We didn't set out to build an enterprise AI practice. We backed into it. And the reason it works is that about 80% of enterprise AI is just eCommerce infrastructure with different labels. Product catalogues become knowledge bases. Search becomes retrieval. Payment gateways become API orchestration. We've been building this stuff for over a decade - we just didn't know what to call it yet.
What You Need to Know
- eCommerce platforms and enterprise AI systems share foundational patterns: structured data, search, integration, and recommendation
- Teams with deep eCommerce experience can transition to AI delivery faster than teams learning these patterns from scratch
- The hardest parts of enterprise AI aren't the models - they're the same problems eCommerce solved years ago: data quality, integration reliability, and search relevance
The Overlap Nobody Talks About
When people think about enterprise AI, they picture machine learning engineers training models. That's maybe 10% of the work. The other 90% is infrastructure that eCommerce developers have been building for years. We've written about why enterprise AI projects fail, and it's rarely the models.
Hassan spent six years at PB Tech, New Zealand's biggest tech retailer. He built product search with Apache Solr, managed a catalogue of tens of thousands of SKUs, integrated payment gateways, and kept REST APIs running at scale. At commercebuild, he did the same thing with Elasticsearch and Laravel, adding payment processing and inventory management into the mix.
Every one of those skills maps directly to enterprise AI.
Product Catalogues and Knowledge Bases
An eCommerce catalogue is a structured data system. Products have attributes, categories, relationships, and metadata. You need consistent data models, validation, and the ability to update records without breaking downstream consumers. You need to handle duplicates, merge records, and maintain data quality at scale.
A knowledge base for enterprise AI is the same thing. Documents have attributes, categories, relationships, and metadata. You need consistent chunking strategies, validation, and the ability to update the knowledge base without breaking retrieval quality.
Hassan built catalogue management systems that handled product variations, multi-currency pricing, and supplier data feeds from dozens of sources with inconsistent formats. That's exactly the kind of data wrangling that enterprise AI knowledge bases require.
Search to Vector Search
This is the most direct mapping. Solr and Elasticsearch use inverted indices, relevance scoring, faceted search, synonyms, and boosting. Enterprise RAG systems use vector embeddings, similarity scoring, filtered retrieval, and re-ranking. The concepts are different, but the problems are identical: given a query, find the most relevant results from a large corpus.
When I first worked with vector search, it felt familiar. Different engine, same problems. How do you handle typos? How do you weight recent results? How do you deal with ambiguous queries? I'd solved all of those in Solr and Elasticsearch already. The answers were different but the questions were the same.
Hassan Nawaz
Senior Developer
Hassan spent years tuning Solr relevance at PB Tech. Search isn't a solved problem - it's an ongoing calibration. You adjust weights, add synonyms, tweak boosting rules, and test with real queries. RAG retrieval is exactly the same iterative process. The teams that treat it as a one-time setup get poor results. The teams that treat it like search tuning get good ones.
Payment Gateways and API Orchestration
Payment gateway integration is brutal. Every provider has a different API, different authentication, different error codes, different webhook formats, and different sandbox behaviours that don't match production. You learn to build abstraction layers, retry logic, circuit breakers, and idempotency checks. You learn to never trust a third-party API to behave consistently.
Enterprise AI is the same. You're calling OpenAI, Anthropic, Azure, and custom model endpoints. Each has different rate limits, different token counting, different error responses, and different latency profiles. The abstraction patterns Hassan built for payment gateways at commercebuild - provider interfaces, retry with exponential backoff, graceful degradation - are the same patterns we use for AI model orchestration.
Recommendations and AI Suggestions
"Customers who bought X also bought Y" is a recommendation engine. It uses purchase history, browsing behaviour, and product similarity to suggest relevant items. Enterprise AI suggestions work identically: given a user's context, history, and current task, what information or action is most relevant?
The data pipeline is the same. Collect user behaviour, process it into useful signals, feed it to a model (collaborative filtering then, LLMs now), and present the results in a way that's useful without being intrusive.
Why This Matters for Hiring
If you're building an enterprise AI team and you're only looking at candidates with "AI" or "ML" in their background, you're missing the largest pool of relevant experience: eCommerce developers. They've already solved the hard infrastructure problems. Teaching them about embeddings and prompt engineering takes weeks. Teaching someone those skills while also teaching them about data pipelines, search relevance, API integration, and production reliability takes months.
We didn't plan this transition. But when we started doing enterprise AI work at RIVER, Hassan's eCommerce background meant he could build production-grade data pipelines and integration layers from day one. The AI-specific parts - model selection, prompt engineering, retrieval strategies - were a small addition to a large existing skill set.
The Lesson
The best preparation for enterprise AI wasn't studying machine learning. It was building eCommerce platforms that had to work reliably, at scale, with messy real-world data and unreliable third-party integrations. That's what enterprise AI actually is, once you strip away the branding.
If your organisation is exploring enterprise AI and wants a team that's already solved the hard parts, let's have a conversation.