How Window Maker Validated AI-Powered ERP Analytics Before Investing in Full-Scale Implementation

Rather than proposing a large platform up front, we proposed a tightly scoped proof of concept. We agreed on a small, well-defined subset of the ERP schema and a focused set of everyday analytical questions to target.

We defined a canonical schema with a whitelist of allowed tables and columns, plus a metadata layer of human-readable aliases that map business terms to the underlying database structure. The non-negotiable design constraint was safety: every generated query had to be read-only and parameterized, restricted to whitelisted tables and columns, so the POC could never modify or over-expose production data.

The goal was deliberately narrow — prove the end-to-end flow from a plain-English question to a correct, safe, data-backed answer, and learn what production would really demand.

We ran the POC against the agreed ERP subset with a small pilot group of business users. We mapped the real questions people actually asked the reporting team and tested whether the assistant could answer them correctly and safely. The assistant ran alongside existing reporting throughout, so its outputs could be validated against trusted, manually produced numbers without any risk to live operations. The entire engagement was delivered in three weeks.

Within the scoped use cases, the assistant reliably translated everyday business questions into correct, read-only SQL and returned answers as tables, simple charts, and plain-language summaries — confirming that the core concept was both technically feasible and genuinely useful. Just as importantly, running it against real questions surfaced the open points that always decide whether an AI analytics system can be trusted in production: Agreed, governed business definitions — what exactly counts as “revenue”, “an order”, or “a closed opportunity” — so metrics stay consistent across the business. Validation and accuracy controls to catch and prevent misleading or incorrect AI-generated numbers. Confidence-aware responses, so the system signals when it is unsure rather than guessing. Robust error handling for ambiguous questions and edge cases. Governance and access enforcement as the data scope widens beyond the POC subset. Surfacing these points early — on a contained POC rather than mid-rollout — was a major part of the value. They became the concrete requirements that shaped the full implementation.

With feasibility proven and the open points understood, the client asked us to prepare a proposal for full production implementation, focused on the Sales Module. The proposed engagement — the AI-Powered Intelligent Sales Analytics & Query Agent — moves the work from “can this work?” to “make this production-grade.” It targets sales data specifically: leads, opportunities, pipelines, orders, customers, and revenue metrics. Critically, the scope is built directly around the open points the POC exposed — adherence to approved sales definitions and business rules, validation and accuracy controls to eliminate incorrect or misleading metrics, governance enforcement, confidence-aware responses, and robust error handling — so leadership can treat the output as decision-grade. The engagement is structured as a fixed-cost, milestone-based delivery, with clearly defined deliverables and acceptance criteria for each milestone. The proposal is currently under review with the client.

The proposed production engagement focuses on hardening, scaling, and deploying the solution for full Sales Module usage: Harden and scale the assistant for full production usage within the Sales Module. Encode approved sales definitions and business rules so every metric is consistent and trusted. Build validation, accuracy, and confidence controls to eliminate misleading AI-generated numbers. Enforce governance and access controls across a wider data scope. Expand from the Sales Module into additional ERP areas once production trust is established.

Window Maker didn’t begin with a large, speculative AI investment. They began with a focused, low-risk proof of concept that answered one question: can AI turn our ERP data into safe, self-serve answers? The POC proved that it can — and, just as valuably, it revealed the specific governance, accuracy, and definition challenges that must be solved before scaling. That clarity is exactly what a proof of concept should produce. By validating feasibility and surfacing the open points first, Window Maker moved into a full implementation proposal with a clear-eyed understanding of what production-grade really requires — and a scope built to deliver it. The project shows that meaningful AI adoption doesn’t start by automating everything at once; it starts by proving value safely, then scaling with confidence.