A client approached us about building a model.
We developed a prototype, got approval, and delivered the finished model.
Then, after weeks of effort… silence.
It’s a familiar story for any data professional, whether you’re an analyst or an ML engineer.
So, what went wrong?
Your Model Feels Like a Black Box
Our field is built on modern computing and technological progress. Many of the most powerful tools we have today would have been computationally out of reach just a few decades ago. But with that reliance on cutting-edge innovation comes doubt.
In data science, we can build astonishingly complex models. On my team, our feature library alone contains hundreds of standard features we can feed into each new model build. We fine-tune dozens of hyperparameters and use advanced algorithms that run through hundreds of iterations to squeeze out maximum predictive power. This process can produce models with remarkable accuracy — but there’s a trade-off: no one can explain how they actually work.
There’s a thin line between a strong model and a black box that even its creators can’t explain.
That tension between explainability and accuracy is especially pronounced in healthcare. Our customers and stakeholders are often doctors and clinicians who rely on years of training and deep medical knowledge to make clinical decisions. A model might be excellent at predicting outcomes, but if no one can explain how or why it works, clinicians will question whether to trust them. Given the choice between a proven clinical process and a black box model with cryptic features and algorithms, most doctors will go with the proven process every single time.
So, how do you fix this? In my experience, the best approach is a simple, easy-to-read model brief — a short set of slides that walks the customer through the whole model story. It begins by defining the population of interest, the target variable, and the features, and then wraps up with proof-of-concept performance and validation. Throughout, I make sure to frame metrics in the context of the business question, stepping into the customer’s shoes. I steer clear of statistical jargon and keep everything grounded in their goals. If the model is complex, I stick to a high-level explanation of the algorithm and clearly communicate why I chose such an extensive — or deliberately simple — feature set. Putting together a thorough model brief is one of the most important steps in demystifying the model and letting stakeholders understand it in their own terms.
Your Solution Took Too Long to Arrive
Building a working model is time-consuming. Between back-and-forth conversations with customers, unexpected complications, and the inherent complexity of the work, creating an effective, useful model is anything as a quick job. And then there’s deployment — a whole process on its own.
The real world doesn’t stand still. Customers continue running their daily operations with the tools they already have — the tools they were using before they ever came to you. If the model takes too long to build, they may abandon the request entirely or find workarounds that don’t involve a predictive model at all.
We see this constantly in healthcare. A stakeholder asks for a model, but after a series of roadblocks — delayed communication, data access problems, deployment issues — weeks of development balloon into months. By the time you’re ready to present the validated results, you set up the meeting only to hear: “We no longer need the model — we worked it out on our own.” Hospitals are fast-moving environments. Staff don’t have the luxury of waiting around for months. They will find creative ways to improve patient care even if it means giving up on a sophisticated predictive model.
There’s a motto I follow: “Don’t let perfection get in the way of good enough.” Move fast. Brainstorm, refine, review — but always keep momentum. Chasing perfection can block you from delivering real value. The world doesn’t wait, and if you spend too long stuck in the build phase, it will leave you behind. Ship version one. If you discover a better approach down the road, put it first on the list for version two. A good solution now is almost always better than a perfect one later.
If progress is slower than expected, communicate with customers early and often. Keep them in the loop about where things stand, and give them a preview to keep them engaged and excited about what’s coming. Use that time to push through building v1 and get it into their hands.
Your Model Is Too Hard to Use
Building a strong predictive model is only half the work. In most industries, stakeholders are stretched thin. In healthcare, doctors and nurses are overwhelmed with patient care. If the data science team pitches their latest, most accurate model to a care team on the floor, but checking the predictions adds extra steps and slows them down, that model will never get adopted. The same principle applies everywhere. Stakeholders want solutions that boost efficiency, performance, and productivity — not ones that pile more complexity onto their already packed schedules.
If predictions create friction, you’re paving the way to abandonment, not adoption.
Delivering predictions in a user-friendly way is one of the hardest challenges data scientists face. Building precise, accurate models is our strength, but weaving those models into customers’ daily routines doesn’t come naturally. This part isn’t about numbers, probabilities, or statistical knowledge — it’s about operations, business sense, and a solid grasp of how requestors actually spend their days.
In a hospital, that means integrating into Epic, the electronic health record system used across the organization. Instead of asking overburdened clinicians to log into a separate platform to see predictions, they can find them right inside the patient’s chart, alongside their other clinical tools and data. In any other industry, the same principle holds: don’t disrupt the current workflow. Fit into it.
In Conclusion
One of the most disheartening experiences a data scientist can have is watching months of hard work go unused. It happens far more often than anyone would like to admit, and it’s tempting to point fingers at the customer. After all, that’s easier on the ego.
The truth is, there may have been some critical elements the data scientist overlooked during development. Recognizing the most common pitfalls can help ensure your models cross the real finish line: actual adoption.
