AI Project Risk
Nobody Acts on a Confidence Score
Overview
I took a predictive model that could flag at-risk projects and designed around the thing it couldn't do: explain itself. By translating raw output into interpretable signals managers could reason about, the feature surfaced 90% more critical risks and cut project delays about 30%, showcasing my ability to make probability based systems legible enough to act on.
Problem
A raw prediction ("this project is 73% likely to slip") is a conversation killer. PMs can't act on a number they can't interrogate. The design problem wasn't displaying risk, it was answering "why is this flagged, and what do I do about it?" without requiring anyone to understand the model.
Constraints
Every signal had to be explainable. No black-box scores.
Flagging more risks could lead to alert fatigue; more wasn't automatically better.
Risks had to be actionable. Surfacing them wasn't enough.
Research
Why: Managers were finding out about project risk too late. The data to catch it earlier already existed, scattered across systems nobody was reading in time.
Methods: Contextual sessions with 6 project managers to understand how they currently detected trouble and what made them act versus wait. I also worked closely with the data and engineering teams to understand what the model could and couldn't say confidently, which shaped the design as much as the user research did.
What we found: Managers didn't lack signals, they lacked reasons to believe them. A score or probability with no explanation behind it got ignored, because acting on it meant making a case to someone else that they couldn't support. The question wasn't "is this project at risk," it was "can I defend escalating this."
What changed because of it: I designed the interface around interpretable signals rather than model output. Each flag shows what's at risk, which factors contributed, and what action is available, so a manager can evaluate it rather than take it on faith. The model didn't get more accurate. People just started trusting it enough to act, which is where the 90% increase in surfaced risks came from.
Key Decisions
Signals, not scores. Instead of exposing raw predictions, each risk renders as an interpretable signal: what triggered it, and a suggested action.
Designing against noise. To address alert fatigue, I tiered severities with color codes, and added a risk tolerance dropdown for the PM to decide how critical risks were on a project basis.
Working with the model, not just around it. The initial model was limited, and the data science team wanted to hardcode some of the risks. I pushed back on this approach because it would hurt the feature in the long run, particularly as we wanted to scale and add more risk types. I wanted the risks to be relevant across different industries/client types, and learn from past projects. They were able to add additional models to accomplish this.
How I worked with AI
I used Claude to simulate skeptical PM questions for each risk and to pressure-test the explanation copy.
AI drafted an explanation that was technically accurate but useless to a PM. I had to validate them against PM personas.
Another designer had started this feature before I took it over. Rather than begin from zero, I used Claude to quickly generate variations on their existing designs so we could settle on a direction fast.
Try it → Open a high-risk project, expand a risk signal, and ask Dela for a mitigation plan.
Recreated demo, rebuilt from scratch in Lovable. Production data, client names, and branding removed. Interaction reference only.
The design bet here: a PM who can see why a risk was flagged will trust and act on the recommendation that follows it. A number alone asks for faith. Evidence asks for a decision.
Outcome
Surfaced 90% more critical risks than the previous manual process compared to the legacy workflow. PMs rated the risks as actionable in testing.
Teams using it saw ~30% fewer project delays within the next quarter, tracked by on-time milestone completion.
The pattern that stuck: AI that supports PM judgment instead of replacing it. Mid-market client adoption with enterprise on the way.
