Featuring Insights from Raheel Gandhi, Data Analytics Program Management Veteran
For a decade, the mantra in data science was simple: “Accuracy at all costs.” If a neural network could predict churn or credit risk with 99% precision, we didn’t much care what happened inside the black box. We toasted to the results and ignored the mechanics.
But in 2026 the regulations are only getting more stringent. With the full maturation of the EU AI Act and the ripple effect of global transparency mandates, the “black box” has become a liability. Today, if a Program Manager (PM) cannot explain why a model reached a conclusion, they aren’t just facing a technical hurdle, they are facing a legal and ethical backlash.
We sat down with Raheel Gandhi, a data analytics and program management leader with over ten years of experience across global enterprises like Omnicom and Linkedin. His experience draws also from serving as a judge for multiple globally reputed conferences as well as being a Senior IEEE member. Raheel has built end to end analytics for Fortune 500 companies in his agency days as well as lead several business critical initiatives during his long stint at LinkedIn always keeping one eye on how ethics and transparency are at the center of everything he delivers. We invited him to share his experience navigating the shift from “pure math” to “principled analytics,” across multiple global enterprise projects, and discuss how to lead when the algorithm remains a mystery.
- The end of “Trust Me, I’m an Expert”
A few years back, PMs could get away with hiding behind a complex tech stack. If a stakeholder asked why a customer got rejected for a loan, you could basically just shrug and blame it on “complex variables” no one really understood.
“Those days are over,” says Raheel. “By 2026, a Data Analytics Program Manager is essentially the ‘Chief Ethics Officer’ for their own project. You’re the one stuck in the middle translating the ‘mystery’ the data scientists built into the ‘map’ the regulators are demanding.”
RG: “This shift has been happening for some time.I remember how Cambridge Analytics changed the landscape and I had to pivot the direction of multiple global projects. AI based targeting ad platforms began opening up their black box to showcase their competitive advantage which very quickly became table stakes. All of our clients would not only want to see the ROI on their marketing campaigns but also details on the targeting and attributes used to do so. Ever since, the expectation has only increased and companies are increasingly becoming attentive to the ‘how’ just as much as the ‘what’.”
- Managing the Unexplainable: The Trust Paradox
The core challenge of modern PMing is the ensuring trust: How do you maintain stakeholder confidence in a system that, by its very nature, involves high-dimensional math that no human can visualize?
Raheel shared how he manages the shift from Product Trust (trusting the output) to Process Trust (trusting the guardrails) using transparency and breaking things down to manageable chunks to build buy-in from all stakeholders for the full solution.
RG: “It starts with recognizing that not all stakeholders in the room are at the same level of fluency in new AI or algorithmic products. And you do need everyone at the table bought in nowadays. Hence, a couple ways I’ve seen success is by building proof of concepts before going all in. There are also tools now in the market that can take one data point and in essence, follow it through the model, which definitely helps demystify AI rather than trying to explain the full model. Think of it as seeing a real example vs trying to decipher the entire model.”
Strategies for “Black Box” Stakeholder Management:
- The “Shadow Model” Approach: Running a simpler, “interpretable” model (like a decision tree) alongside the complex one to see if they generally align.
- Local Interpretable Model-agnostic Explanations (LIME): Using tools that explain individual predictions rather than the whole model.
- Confidence Intervals as Communication: Instead of saying “The model says X,” the PM says “The model is 85% certain of X based on these three primary drivers.”
- The 3 Red Flags: Spotting a Regulatory Disaster Before it Starts
Not all black boxes are created equal. Some are just complex; others are ticking time bombs. Raheel identifies three specific “Red Flags” in a data pipeline that suggest a project is heading toward an ethical or regulatory wall.
RG: “If the team can’t explain why a specific variable is being used, it’s a red flag for disparate impact, since that opens up the model to direct or proxy based biases. Another red flag is no intentional separation of the model’s inputs and outputs. If the predictions made by the model are being fed back in, then there is a high risk of an echo chamber effect making things worse not better. And the last one is surprisingly not technical, but cultural. If there is no human intervention in making decisions based on the output, the consequences can be astronomical especially when the decisions involve high impact outcomes like denying insurance claims for example.”
He then shared some examples that help bring these to life in real world scenarios.
Red Flag #1: The “Proxy Variable” Trap
When a model is told not to use protected classes (like race or gender) but finds “proxies” (like zip codes or shopping habits) that replicate the same bias.
The PM’s Duty: Audit the feature set for “redundant encodings” that smell like bias.
Red Flag #2: The Feedback Loop
When a model’s predictions influence the future data it learns from. (e.g., An AI predicts a neighborhood is “high crime,” leading to more policing, which leads to more arrests, which “confirms” the AI’s bias).
The PM’s Duty: Look for “degenerate feedback loops” where the model is essentially grading its own homework.
Red Flag #3: The Lack of “Human-in-the-Loop” Circuit Breakers
If the model can execute a high-impact decision (like cutting off a user’s access or denying a claim) without a clear path for human intervention or “right to explanation” requests.
The PM’s Duty: Ensure there is a “Kill Switch” or a manual override for every automated decision.
- The EU AI Act: From “Fine Print” to “Front and Center”
By 2026, the EU AI Act has categorized AI into levels of risk. Most enterprise-level data analytics fall into the “High Risk” category, requiring:
- Traceability of results.
- Detailed documentation.
- Human oversight.
This has changed the PM’s daily stand-up. We are no longer just asking “Is the code done?” We are asking “Is the documentation audit-ready?”
- Becoming the “Chief Ethics Officer” of the Program
If you are a Program Manager today, you are the last line of defense. You sit at the intersection of the tech teams of SWEs, Data Science, Design, etc (who wants to innovate), the Legal Team (who wants to protect), and the Customers (who want to be treated fairly).
Raheel argues that the best PMs in 2026 are those who aren’t afraid to “break” the project if it doesn’t meet the “Explainability” bar.
RG: “Being a leader in this space means knowing when to say: ‘This is 5% more accurate, but it’s 100% less explainable. We aren’t pushing it live.”
- Conclusion: The Future is Transparent
The “Black Box” isn’t going away—the math is only getting more complex. However, the management of that black box is becoming more transparent. The successful Data Analytics PM of the next decade won’t be the one with the fastest model, but the one with the most “accountable” model.
As we navigate the complexities of 2026 and beyond, remember: If you can’t explain the why and the how, the what doesn’t matter.
RG: “At the end of the day your job isn’t to just be the cheerleader, but rather its most constructive skeptic. Data is powerful, but it needs us to bring the conscience. Our greatest value lies in our intuition, in noticing when the result seems right but feels wrong and then asking the questions that can’t be answered by just maths.”
