The scenario below is a composite, built from patterns we've seen play out across multiple healthcare and med-device organizations. It's not a real client, and the numbers are illustrative, not reported figures. The point isn't the company — it's the pattern, because this same story repeats itself in some form at almost every mid-size healthcare or industrial-health organization that's tried to bolt AI onto its list-building process in the last two years.
The Situation
Picture a mid-size medical device manufacturer selling a category of diagnostic equipment into hospitals, ambulatory surgical centers, and independent imaging groups. Their sales team was capped at maybe a dozen reps covering the country, and marketing was under pressure to keep the pipeline full without growing headcount. Someone on the leadership team had read enough about generative AI to conclude — reasonably, on the surface — that AI could "build the list" faster and cheaper than the manual process the team had relied on for years: a mix of a purchased database, LinkedIn Sales Navigator pulls, and a part-time researcher who cross-referenced facility websites by hand.
The pitch internally was simple. Feed an AI tool a target profile — facility type, bed count, imaging modality, geography — and let it generate a prospect list with contacts, titles, and firmographic detail. Cut the researcher's workload, speed up the cadence of new segments, free up the reps to sell instead of qualify.
What Didn't Work
The first list came back fast. That was the good news. It was also, in hindsight, the warning sign.
The AI tool — a combination of a large language model doing entity extraction and a scraping layer pulling from public web sources — produced a list of a few thousand facilities and named contacts within days. On paper it looked like a win. In practice, three problems showed up almost immediately.
First, the facility data was stale in ways that weren't obvious until reps started calling. Ambulatory surgical centers had merged, imaging groups had been acquired by private equity roll-ups and rebranded, and a meaningful chunk of "independent" facilities on the list were now part of larger systems with centralized purchasing — which changes who the actual buyer is entirely. The AI had scraped current-looking web pages, but healthcare ownership structures change faster than most websites get updated, and the model had no way to know that a page it was confident about was six months stale.
Second, the title-matching was confidently wrong in a specific way that's easy to miss if you're not close to the data. The model was very good at finding someone with "Director of Imaging" or "VP Procurement" in their title. It was much worse at knowing whether that person actually had budget authority for this device category at that facility, versus a similar title that was purely clinical or purely administrative. The AI wasn't hallucinating names — the names were real people who did work there — but the functional fit was often wrong, and wrong in a way that looked plausible on a spreadsheet.
Third, and least visible at first: the tool had no concept of what "good" meant for this business specifically. It generated a list that matched the literal filter criteria, but it had no signal on which of these facilities had budget cycles aligned to the sales team's quarter, which had recently bought a competitor's equipment, or which were even open to new vendors given group purchasing organization contracts. Those are the qualifiers that actually predict whether a list converts, and none of them live in the kind of public data an AI tool can quickly parse.
The result, a few weeks in: reps burned time on calls to facilities that had merged out of relevance, pitched people who didn't have the authority the title implied, and the "faster list" ended up requiring almost as much manual cleanup as building one from scratch — just after the fact instead of before.
What Changed
The team didn't abandon AI. That's the part worth noting, because the easy narrative is "AI failed, back to manual." What actually happened was a re-scoping of what the AI was asked to do.
Instead of using AI to originate the list end to end, they used it for the parts it's genuinely strong at: fast pattern-matching across large volumes of structured and semi-structured data, and rapid first-pass entity resolution — figuring out that "Riverside Imaging Center" and "Riverside Diagnostic Partners LLC" were probably the same organization under a rebrand, something that used to take the researcher hours per batch.
The verification and qualification layer — confirming ownership structure, confirming the contact's actual role in purchasing, checking recency against something more reliable than a scraped webpage — stayed human, but became faster because the AI had already done the tedious first sort. This is roughly the model NPLUS Global uses when we talk to clients about where automation actually earns its keep in healthcare data: AI compresses the sorting and matching work, but a human (or a verified data pipeline built by humans) still has to close the gap between "looks right" and "is right," especially in an industry where entity structures shift constantly and titles rarely map cleanly to authority.
They also narrowed what they asked AI-generated lists to promise. Instead of treating the output as sales-ready, they treated it as a shortlist for verification — which sounds like a small semantic shift but changed how the team behaved. Reps stopped getting handed lists with implicit confidence they hadn't earned.
The Outcome
The honest version of the outcome isn't a dramatic transformation. Time-to-list for a new segment got noticeably shorter — not because AI replaced the researcher, but because it removed the most repetitive part of their job and let them spend their hours on the judgment calls that actually required a person. List quality, measured informally by how often reps reported dead-end calls, improved compared to the pure-AI-generated batch, though it never quite matched the old fully manual process in precision — it matched it in speed while getting close enough in accuracy that the tradeoff was worth it.
The bigger shift was cultural. The team stopped asking "can AI build our list" and started asking "which specific step in building our list is AI actually good at." That's a less exciting sentence than a vendor demo, but it's closer to what's actually happening across healthcare data operations right now — incremental, uneven, and genuinely useful only when someone is honest about where the tool's confidence outruns its accuracy.
Ready to see what we can build for your ICP?
Send us your ICP — sample in 2–3 hours, full delivery in 48–72 hours.
Request a free sample →