What AI actually does in food formulation: a practical guide for UK food brands

A protein bar brand uses an AI tool to model ingredient substitutions. It cuts reformulation trials from twelve to three. The time from brief to bench sample drops from six weeks to two. The finished product passes NPM scoring and hits the cost target.

That's a real outcome. It happened in 2025. But it happened at a large manufacturer with structured R&D data going back years, a dedicated data team, and four-figure monthly software spend.

If you're working at an SME food brand, that story matters. So does the gap between it and where most smaller teams actually are. AI in food formulation is real and accelerating, and a lot of it is inaccessible or overstated depending on where you're standing.

This post is about what's actually useful now, for a food team that doesn't have Unilever's data infrastructure.

What AI in formulation is actually doing in 2026

The honest starting point: AI in food formulation isn't one thing. It covers a range of tools with very different capabilities, and conflating them leads to both overhype and unnecessary scepticism.

The main application types:

Predictive reformulation modelling. Tools that take an existing formulation, apply a change (reduce sugar by 20%, swap soy for pea protein, remove a specific E-number), and predict how the finished product will perform across nutrition, cost, texture, and shelf life before any bench trial. Unilever uses this across multiple product lines. For Hellmann's squeeze bottle development, AI modelling saved months of physical testing. Nestlé and PepsiCo have been doing similar work since at least 2022. Mondelēz's AI product development tool has helped create 70 SKUs, produced two to five times faster than through traditional methods.

Ingredient substitution recommendations. Given a target - reduce HFSS score, hit a clean-label brief, meet a cost per kg target - these tools suggest ingredient swaps with predicted outcomes. This is where Nibblr currently sits. The platform surfaces substitution options mapped to your formulation targets, compliance requirements, and product category. Two early examples: a sauce brand used Nibblr to identify non-seed oil and vegetable purée alternatives that replaced a high proportion of seed oils within existing cost parameters and without sacrificing flavour. A breakfast brand sought to increase its official fruit, vegetable and nut content to improve its NPM score, and Nibblr identified additions and swaps that achieved this while maintaining product integrity. We wrote about both in more detail in the IFST's Food Science and Technology journal in December 2025.

Generative formulation from scratch. AI systems that take a spec (protein 20g, sugar under 5g, free-from gluten, target texture: chewy bar) and generate candidate formulations. NotCo pioneered this with its Giuseppe platform, applying it commercially for Kraft Heinz and others. More accessible versions are emerging, but quality depends heavily on training data. For most food SMEs, this remains a large-manufacturer capability for now.

Nutrition scoring and compliance modelling. Tools that calculate NPM score, front-of-pack nutrition labelling, allergen status, and HFSS classification as a formulation changes. The most accessible entry point for smaller teams, and the foundation on which everything else builds.

Where Nibblr is heading: ingredient substitution today, predictive modelling as your product data matures. The reason that progression matters is explained in the next section.

The real constraint: your data is the bottleneck

Every AI model in food formulation is only as good as the data it's trained on or the data you feed it.

This is where the SME reality diverges sharply from the Unilever story. A peer-reviewed paper published in Translational Food Sciences in December 2025 makes this explicit: SMEs often lack the infrastructure for comprehensive data collection and management, and this limits their ability to effectively use AI for formulation.

The specific problems most teams hit:

Data is siloed. Most food brands have formulation data spread across lab notebooks, Excel files, email chains, and the heads of individual developers. AI tools can't learn from data they can't access.

Data is inconsistent. Ingredient specifications recorded in different formats across different suppliers over five years will produce unreliable model outputs.

Historical trial data is often lost. Every failed bench trial contains useful information. Most small teams don't capture it in a form an AI system can use later.

This is exactly where getting structured pays off, not just now but in compound ways over time. A team that starts capturing formulations, trial outcomes, ingredient specs, and compliance results consistently today will be in a position to run meaningfully better predictive modelling in 12 to 18 months. A team that doesn't will still be starting from scratch when the tools are even more capable.

Nibblr's spec and product data management is built with this in mind. The aim isn't only to help you manage today's compliance workload. It's to structure your data in a way that makes the next generation of AI formulation tools actually useful when you reach them, rather than hitting a wall because your ingredients, trials, and outcomes are in three different Excel files owned by three different developers.

Where it genuinely helps right now, even for smaller teams

Regulatory and nutritional modelling. Tools that calculate NPM score, HFSS status, calorie content, and allergen flags as you adjust a formulation save hours per iteration. Catching a formulation that would push a product into HFSS scope before you've ordered raw materials is exactly the kind of expensive error these tools prevent.

Ingredient substitution recommendations. For a developer working on reducing emulsifier load in a baked product, finding a binding agent that keeps the product below an NPM threshold, or identifying a seed oil alternative that holds up at a given cost, an AI-assisted substitution tool narrows the candidate list faster than manual literature searching or supplier calls. The key point: substitution recommendations still need bench validation. They narrow the list; they don't replace the trial.

Costing and optimisation. Tools that model ingredient cost against a nutritional or sensory target help teams decide which reformulation routes are worth physical trials before committing. Particularly useful when ingredient prices are volatile, which has been a consistent reality for UK food brands since 2021.

Spec and data management. Not glamorous. Completely foundational. A searchable, consistent record of your own formulations, trials, and ingredient specs is the raw material that makes more sophisticated AI applications possible. Investing in this now is investing in what your R&D capability looks like in two years.

What AI still can't do

AI cannot taste, smell, or assess texture. A model might predict that swapping 30% of sucrose for a polyol blend will maintain sweetness, but it won't catch the cooling aftertaste that kills repeat purchase. That sensory gap is structural.

AI cannot validate against your specific manufacturing context. A formulation that works on paper may behave differently during pasteurisation, at your fill weight, with your specific equipment. AI models generalise. Your factory is specific.

AI cannot reliably handle regulatory edge cases. Compliance modelling is useful for well-understood rules (NPM scoring, calorie content, allergen flags). It's much less reliable on novel food status, health claim eligibility, or category-specific labelling edge cases. A human who understands the regulatory text remains essential.

AI cannot fill data gaps with good guesses. If your formulation history is thin, inconsistent, or inaccessible, an AI tool will extrapolate poorly. A confident-sounding wrong answer is arguably more dangerous than no answer at all.

What this means in practice

The teams moving faster on reformulation in 2026 aren't necessarily the ones with the most sophisticated AI. They're the ones who've sorted their data well enough that AI can actually work with it, and who've started with the use cases where the tools are genuinely ready.

Start with your data. Audit what you have, where it lives, and whether it's consistent. A searchable formulation history beats fragmented email folders, and it compounds: 18 months of structured data makes predictive modelling viable when you're ready to go there.

Start with compliance and substitution. Nutritional modelling, HFSS and NPM scoring in real time, and ingredient substitution recommendations are accessible now. The decisions they inform are frequent and the cost of getting them wrong is high.

And be honest about what the tool is predicting. No AI formulation tool has tasted your product. Its outputs are hypotheses for bench trials, not replacements for them.

Frequently asked questions

What is AI food formulation?

AI food formulation refers to the use of machine learning and predictive modelling to assist with recipe development, ingredient substitution, nutritional optimisation, and compliance modelling. Tools range from ingredient substitution recommenders that surface evidence-backed swap options to predictive models that simulate how a reformulation will affect nutrition, cost, texture, and shelf life before any physical trial.

Can AI replace food scientists and product developers?

No. AI tools assist formulation work by narrowing options, modelling outcomes, and surfacing information faster. They cannot taste, smell, or assess texture, and they cannot account for the specific behaviour of your ingredients and equipment in your manufacturing environment. Human expertise is essential for sensory validation, edge-case regulatory decisions, and any context where judgement matters.

Which AI formulation tools are most useful for UK food SMEs right now?

As of mid-2026, the most accessible applications are ingredient substitution recommendation tools (which surface swap options mapped to nutritional, compliance, and cost targets), compliance modelling tools that calculate NPM score and HFSS status as formulations change, and spec and data management platforms. Full-stack predictive AI platforms used by large manufacturers typically require substantial proprietary data to work well and are not yet practical for most SME teams without first investing in data structure.

What data does a food team need before AI formulation tools work well?

Structured, consistent, accessible formulation data: records of past formulations, bench trial outcomes (including failures), ingredient specifications, processing parameters, and cost history, in a format that can be searched and exported. The quality of AI formulation output is directly limited by the quality and accessibility of the data fed into it. Teams that invest in organising this now are building the foundation for predictive modelling capability later.

How does AI help with HFSS compliance and reformulation?

AI-assisted compliance modelling tools can calculate a product's NPM score in real time as ingredients and quantities change, flag when a reformulation crosses the HFSS threshold, and model the nutritional impact of substitutions. Teams can identify compliance issues before committing to raw material orders or production trials, reducing the cost of late-stage reformulation.

What's the biggest risk of using AI in food formulation?

Treating confident-sounding model outputs as reliable predictions when the underlying data is thin, inconsistent, or out of date. AI models trained on general food science literature may extrapolate poorly to your specific ingredients, processing conditions, and product format. The risk is not that AI gives no answer. It's that it gives a plausible-sounding wrong answer that still requires a sensory trial to catch.