoped AI tool can sometimes create the business case that finally gets neglected data cleaned, systems integrated or workflows redesigned. And imperfect data may be perfectly adequate for a low-risk use where a human verifies the result. The relevant question is not whether the foundation is perfect. It is whether it is strong enough for the consequence of being wrong.

Build accordingly

The question is no longer whether your organization can access a powerful model. Increasingly, everyone can. The question is whether the system underneath it can support what you are asking that model to do.

You can fine-tune the model. You cannot fine-tune your way around missing data, broken integrations, weak governance, or an operating model that was never designed to carry AI. Build accordingly.

This is a field note from our practice. If you’re weighing an AI investment and want an honest read on whether the layer beneath it will hold, start a conversation.

Sources

The barrier frequencies come from Al-Ganad A, et al., “Deploying medical AI in low-resource settings: a scoping review of challenges and strategies” (Frontiers in Digital Health, vol. 8, 1 April 2026), drawn from the review’s thematic synthesis across 44 included studies; the review does not rank model choice as a separate barrier category. Enterprise figures and the “proof-of-concept trap” come from the Deloitte AI Institute’s State of AI in the Enterprise: The untapped edge (January 2026), a survey of 3,235 leaders across 24 countries. For further context, see “Can artificial intelligence revolutionize healthcare in the Global South?” (DIGITAL HEALTH, 2025).