Palate in the Machine | #PSC 247
OVERVIEW: Episode 247 takes a look at the uses of AI tools in confectionery recipe formulation (and reformulation). What can AI tools tell us about human taste? Bueller? Bueller? [updated]
Timing and Links to Watch & Comment
Links below to watch LIVE and to view the archived episode.
Watch and comment LIVE or view the archived episode!
Subscribe (free!) to the @PodSaveChocolate YouTube channel, like this video, comment, and share this episode to help grow the #PSC community.
Watch and comment LIVE or view the archived episode on LinkedIn.
Join my network on LinkedIn to receive notifications and to refer business to each other.
Watch and comment LIVE or view the archived episode on TheChocolateLife page on Facebook (for 30 days, then watch the archive on YouTube).
Follow TheChocolateLife on Facebook to receive notifications and catch up on other content.
Episode 247 Overview
AI is now in every link of the cacao «» chocolate supply web, from pod disease to gift recommendations. The most useful (grading, fermentation, inspection) get the least press.
The most heavily promoted uses (formulation and retail) are the least transparent.
In this episode I will be taking a look at the uses of AI in confectionery formulation and reformulation.
The TL;DR
The story is not whether AIs can taste chocolate. (They cannot, yet.)
The story is that tools for predicting “acceptable” have arrived in an industry where “acceptable” comes with a legal definition, a price, and a margin.
- AI tools really arrived in chocolate R&D when cocoa got expensive, and that timing shapes what the tools are asked to do. Cocoa futures hit records in late 2024, and most of the reformulation work started then.
- Tools attempt to predict “acceptable.” And “acceptable” has a legal edge. Sensory models can estimate how much cocoa consumers will tolerate losing for a given cost. US and UK law sets a minimum cocoa content for “chocolate,” and Nestlé has already reformulated two UK bars (Toffee Crisp and Blue Riband) below it. Nestlé began calling the coating “milk chocolate flavour[ed],” insisting the new versions were “carefully developed and sensory tested.” Both things can be true at once, and that is the point.
- Nobody discloses training data. Cargill, Mondelēz, Barry Callebaut/NotCo, and others all say humans get the final say. None says what panels, markets, or products trained the model(s).
- Formulation models start at the factory door. Formulation AI learns from consuming-country panels. No formulation model (at least among those found in this research) uses farm, fermentation, or producing-country tasting data.
- The best upstream work is being done in producing countries. Teams in Côte d’Ivoire and Ghana are building swollen-shoot detection datasets. Cut-test AI now passes 90% defect accuracy in one commercial tool. But Mars’s own bean-scoring app is closed to outsiders.
- The flavor houses hold the deepest data. Givaudan, Symrise, and dsm-firmenich run AI on sensory and receptor-level data for sugar reduction and flavor design. That work reaches the consumer only as “flavoring” on a label.
- Factory, inspection, and supply AI is real and quiet. Sorting cameras, line control, bloom detection, satellite crop-risk readings, and EUDR deforestation checks are in daily use. They carry the most checkable claims.
- “Evidence” is mostly self-reported. Outside peer-reviewed upstream studies, nearly every number comes from the company or vendor. At least one widely circulated AI claim has no source at all.
Until proven otherwise, every claim of AI-assisted “innovation” – as of the time of writing – is worth reading as a claim about cocoa content.
Evaluating Claims: Five questions to ask
- What data trained the model, and from which markets and panels?
- Did the panel test acceptance or detection of a change?
- Is there a named customer, a date, and a comparison baseline?
- Has anyone outside the company tested it?
- What changed on the label?
Who Claims What?
Thirty companies and programs, grouped by where in the chain the AI sits, found during the course of researching today’s episodes.
Company or program | What they say AI does | Chain stage | Chocolate-specific? | Training data disclosed? | Date of claim |
|---|---|---|---|---|---|
Deep Cacao (Côte d’Ivoire) new | Early CSSVD detection from field images | Farm | Yes | 8,000+ geotagged images; national dataset | 2025 |
Cocoa of Excellence / Alliance Bioversity-CIAT new | Cut-test grading from images | Post-harvest | Yes | 1,000+ samples, 55+ origins | 2023 |
QVisions / Rausch new | Automated cut-test defect detection | Post-harvest | Yes | “Thousands of images” from key regions | 2025 |
Mars Cocoa Fresh AI new | Bean-quality scoring app | Post-harvest | Yes | No; Mars users only | 2025–26 |
Treefera new | Satellite and AI crop-risk readings | Supply | Yes (cocoa coverage) | Satellite, weather, proprietary science | Nov 2026 |
EUDR platforms (TraceX, Dimitra, ESA) new | Deforestation and plot verification | Supply | Cocoa among commodities | Satellite imagery; methods vary | 2025–26 |
Cargill | Sensory modeling for flavor, texture, melt; AI sorting cameras | Formulation; factory | Yes | Soft-cake study: 8 recipes, 5 segments; no sample sizes | Oct 2026 |
Mondelēz | Flavor combinations, ingredient interactions; recipe tool for Cadbury, Toblerone, Oreo | Formulation | Brands named; published results are biscuits | Internal R&D and consumer-test history, no detail | Nov 2026 (and 2023, 2025) |
Barry Callebaut / NotCo | Giuseppe platform in BC’s data cloud; HFSS and Nutri-Score work | Formulation | Yes | “Proprietary dataset… over 10 years,” no detail | Nov–Dec 2025 |
NotCo (unnamed clients) | Cocoa-free, sugar-free HFSS, cost-optimized chocolate | Formulation | Yes | No | Undated |
AKA Foods | Retrieves past formulations and sensory data; does not reformulate automatically | Formulation | Pitched for chocolate | Customer’s own data, kept private | Dec 2025 |
Hill Country Chocolate | “Ganache AI” predicts ingredient interactions | Formulation | Yes | Not disclosed; accuracy unsourced | Aug 2026 |
Mars | LEAP (PIPA) knowledge graph for ingredient discovery | Ingredients | No | Literature, omics, Mars data | Jan 2026 |
Nestlé | Natural colors; packaging materials with IBM | Ingredients; packaging | No | Chemical language models (packaging) | 2025–26 |
Givaudan ATOM new | Flavor formulation, sugar reduction, vanillin replacement | Flavor supplier | Confectionery among uses | “Two decades of research,” no detail | 2021 |
Symrise Symvision AI new | Sensory and formula prediction, reduced sugar | Flavor supplier | Not stated | Proprietary sensory and consumer studies | Nov 2025, Oct 2026 |
dsm-firmenich / Boltz new | Designs flavor molecules against taste and odor receptors | Flavor supplier | Not stated | Proprietary receptor library | Sep 2026 |
Gastrograph AI (NIQ) new | Predicts demographic preference from small panels | Consumer insight | No | “Largest sensory dataset,” 20+ countries | Ongoing |
Celleste Bio new | Selects cell lines; plans tunable cocoa butter | Cultured cocoa | Yes | No | Dec 2024, Apr 2026 |
COCO-AI (EU) | AI agent optimizes cocoa cell growth | Cultured cocoa | Yes | Project record only | May 2026 |
Bühler new | Data platform and automated refining control | Factory | Yes | Line sensor data | 2025–26 |
Scorpion Vision, Keyence new | Tray, bar, and defect inspection | Packing | Yes | Customer images | 2023–26 |
Lotte | Chose calamansi flavor for chocolate Toppo | Formulation | Yes | Consumer preference data | Dec 2017 |
Meiji | Kinotake Mother preference AI; KAORIUM scent recommendations | Formulation; retail | Yes | 1.67M preference points; 12 aromas | Jan, Apr 2026 |
Valio / Kultasuklaa | Social-listening AI shaped a 30%-less-sugar bar | Formulation | Yes | 1.5M+ social posts | Nov 2022 |
Ferrero (Nutella Unica) new | Algorithm generated 7M unique jar designs | Packaging; marketing | Spread | Pattern and color library | 2017 |
Mondelēz (Cadbury India) new | Generated celebrity likeness in localized ads | Marketing | Yes | Store location data | 2021 |
Hershey | AI-directed ad spend | Marketing | No (not R&D) | Sales and media data | 2024, May 2026 |
Hotel Chocolat (Mars) | AI gift assistant | Retail | Retail only | No | Nov 2026 |
Foreverland | Unverified. AI claim appears only in an unsourced blog post | n/a | n/a | n/a | Mar 2026 |
Cargill: “complements, does not replace”
On October 26, FoodNavigator and ConfectioneryNews ran the same Cargill story, “Beyond cocoa: The race to futureproof chocolate.” Cargill says AI-powered sensory modeling predicts how a formulation change will affect flavor, texture, creaminess, and melt. Michiel Van der Bom adds that the tool “complements, and does not replace, consumer testing and expert tasting.”
Cargill’s own product page credits “statistical modeling and predictive analysis tools” and never says AI. Its headline number comes from a soft cake study: eight recipes, five consumer segments, and 90% of consumers happy with a coating instead of chocolate.
There are no per-segment sample sizes, no full method, and no independent check. The panel was asked whether people would accept a substitute, which is not the same as asking whether they would notice one.
Mondelēz: “real humans still get the final say”
In a November 12 interview, chief R&D officer Norberto Chaclin said AI evaluates flavor combinations, predicts ingredient interactions, and identifies functional ingredients. Products still go through human sensory testing.
This is not new. Joe Manton, global lead for digital R&D, described in 2023 how testing began in 2019, trained on historical data about concepts, recipes, ingredients, and consumer tests. This year Mondelēz reportedly added agentic AI from PRS In Vivo to run multi-step development workflows.
Mondelēz names Cadbury and Toblerone among the brands its recipe tool supports. But the result it publicizes, 60% of recipes doing better on nutrition, sustainability, and cost, refers to biscuits.
Barry Callebaut and NotCo: the most specific supplier
Barry Callebaut partnered with NotCo AI in late 2025. NotCo’s platform, Giuseppe, is “trained on a unique proprietary dataset of food and consumable properties accumulated over 10 years.” Neither company has said what is in it.
Per BakeryAndSnacks, Giuseppe runs inside Barry Callebaut’s own data cloud and can flag UK HFSS problems early. NotCo’s case-study page lists a cocoa-free chocolate with “zero cocoa and 40% less sugar, delivered in 4 months,” a sugar-free HFSS chocolate, and a cost-optimization project. No clients are named.
AKA Foods: the modest one
AKA Foods says its AKA Studio platform can shorten reformulation from “years to weeks” (BakeryAndSnacks, also run by FoodNavigator). Every performance claim comes from the company, with no named customers.
Yet CEO David Sack also said the software “doesn’t reformulate products automatically.” It helps R&D teams find old formulations, sensory results, and project files. That may be the most honest description of AI in food R&D on record.
Mars, Nestlé, Hershey, Ferrero, Lindt: less than the headlines
Mars uses LEAP, a knowledge-graph research assistant from UC Davis spin-out PIPA, to discover ingredients. Darren Logan of the Mars Advanced Research Institute told FoodNavigator that AI “significantly improves hit rate” in some cases. He gave no figures, and the work concerned plant-based and bioactive ingredients, not chocolate.
Nestlé’s public AI claims this year cover natural colors and packaging with IBM Research. Hershey uses AI to decide where ad dollars go, not what goes in a bar. The original research found no public claim from Ferrero or Lindt on formulation.
Hill Country Chocolate: a craft maker that sounds like Mondelēz
Hill Country Chocolate, a bean-to-bar maker in Fredericksburg, Texas, describes a system it calls “Ganache AI.” It claims 85–90% first-trial accuracy and development in “weeks instead of months,” with no method cited.
It also says “this isn’t about replacing the chocolatier’s palate.” When a small Texas maker and the maker of Cadbury reach for the same reassurance, it is worth asking who they are reassuring.
The fair case for speed
Speed is not a vice. Sensory science has long used trained panels, and a model that narrows the field before the panel meets is a legitimate tool. If AI keeps flavor intact while a recipe adapts to a tougher supply, eaters win.
The test is the label. Same model, same panel, opposite result depending on intent. The only way to tell which happened is to compare this year’s ingredient list with last year’s.
Future Episodes
I will be traveling for the Thanksgiving holiday.
Topic TBD
Episode Hashtags
#cocoa #cacao #cacau
#chocolate #chocolat
#PodSaveChoc #PSC
#LaVidaCocoa #TheChocolateLife
The Archives
To read an archived post and find the links to watch archived episodes, click on one of the bookmark cards below.

Episodes 201-300 (links to older PSC archives are in the top nav).




