TL;DR. Nestlé’s AI runs on a single global ERP and a recipe-specification library it spent six years harmonizing. Mid-market food, beverage, supplement, and pet food companies — those at $50M–$500M in revenue — cannot replicate that path and shouldn’t try. The asymmetric advantage for mid-market: large language models collapse the data-harmonization step that took Nestlé six years. A $100M brand can now get to working AI on supplier and procurement data faster than a $100B brand did.
What Nestlé actually built
Nestlé’s R&D group has built a recipe optimization tool that balances ingredients, nutrition, cost, and sustainability for product developers. Product ideation has compressed from three months to three weeks. The tool has generated approximately 1,300 product ideas, with around thirty currently in development.
That story usually gets told as an AI story. It is mostly a data story.
The infrastructure underneath the recipe optimization tool includes:
- A single global ERP system running across markets, with a unified data foundation
- Recipe-specification harmonization that has reduced specifications by approximately 50% since 2019 (and around 10% since 2022) — coordinated across R&D ingredient experts, material-recipe managers, and procurement teams
- Digital dashboards for raw-material requirements feeding ingredient sourcing decisions
- A multi-year IBM Research partnership for machine-learning-driven packaging materials discovery
- Selection into the Frontier Firm AI Initiative with Harvard’s D^3 Institute and Microsoft
The recipe optimization tool sits at the top of that stack. Without the ERP backbone and the harmonized spec library underneath it, the tool would have nothing to optimize against.
The mid-market stack looks nothing like that
If you run procurement, QA, or R&D at a $50M–$500M food, beverage, supplement, or pet food company, your reality looks different:
- You’re on NetSuite, Sage, QuickBooks Enterprise, or a fragmented set of systems — not a single global ERP
- Your spec sheets live in three places, in three formats, with three sets of conventions
- Procurement, QA, and R&D each email the same suppliers independently. No one sees what came back
- Your largest current “AI investment” is whatever your CFO is using to draft board decks
This is not a deficiency. It is the rational state of a $200M company. Mid-market food brands don’t have $100M to spend on six-year ERP transformations, and they don’t need to. What they need is software that delivers results without first requiring a multi-year data-harmonization project.
Why the Nestlé playbook breaks at mid-market scale
Three structural reasons large-CPG AI doesn’t translate downstream:
1. The capital ratio is different. Nestlé spends well over a billion dollars per year on R&D. A $200M brand might spend $4M. A model that requires a 50-person data science team is not a $200M company’s tool.
2. The data infrastructure is different. Roughly 70% of food and beverage organizations cite lack of qualified AI staff as a barrier to adoption. That figure is real but downstream of the deeper problem: legacy mid-market systems were never designed to be the data backbone for ML pipelines, and harmonizing them is a multi-year capital project.
3. The use cases are different. Nestlé’s AI optimizes recipes across thousands of SKUs. Your team is managing fifteen active formulations. The problem isn’t generating 1,300 product concepts. The problem is confirming whether your secondary supplier for organic cane sugar still has a current SQF certificate.
The shortcut: LLMs collapse the harmonization step
Here is what changed in the last eighteen months.
Before LLMs, AI required structured data. To get value, you first had to clean, harmonize, and stage your data in a format a model could consume. That is the six years Nestlé spent. That is why a $200M company couldn’t follow.
LLMs read unstructured input directly. A spec sheet PDF, a supplier email, a CoA scanned upside down, a forwarded reply with a price list embedded three messages deep — all of it is now extractable into structured data at sufficient accuracy for procurement, QA, and R&D workflows.
This is the asymmetric advantage mid-market gets.
Email was always the system of record in mid-market CPG procurement. The internationally accepted way of moving CoAs, spec sheets, and pricing between food brands and their suppliers is the inbox — not portals, not EDI. 15 companies tried and failed to displace email with supplier portals over the last decade. They failed for the same reason: suppliers don’t know we exist. They have no incentive to log into a new system on behalf of one of their customers.
LLMs let mid-market work with the data layer they already have. No portal. No supplier behavior change. No six-year transformation.
What mid-market procurement, QA, and R&D leaders should focus on
If you are evaluating AI at a $50M–$500M food, beverage, supplement, or pet food brand, the practical focus list looks like this:
Skip:
- Computer vision QA on production lines (capex-heavy, requires line redesign, often impossible at co-mans)
- Custom demand-forecasting models (need years of POS data and an in-house data team)
- Dynamic pricing engines (need an SAP backbone you don’t have)
Focus:
- Document extraction from supplier emails — CoAs, spec sheets, allergen statements
- Certificate expiration monitoring — GFSI, SQF, BRC, organic, kosher, halal, HACCP
- RFP normalization across supplier responses with different units, IncoTerms, and payment terms
- Cross-functional supplier coordination across procurement, QA, and R&D
The pattern: pick AI that works on the data you already have, not AI that requires you to build clean data first. That is the coordination tax — the invisible cost CPG companies pay when R&D, QA, and procurement all email the same suppliers independently with no shared record of what came back. It is also the highest-leverage place to apply LLM-based tools today.
The bottom line
Nestlé’s AI works because Nestlé spent six years building the data foundation underneath it. Mid-market brands cannot copy that path and shouldn’t try.
The right path is to look at where your institutional knowledge actually lives — in inboxes, in shared drives, in supplier email threads — and apply AI that reads it as-is. That is a different category of AI than the one Nestlé built, and it is the right one for $50M–$500M food, beverage, supplement, and pet food companies.
You don’t need a single global ERP. You need software that reads your inbox.
