Food & Ingredients · Food Ingredients & Ingredients

AI-Powered Product Recommendations for Food Ingredients

Starting point

An international food and ingredients corporation with more than ten subsidiaries faced a scaling problem: product knowledge — formulations, application parameters, dosage recommendations — resided in the heads of individual experts and in thousands of unstructured documents. Sales representatives needed anywhere from 30 minutes to several hours per customer inquiry to provide a well-founded product recommendation. With four subsidiaries, each with its own portfolio, specialized terminology, and varying data depth, a unified solution was far from trivial.

What we did

We built AI-powered product finders in parallel for four business units of the corporation on an enterprise GenAI platform. Each agent operates with a dedicated, hierarchical dual-source system and was iteratively optimized across five prompt versions — validated through business unit workshops and structured testing with 186 data points. The finders are integrated into everyday sales operations: they read customer inquiries directly from Outlook, analyze them, recommend products, and draft the response email. Timeline: 11 months to production.

Results

business units in productive use

78 %

responses above expert level

bis 5 Std.

time saved per customer inquiry

40.000+

indexed documents (largest unit)

What we learned

An AI product finder is only as good as its data foundation — not its model. The first version delivered disappointing results because too many unsorted documents were connected. Only the reduction to curated, hierarchically structured sources produced the breakthrough. Less data, better curated, beats more data, poorly structured. This is a pattern we see across industries.

This is the summary. How we approached it methodologically — which architectural decisions we made, what we discarded and which patterns can be transferred to other contexts — we discuss in a personal conversation.

Not because we want to sell you something. But because this depth is what our clients engage us for — and it does not belong on the open internet.