We built an AI that sometimes refuses to answer. That sounds like a questionable product decision, especially when its job is to fill in missing data.
inRecipe enriches industrial item data for standard parts and commercial components: weights, dimensions, materials, sales descriptions and intended uses. It finds and reads them in manufacturers’ documents and standards tables, and every value it fills carries its source.
In these projects, the most interesting results for me have been the fields that stayed empty. Each of them gets a reason too.
Every empty field has a reason
An empty field with its reason is a useful result in itself. It tells you what to request from the supplier, where access is needed and which item needs a closer look. It also separates the gaps another search might close from the ones it never will.
The part-number case is the one I find most telling. A part can be bought for years under a code that is not the maker’s part number at all. Purchasing works, but no document will ever match it, and searching harder for its weight will not fix an identification problem. Some answers inRecipe only proposes: a customs code stays a proposal until a person confirms it.
Where it starts to hurt
I have not met many people who get excited about an item register. Interest arrives when missing information gets in the way of their own work:
- Purchasing
The part cannot be ordered.
- Customs
A shipment waits because the customs code is unclear.
- Aftermarket
A spare-part export is delayed for want of origin data.
- Procurement
The same part sits under three codes in three plants.
- Sales
The carbon footprint for a quote stalls on an unknown weight.
- Maintenance
The last-time buy is missed because nobody saw the maker’s notice.
This is where a gap in item data becomes a business problem, often years after it was created. More on where those costs land in the hours are the smallest number.
The source travels with the value
A plausible but unsupported value hides all of this. A value inRecipe reads from a document is accepted only with its page and a quote that is checked to appear on that page, so a weight copied from the wrong product family cannot pass as verified because it looks right. A reviewer can see whether it applies to the exact component, and the next disagreement starts from the evidence, not from memory.
Start with the gaps that get in the way
An industrial item register can hold millions of rows, and fixing all of it at once is the wrong place to start. Start with the fields whose absence gets in the way most: what delays purchasing, what keeps logistics asking, what someone researches again for every quote. Enrich what the sources support, confirm the proposals, and work the rest from their reasons.
Sometimes the most useful answer an AI can give is: “I could not verify this. Here is why.”
- An empty field with a reason is a result: it says what to ask the supplier, where access is needed and which item needs a closer look.
- Some gaps are not search problems: a code a part was bought under for years may not be the maker's part number.
- Missing item data becomes a business problem when it stops someone's work.
- Start with the fields whose absence gets in the way most, not with the whole register.
Further reading
More than 20 years turning data and AI into business value and working solutions. CEO of Data Design, previously at Silo AI. At inRecipe he leads the product and the Starts.
