On the importance of connective tissue between technical solutions and business problems. Also on Indiana Jones, Monty Python and the Holy Grail, and such other important matters.
Sometimes, enterprise data science can feel like one of those scenes in the Indiana Jones movies where Harrison Ford is walking through a tunnel and keep encountering skeletons of people who tried this before him. Except here, the skeletons are sophisticated algorithms that never saw production, elegant ML pipelines that solved no real problems, and powerful AI systems that gathered dust while business users struggled with Excel. The culprit isn’t usually bad technology—it’s the missing connective tissue between brilliant technical solutions and messy business realities.
Consider some common scenarios where this connective tissue makes or breaks the entire endeavor. (We don’t even have to get into LLMs because by now, more people have probably heard of Agents and Model Context Protocol than of Indiana Jones.)
Recommendations, and what happens next
An AI personalization platform would need a powerful recommender system. Building a world-class recommendation algorithm is a well-understood problem. Collaborative filtering, matrix factorization, deep learning approaches — the technical literature overflows with solutions that can predict user preferences with impressive accuracy. But the key isn’t the algorithm; it’s everything else around it.
A powerful recommender needs a use-case layer that translates algorithmic output into business-relevant scenarios. “Users who liked X also liked Y” is a technical insight. “Hey, we know you love your beach holidays, so here are some destinations we recommend” is a business conversation. The difference is the connective tissue that understands customer segmentation, purchase history context, and campaign strategy.
This means your recommendation system needs to handle queries like “give me recommendations in this customer’s favorite category” or “suggest items from categories this customer has never explored.” It needs to understand seasonal patterns, inventory constraints, and margin considerations. Most importantly, it needs to sync with your outreach mechanisms.
The campaign layer running customer lifecycle management nudges is where this gets interesting. A recommendation with generic context is just noise. But “Wi nøt trei a høliday in Sweden this yër?” — now that’s a conversation starter that acknowledges the customer’s history while giving them travel recommendations and introducing them to Monty Python and the Holy Grail (seriously, go watch it, and don’t miss the opening credits). The recommendation engine provides the intelligence, but the business context provides the personality and timing.
ML Platforms will build your model. But what model?
MLOps platforms have reached impressive maturity. You can spin up environments that contain the latest and greatest in gradient boosting algorithms, handle model versioning, automated retraining, and deployment pipelines with remarkable sophistication. The infrastructure problem is largely solved.
But infrastructure isn’t a business problem. The business problem sounds something like: I want to send out a Winterwear campaign to an appropriate subset of customers. Apart from not sending it to people in Chennai, what would you recommend? (Actually, even the Chennai part isn’t true – I’ve seen people wearing monkey caps on morning walks in December.)
Now, an obvious counter-argument here is that business problems can be very diverse. And here’s where we put on our statistician’s hat and say, “Just because something is diverse doesn’t mean it is uniform across all that diversity.” In other words, yes, there’s a whole bunch of different predictive modeling problems, but the vast majority of those fall into a small group of problem templates.
The translation from there to a technical problem looks something like this: When was the last time people bought Winterwear? Last winter, of course. Didn’t the word Winterwear give you a clue? And what did we know about our customers prior to that winter that could predict their winterwear purchase? Recency, frequency, favourite category, any past history of winterwear purchase…
So now we get to the hard part: feature engineering, understanding what constitutes a proper training sample, and so on. The hard part is knowing that in retail, you need to account for seasonality differently than in finance, or that customer churn models in a fund house involve solving for explicit account closure whereas they require solving for prolonged inactivity in retail.
A truly business-centric MLOps platform doesn’t just provide generic model training capabilities — it provides configurable templates for the predictive modeling problems that matter in that industry vertical. Without these business-specific templates, you’re solving the easier part of the problem (the infrastructure) while leaving business data science teams to tackle the harder part (the domain expertise). You end up with beautiful MLOps platforms that sit largely unused because the cognitive load of translating business problems into technical implementations remains too high.
We’ve seen this approach work in practice. One fashion retailer using SOLUS AI saw their campaign conversion rates improve by 4-5x when they moved from generic promotional campaigns to contextual, AI-driven recommendations that understood both customer preferences and seasonal timing. The technology was the same; the connective tissue made all the difference.
Building the Bridges
The pattern across these examples is clear: the technology is rarely the bottleneck. What’s missing is the thoughtful design of interfaces, templates, and integrations that translate between technical capabilities and business needs.
This connective tissue requires a different kind of thinking than pure algorithmic development. It requires understanding business workflows, organizational constraints, domain-specific requirements, and human factors. It requires building abstractions that hide complexity without sacrificing power.
Most importantly, it requires recognizing that in business-centric products, the measure of success isn’t technical elegance—it’s business impact. And business impact happens in the messy middle ground where sophisticated technology meets practical human needs.
The companies that succeed in data science aren’t necessarily those with the most advanced algorithms. They’re the ones that build the best bridges between what’s technically possible and what’s business-relevant. They understand that you can’t throw a technology solution over the wall and hope that a business problem will catch it.
We get this:
At SOLUS AI, we call this approach ‘Systems of Intelligence’ – platforms that sit between your data systems and customer engagement channels, providing the crucial translation layer that turns AI capabilities into business results. Because in the end, the most sophisticated algorithm is only as valuable as its ability to drive real business outcomes.
