6 min 
A few months into an AI/ML-powered Next Best Action rollout, I watched a sales rep dismiss a recommendation almost before it finished loading. I’m exaggerating here, only slightly, but you get the point.
“Nope.” No hesitation. No curiosity. Just one click, and it “disappeared” from their CRM.
The easy explanations would have been that the models wrongly predicted this physician or alternatively that the rep ‘doesn’t trust AI.’ To be fair, sometimes that first explanation is right — some AI tools genuinely aren’t tuned well enough for the job, and no amount of framing this as a human problem should let a bad model off the hook. That’s a real failure mode, and it deserves its own conversation. But it wasn’t the failure mode I kept running into. After spending months watching these interactions play out from a front row seat, it became obvious that for the recommendations I was watching — where the model itself was reasonably sound -this wasn’t really an AI problem. It was a human one.
Note: I’m writing this from inside pharma, because that’s the industry I watched it happen in. But my hunch, and it is a hunch, not a claim (I’m a scientist and I don’t have the data to back this up across other industries), is that none of this is actually specific to Pharma. The same resistance likely shows up anywhere expertise is asked to defer to a system that just “showed up.” Details change by domain; the resistance probably doesn’t. Of course, not every rep reacts this way — some are first in the door for anything new, a CRM, a call model, a device, doesn’t matter. This write-up isn’t about them.
The conversation usually starts with AI. But the real challenge isn’t AI at all. It’s asking people to change how they do their jobs. And that problem is older than software.
Tell a kid to stop doing something a certain way and watch what happens. Tell an adult sales rep who’s been managing the same territory for eight years that a model now has an opinion about which physician they should visit next. The reaction isn’t all that different. Nobody hears “do it differently” and immediately thinks, “Great, let me abandon what’s been working for me all these years.”
That’s not unique to pharma. It’s not unique to AI/ML systems that recommend what you should do next, either. I mean, I refused to use Grammarly back in the day even for my thesis. Not because it wasn’t good — it probably would have caught things I missed. But writing is part of how I think, and every suggestion feels like a tiny editor peering over my shoulder saying, “Are you sure?” Maybe that’s irrational. I don’t care. There’s pride. There’s ownership. There’s a voice I don’t want an algorithm smoothing out. Recommendation systems, whether they’re suggesting how to write, what code to generate, or which HCP to visit next aren’t just offering advice. They’re asking someone to trust a second opinion over instincts they’ve spent years developing. If I feel that way about writing, why would I expect a field rep who’s spent a decade building instincts in a territory to instantly embrace a recommendation about who to visit next?
The question isn’t, “How accurate does the model need to be before reps trust it?” The real question is much older than machine learning: “How do you get someone who’s already good at their job to accept a second opinion they didn’t ask for?”
I don’t think the answer starts with making the model smarter. I think it starts with acknowledging that the rep knows things the model might miss. They know an HCP just had a difficult quarter and isn’t taking meetings. They know a territory was reshuffled last month and the CRM hasn’t caught up yet. They know a physician is technically still in the data but retired in practice. They know an office switched systems and everything has been chaotic for weeks.
None of that shows up cleanly in a dataset. Not until the rep tells the system. The model brings pattern recognition across thousands of territories and interactions; the rep brings context that exists in the field. Neither is complete on its own. The value comes from the feedback loop: the model informs the rep, the rep teaches the model, and each makes the other better.
What surprised me most wasn’t whether reps engaged with recommendations — adoption was already solid, and reps were acting on the majority of what surfaced. What surprised me was what we learned by looking closely at the ones they didn’t act on. This is a pattern I’ve seen play out across recommendation systems generally, not something unique to any one engine: When every dismissal is grouped into a single bucket like “incorrect recommendation”, valuable feedback gets lost. Not every rejected recommendation is rejected for the same reason and understanding the ‘why’ matters. For example, a rep may be signaling that the underlying data has gone stale — for instance, an HCP has retired or moved territories (i.e., data case). In another case, the rep may simply see the situation differently than the model — not because the model is wrong, but because they’re working from information the model doesn’t have.
The perspective cases are the more interesting ones. The rep is challenging the recommendation, but that doesn’t automatically mean the model is wrong. Maybe the rep has local context the model doesn’t. Maybe the model is surfacing an opportunity the rep hadn’t considered. Or maybe the rep simply prefers calling on physicians they’ve had success with before. Those are very different reasons, and they deserve very different responses. The data cases are more straightforward — fix the underlying record, and the recommendation corrects itself
Once those signals are separated, everything changes. One points to a data quality problem. The other is an opportunity to understand behavior, business rules, and where the model and the rep are seeing the world differently. Every dismissal is data — the only question is whether it’s treated that way.
I think that’s the piece people miss when they frame this as “getting reps to trust AI.” It assumes trust flows in one direction — as if the only job is convincing the rep that the recommendation deserves a chance.
But that’s backwards.
Trust isn’t built when the rep follows the recommendation. It’s built when the system shows it’s willing to learn from the rep. The recommendation is only half the interaction. The dismissal is the other half. What happens after the rep agrees, disagrees, or dismisses a recommendation is what determines whether they’ll engage with the next one.
I’ve started thinking that we spend too much time asking how to make AI more persuasive and not enough time asking how to make it more teachable.
If a rep tells the system, “This recommendation is wrong because this physician retired,” and the same recommendation keeps appearing next week, the lesson isn’t that the rep should trust AI more. The lesson is that the AI isn’t listening. And people stop talking to things that don’t listen. That’s true whether you’re building enterprise software, raising children, or working with a colleague.
That’s why I don’t think the goal is to build a recommendation engine that always wins the argument. Sometimes the model will have spotted an opportunity the rep hadn’t considered. The rep will know something the model didn’t know yet. The interesting question isn’t who was right — It’s whether the system became smarter because of the interaction.
Giving reps a way to explain why they disagree transforms feedback from a dead end into a learning mechanism. Whether that feedback corrects stale data, refines business rules, or captures context the model didn’t have, it improves future recommendations. In doing so, the system stops feeling like something that’s talking at them and starts feeling like something that’s learning with them.
Trust doesn’t come from asking people to ignore their experience. It comes from respecting it. I still think about that rep who clicked “Nope.” At the time, I thought I was watching someone reject AI. Looking back, I think I was watching someone ask a perfectly reasonable question: “If I’m going to learn from you, are you willing to learn from me?
That’s a much harder problem than building a better model. It’s also a much more human one.
Dolonchapa Chakraborty, PhD
Director, Product and Engagement Lead
