Reducing Installation Risk: From Survey Data to Predictive Intake

The company relied on customers to self-identify wall types, measurements, room dimensions, and installation suitability. Those inputs were inconsistent and created clarification work, failed installations, repeat visits, cancellations, and customer frustration.

• About a third of surveys failed automatically or required clarification before an installation could be scheduled.
• Two conditions accounted for most of the failures, and both required construction knowledge to answer correctly.
• One failed survey could contain multiple failure conditions, so the failure percentages overlap and are not mutually exclusive.
• Failure concentrated in specific markets, which pointed to housing stock rather than customer carelessness.
• The root problem was not that customers completed the survey incorrectly. The process asked untrained customers to make construction and measurement judgments that often required specialized knowledge.

1. 01
   Stage 1 · Diagnose

   Full survey pipeline analysis

   Every survey in the analysis window was classified by pass, fail, and failure condition, with legacy records excluded because they lacked consistent pass and fail logic.

2. 02

   Root-cause breakdown

   Failure conditions were counted as a share of failed surveys, with the overlap made explicit so the drivers would not be double counted.

3. 03

   Fulfillment and cancellation analysis

   Survey outcomes were traced through to fulfillment status and installation-related cancellations to connect intake quality with business impact.

4. 04
   Stage 2 · Operationalize

   A knowledge-backed operations assistant

   I created and launched a custom, knowledge-backed GPT to support Order Operations and installation-feasibility decisions, grounded in available internal materials rather than general web knowledge.

5. 05

   Structured rollout and handoff

   Configuration, knowledge loading, sample-question testing, refinement, gap filling, a user walkthrough, usage monitoring, and SOP updates based on the questions the assistant could not answer.

6. 06
   Stage 3 · Scale

   A predictive intake concept

   A proposed intake product: customer-submitted photo or video, automated wall-material classification, AR-assisted measurement, confidence scoring, structured risk data, a recommended installation pathway, installer verification, and market-based risk prioritization.

7. 07

   Implementation planning

   A product requirements document, functional and non-functional requirements, user flow, technical dependencies, prototype and pilot and rollout phases, budget estimates, risk mitigation, and success metrics. This was not built.

The problem was not customer training. It was system design.

Rather than wait for an engineering build, I created and launched a custom, knowledge-backed GPT to support Order Operations and installation-feasibility decisions. It was grounded in internal installation requirements, SOPs, escalation workflows, operational plans, known installation risks, and existing team knowledge. It reviewed installation information and customer-submitted images to assist with feasibility assessment. The assistant went live and was used as an internal decision-support resource, although adoption remained limited.

• Configure the custom GPT and system instructions
• Load the relevant operating knowledge
• Test sample questions for accuracy, tone, and clarity
• Refine responses based on early testing, then add missing workflows and materials
• Conduct a user handoff and walkthrough
• Establish an expectation to consult the assistant before escalating routine questions to a manager
• Monitor usage and answer quality, and capture unanswered or recurring questions
• Update SOPs and instructions based on identified gaps
• Formalize the tool as an operations assistant and audit whether it helped the user operate more independently

Combined completed operational analysis with a live AI-assisted decision-support tool, then translated the remaining system limitations into an implementation-ready product concept for predictive installation intake.

• The analysis reframed the problem. The intake process was asking an untrained customer to make construction and measurement judgments that required specialized knowledge.
• The assistant made internal knowledge accessible immediately, with no engineering dependency. It still relied on a user actively consulting it and interpreting its guidance.
• The predictive intake concept was documented to a level a build team could pick up, but it was not built.
• Projected reductions in wall-type misclassification, installation failure, and repeat visits were planning targets for the future-state product. They are not measured results, and the assistant did not deliver them.