What Resolve AI taught me about asking the right question

One of the most useful moments in building Resolve AI came when we had to rethink the problem. We started with a broad ambition around service problems and customer disengagement across Orange's digital ecosystem. It sounded promising, but we still needed a precise answer to a basic question: what decision would our product help someone make?

I worked on Resolve AI as a software developer and assistant team lead during the Orange Summer Challenge 2026. Alongside the development work, I helped research the case study, refine the problem statement, prepare presentation material, and support the team through the pitch process.

Narrowing the first use case

Feedback from Orange pushed us toward one journey: disengagement from data bundles. That gave the prototype a clearer starting point. We could focus on a change in a customer's activity, investigate what might explain it, and recommend a useful follow-up.

This change affected the product requirements as much as the presentation. A focused use case made it easier to decide which information mattered, what the interface should show, and what a successful demonstration needed to prove.

A behavior change needs an explanation

A decline in bundle activations tells us that something changed. It does not tell us why. A customer might be spending less, experiencing coverage problems, using another service, or dealing with a lost phone or SIM. Treating every decline as the same problem would lead to poorly matched responses.

That distinction became central to our thinking. Behavioral information could help identify where to look. Customer feedback and other relevant evidence would help us understand the situation before recommending action.

Giving AI a specific job

Our proposed workflow moved from behavior detection to feedback classification, then to a recommended next action and an outcome record. We considered information from customer calls and other permitted feedback channels as possible inputs. Those inputs would still need appropriate access and review before any real deployment.

The AI component had a defined role: help organize feedback, summarize it, and make the recommendation understandable. Staff would review the proposed action. We wanted the product to work with Orange's existing communication channels and processes, so the team using it could connect a recommendation to an actual intervention.

Being honest about the prototype

We used synthetic data and simulated integrations for the MVP. That let us demonstrate a workflow, but it could not establish real churn prediction accuracy or a measured improvement in customer retention. A production version would need approved data, validated assumptions, and evaluation against actual outcomes.

This mattered during the pitch. A convincing demo should make its assumptions clear. Otherwise, a polished interface can make an untested idea look more certain than the evidence supports.

The lesson I am keeping

Resolve AI placed second, and our team received a $1,500 prize. I am proud of that result, but the product lesson will stay useful long after the competition.

I now pay closer attention to the decision behind a feature: who needs it, what evidence they need, and what they can do next. The project became stronger when we were willing to narrow the scope and explain the limits of what we had built.