AI in Food R&D: Efficiency Over Alchemy
AI in food isn't new, and it isn't a chef. It's a filter, a model, and a time-saver - when you have the data to feed it.
McCormick has used AI in flavor development for nearly a decade, reporting 20%-25% shorter timelines by surfacing promising flavor combinations and narrowing which ideas deserve a physical prototype. At Unilever, systems test thousands of recipes digitally in seconds; Knorr Fast & Flavourful Paste was developed in roughly half the usual time. AI also modeled how formulations behave in Hellmann's Easy-Out squeeze bottle, shaving months off lab work.
Even back in 2017, a Google Brain team (now part of Google DeepMind) helped craft a recipe for the "perfect" chocolate chip cookie. That tells you something: AI speeds decisions and expands search. It doesn't taste.
Humans are still the tastemakers
Inside large food companies, the message is consistent: AI is a co-creation tool. "Human creativity and judgment lead the way, and AI is a tool to help us amplify our impact," said Annemarie Elberse, who leads ecosystems, digital and data for foods R&D at Unilever. McCormick's chief science officer, Anju Rao, put it more directly: AI inspires their flavor scientists, but "our greatest asset will always be our people."
That's not PR gloss. It's a boundary condition for product teams: let models compress your search space; keep humans in the loop for target setting, trade-offs, and taste calls.
Startups promise "virtual sensory." Can they deliver?
A wave of platforms - Zucca, Journey Foods, NielsenIQ, AKA Foods, and others - pitch "virtual sensory" screening to reduce taste panels, lower failed launches, and compress cycles. On paper, that mirrors what big players say they already do: digitally triage ideas and prioritize the best physical trials.
Market estimates peg AI in food and beverage at roughly $10B in 2025, growing past $50B by 2030. But some early efforts have already pivoted. McCormick's initial work with IBM came alongside projects like Chef Watson; IBM says it's "not actively focused in this area anymore." The lesson: claims move faster than validation.
The bottleneck: data and biology
Brian Chau, a food scientist and founder of consultancy Chau Time, says many platforms are still collecting data rather than producing truly predictive outputs. "They need to attract investors, they need to build datasets, and they need real industry partners before any of this really works at scale," he said. Without proprietary formulation and sensory data from manufacturers, most tools feel like general-purpose AI with food terms sprinkled in.
The harder constraint is human perception. Dr. Julien Delarue, professor of sensory and consumer science at the University of California, Davis, is blunt: "Trying to predict what people will perceive from a complex mixture of compounds - the answer is no." People don't taste the same way. Genetics, culture, experience, and memory all bend perception. "There is no such thing as the average consumer," he said. For context on his field, see UC Davis Food Science and Technology's work in sensory science: UC Davis FST.
The fix would require granular, person-level data at scale - who tastes what, how, and why - connected to formulations and process conditions. That's a massive lift, and a tough sell for companies guarding IP.
Where AI actually helps your team today
- Hypothesis generation: Mine historical formulations, sensory notes, and consumer feedback to propose viable flavor systems and processing paths.
- Search-space reduction: Rank candidate recipes by predicted liking, cost-to-serve, stability, or nutrition targets before you hit the bench.
- Constraint balancing: Work within tighter limits for sodium, sugar, allergens, sustainability, and cost - and still keep an acceptable sensory profile.
- Packaging and process simulation: Model flow, squeeze, and spread behaviors (think: squeezable formats) to cut physical tests.
- Quality and off-flavor detection: Flag likely defect pathways from ingredient interactions, storage conditions, or process drift.
- Experiment design: Optimize DOE to learn more per test; stop unpromising lines earlier.
- Portfolio triage: Predict launch risk and prioritize concepts with a better shot at meeting sensory, cost, and operations constraints.
What AI won't do (and you shouldn't ask it to)
- Replace sensory panels. You still need trained panels and consumer tests for final validation.
- Perfectly predict taste from chemistry alone. Complex mixtures plus human variability make that a dead end with today's data.
- Substitute for human judgment. Scientists set goals, define constraints, and interpret ambiguous results.
Field notes from practitioners
David Sack (AKA Foods) positions his platform as an internal knowledge system, not a scientist replacement. The value: consolidating past formulations, panel data, and tacit know-how that usually lives in notebooks and heads.
Jason Cohen (Simulacra Data; founder of Analytical Flavor Systems, acquired by NielsenIQ in 2025) says their best returns come from identifying off-flavors early, narrowing formulation options, and prioritizing what to test - not from skipping human perception entirely. The pattern is consistent: AI trims the tree; people pick the fruit.
Practical playbook for R&D leaders
- Inventory your data. Collect formulas, batch records, process parameters, sensory notes, and consumer data into an accessible system (ELN/LIMS). Fix naming, units, and versioning.
- Codify your sensory language. Build or adopt a shared lexicon and scaling approach across teams and products so models learn consistent signals.
- Link data to outcomes. Tie formulations and process conditions to panel results and in-market performance. Without labels, models guess.
- Protect IP while collaborating. Use clean rooms, federated learning, or model-to-data setups with vendors. Spell out data ownership and model retraining rights.
- Pilot with paired sprints. Run digital screening and bench work in parallel for a few focused projects (e.g., sodium reduction, plant-based texture). Measure hit-rate improvement and time saved.
- Choose vendors for fit, not flash. Ask for: data model details, on-prem or VPC options, fine-tuning with your data, audit trails, and integration with your ELN/LIMS and PLM.
- Define success metrics early. Example: 30% fewer physical trials to reach a consumer-viable concept; 20% reduction in panel hours; improved first-pass yield at scale-up.
- Keep panels central. Use AI to size and focus panels, not remove them. Validate with humans at the points that matter.
- Upskill your team. Train scientists to frame good questions for models and to interpret outputs critically.
If you're building team capability, here's a curated starting point for R&D roles: AI courses by job.
Bottom line
AI helps food R&D teams move faster with fewer dead ends. The biggest wins are in search, screening, and knowledge reuse - not in replacing taste. As constraints on health, sustainability, and cost tighten, the teams that organize their data and keep humans in the loop will ship better products, sooner.
Consumers still make the final call - with their palate, not a model.
Your membership also unlocks:
