98%+
Product-match accuracy, up from 80%
CASE STUDY · AI SAAS
AI quoting across a 300,000-item catalog.
An AI-native SaaS that worked, and read like a black box.
A confidential eight-person team built real AI for electrical distribution. The product functioned. But buyers could not see the value, and investors could not price the company. I rebuilt the commercial layer so the technology finally made sense to the people writing checks.
A quote can arrive as a handwritten list, a PDF or half a part number. The catalog has to resolve it before a model can help.
The problem
Matching incomplete requests to the catalog.
A distributor sells breakers, conduit, gear, and fittings across a catalog north of 300,000 items. The inputs are a mess. PDFs. Handwritten bills of material. Emails with half a part number. Vendor aliases for the same SKU. ERP records missing the attributes you actually need to quote.
The model could read those inputs. The catalog still needed resolved aliases, usable attributes and ERP context to match them to the right products.
- Product-match accuracy was 80%. That measures item matches, not the percentage of entire quotes that were correct.
- The pitch sold the AI, not the outcome. Buyers heard hype, not margin.
- Pricing was flat, so heavy AI users quietly ate the gross margin.
The number that moved
Product-match accuracy: 80% to 98%+.
Product-match accuracy
+18 pts after the data layer was rebuilt
And when there is no exact match, the system surfaces valid alternatives instead of failing silent.
Unmatched or incorrect item matches
20% → 2% or less item-match error rate
The complement of reported product-match accuracy. This is not a measured whole-quote error or human-review rate.
The signature system
Messy input in. A priced, grounded quote out.
This is the pipeline the whole business runs on. Raw distribution chaos on the left. A quote a rep can send on the right. Every stage exists to make the product data trustworthy before the model ever speaks.
Raw inputs
PDF spec sheets
Bills of material
Half-typed emails
Vendor aliases
Partial ERP records
→
NORMALIZE
Catalog intelligence
Resolve aliases, fill attributes, map every input to a canonical SKU.
→
RETRIEVE
Vector / RAG
Pull the right products and valid alternatives from the grounded catalog.
→
ASSIST
LLM quote assist
Draft the quote from retrieved catalog data for review.
→
OUTPUT
ERP-grounded quote
ERP pricing and stock context. Item-match accuracy: 98%+.
The founders sold stage three. The value lives in stage one. My job was to make the whole line legible to a buyer and an investor, then price for it.
The compounding part
The catalog work and the cost to serve.
Resolving aliases, filling missing attributes and identifying valid alternatives made the catalog usable for quoting. Those are data-maintenance tasks; the accuracy improvement does not establish that the model learns automatically from each quote.
I built usage-metered pricing around per-customer model cost, so heavier use could be billed instead of hidden inside a flat fee. That is the mechanism behind cost per customer. It makes variable costs visible; this case does not report a measured gross-margin result.
- Positioning rebuilt around the outcome, a quote you can trust, not the model.
- Catalog-maintenance strategy for aliases, attributes and valid alternatives.
- Usage-metered pricing that meters per-customer model cost to protect margin.
Where it landed
300k+
Items in the live catalog, plus surfaced alternatives
Metered
Usage-based pricing tied to per-customer model cost
8
Person team, now legible to buyers and investors
The point
My work connected the catalog and matching improvements to positioning, investor materials and usage-metered pricing. Buyers could see what the software did and how its use would be charged.
The catalog had 300,000 items. The matching, pricing and commercial explanation had to work together.