According to the study by MIT Project NANDA, 95% of enterprise GenAI initiatives deliver no measurable return. The study spans every industry, not just healthcare contact centers. The study also showed that AI systems don’t learn. They don’t retain feedback, adapt to context or improve over time.

The best way to think about enterprise AI is like a heart transplant. The new organ has to keep adapting to its host or the host rejects it. Most AI deployments get this backward. They ask the host to adapt — rewriting workflows, retraining staff, building out new SOPs. The industry has names for that work: change management. You can’t ask the organization to learn AI. The AI has to learn the organization.

The pattern shows up in healthcare. We see it constantly in patient access and contact centers. Pilots launch with fanfare. Six months later, they’re shelved or limping along on a narrow use case no one talks about. But what about the 5% that are working? What are they doing differently? Let’s take a look. 

The complexity in your rules and overrides isn’t a bug

Every practice runs on layers of operating complexity. Different appointment types for different providers. Eligibility rules that change by payer. Acute-issue routing that overrides the normal scheduling flow. Refill SOPs that vary by site. Half of it lives in the EMR, but the other half lives in the heads of the people who’ve been answering the phones for ten years.

It’s easy to look at all that and call it “a mess.” But it’s not. It’s how the practice deals with a healthcare system that doesn’t meet it halfway — dozens of payers and plan types, state-by-state rules, providers with individual scope-of-practice constraints. Somebody had to figure out how to make appointments happen inside all of it. That configuration is one of the most valuable assets the practice owns. It’s how patients get care, but it’s also what most AI vendors ask the practice to throw away.

The wrong question to ask an AI vendor

Most AI evaluations you’ll have with a vendor start with: What does the AI do? The better question is: Can the AI absorb everything we already do?

The first question leads to a feature checklist and a lengthy implementation that culminates in the practice rewriting its SOPs to fit the AI. The second leads to a partner who learns the practice’s rules, integrates with the EMR you already run and works alongside the staff who already know how things work. 

Very few AI platforms can do the second, because it’s hard. That takes configurability, bi-directional EMR integration and a partner willing to do the work to map out all the implied rules and overrides, written or unwritten.

What to ask before you start an AI pilot

Here are six questions to ask an AI vendor: 

• Will the AI learn all our rules and SOPs or will we retrain staff to match the AI?
• Is this real AI or a rule engine with AI marketing?
• Does it read and write to our EMR in real-time?
• Is it one platform across phone, web, messaging and back-office — or five separate contracts?
• Can we be live in weeks with measurable ROI in 60 days or is this a year-long implementation?
• Will it scale as we add new use cases or will we be shopping again in a year?

The 5% of companies with successful enterprise GenAI initiatives don’t have better AI. They have a better way of buying it, asking better questions and having a better partner. 

Pretty Good AI is exhibiting at #HCCT26 and presenting “The A-Z of Deploying AI Voice in Your Multi-Location Call Center: What to Ask Before You Buy,” with Jeannine Spagna, VP of Patient Experience & Call Center Operations at Clearway Pain Solutions. Find them in the Exhibit Hall, June 3-5 in Atlanta.