All resourcesPractical guide

Reading supplier documents with AI: "hard", "soft" and the grey zone between

Published on 3 November 2026 · 2 min read

The automatic migration of supplier documents by AI — spec sheets, certificates, specifications in PDF or Excel — has gone into production in food & beverage. And the field feedback surfaces something more useful than a plain "it works": a typology of data by how you can, or cannot, verify it.

Three tiers emerge. Blurring them means either over-controlling or letting errors through. Distinguishing them makes validation more reliable without paralysing it.

"Hard" data: strict verification

Some information tolerates no approximation: an additive's CAS number, an inclusion dose, the presence of an allergen, a nutritional value. They engage data integrity and, sometimes, safety. Their verification is hard: controlled format, cross-checking against references, arithmetic consistency.

For these fields, the AI proposes, but validation is strict and automatable: a value that fails the check is blocked, not assumed.

"Soft" data: human judgment

Other information is, by nature, floating: an organoleptic description, a usage recommendation, a free-text marketing mention. It belongs to human judgment and lends itself poorly to a binary rule.

Trying to verify it "hard" produces fake control. The right reflex is to treat it as text to keep and contextualise, not as data to certify.

The grey zone: compiled by AI, decided by a human

Between the two lies the heart of the matter: information the AI compiles or interprets — matching an ingredient to a reference, deducing a category, consolidating a specification from several sources. Neither fully certain nor purely subjective.

This is where the alert level earns its place: the AI attaches a confidence index to each compiled value, and escalates doubtful cases to a human. The gain is not to remove human control, but to concentrate it where it truly decides.

What this means for your product data

Making supplier migration reliable is neither blind trust in the AI nor re-checking everything by hand: it is routing each field to the right validation regime — hard, soft, or to arbitrate with an alert.

A PLM that carries this distinction at field level, traces the source and confidence of every value, and escalates edge cases turns ingestion into a controlled flow. That is crumble-ai's approach: the AI does the heavy lifting, the human decides where that is their role.

Is your portfolio ready?

See how crumble-ai continuously audits your recipes and blocks non-conformities at the source.

Book a demo