Nine out of 10 R&D execs in large CPG companies are using AI at work in some capacity, but just 19% have embedded it into routine workflows. And fewer still say it’s actually moving the needle, according to a new survey by Turing Labs, a startup developing AI/ML tools for CPG innovation teams.
On the face of it, the fact that 91% of 290 senior R&D leaders in the US and Europe* quizzed by Turing Labs are either using AI routinely or actively piloting/scaling it, hardly suggests they are behind the game, acknowledges Turing Labs cofounder and CEO Manmit Shrimali. But dig a little deeper, he says, and most of the activity is on concept generation and basic research and summarization tasks, where generic LLMs excel.
“The big impact is actually on the CPG marketing side, so marketers are now able to substantially reduce their spend on ad agencies, and they’re able to significantly increase in-house ad management. That’s where the wins are happening right now, because that’s where the data is very easy. But that’s not product innovation.”
The gap between AI activity and AI impact
Once you get into more substantive attempts to improve the pace and quality of innovation, however, the problems begin, with 20% of AI initiatives never implemented, 14% implemented and later abandoned, and 24% still in use but delivering minimal or no business impact, according to Turing Labs’ survey. Meanwhile, 57% of respondents said GenAI formulation outputs were too generic to use without significant manual re-work and validation.
The issue, says Shrimali, is two-fold: First, there’s a mismatch between general-purpose tools and highly domain-specific workflows. Formulation work is not just a search or content-generation problem, he says, but involves ingredient chemistry, regulatory constraints, manufacturability, cost, taste, texture, shelf life, nutritional targets, consumer positioning, and the commercial context around the product.
As a result, tools that look impressive in demos often fall short when tested against real problems: “Customers often try dumping thousands of technical white papers in ChatGPT and asking questions and then realizing that they get different responses to the same questions, so these kind of LLM approaches can be a disaster.”
Second, firms that see AI as a means of simply accelerating current processes rather than rethinking how innovation is organized in the first place will not get the full benefits, claims director of engineering Bryan McCarty.
“If I had to pick one thing the CPG industry is doing wrong in adapting AI… it’s trying to use it simply to speed up existing workflows.”
Beyond formulation optimization
Founded in 2019, Turing Labs started out as a formulation optimization platform but has since broadened into what Shrimali describes as an “innovation system” for CPG companies, who he claims can typically be divided into three buckets when it comes to AI.
“The first realize they can build in-house but then discover it takes too long. The second is constantly running pilots and trying everything without the resources to give enough justice to any single platform. The third asks the right questions, picks one vendor and goes very deep. They know that just because they can build something in-house doesn’t mean they want to, plus talent [in AI] is very expensive. And that’s why companies eventually come to us.”
According to McCarty, several large CPG companies have also learned the hard way that trying to build their own system doesn’t always work. Some, he claims, have “poured millions into data lakes and the promise of one massive intelligent system.” But in most cases, he says, this didn’t materialize: “What they got were isolated wins, frustrated product developers, and less confidence that AI could change how products are actually created.”
‘We’re not offering a GPT-wrapper’
Turing Labs, which has raised $19.25 million from backers such as Insight Partners and Moment Ventures and works with large CPG companies and retailer private-label teams, is trying something more ambitious, says Shrimali.
Rather than positioning itself as a generic concept generator, Turing Labs’ platform is built around domain-specific models and workflows spanning the innovation cycle, from commercial KPIs and market positioning through formulation, regulatory constraints, ingredient assessment and commercialization, he says.
“We’re not offering a GPT-wrapper or generic AI. The industry doesn’t lack concepts or ideas. It needs a new operating model to win in the market: a reasoning system on which lever to pull, how to make the winning product, and how to pay for the innovation initiatives.”
It is also helping CPG clients understand risk, says McCarty. “I don’t think that the risk profile that CPG companies take on is correct a lot of time. They often don’t take on things that could be huge wins, that could be billion-dollar brands, but at the same time when they do take a risk, it’s not measured and calculated, it’s based on gut instinct. That’s one area where we can help with better decision making moving forward.”
Beyond chatbots
Typically, says Shrimali, a Turing Labs customer starts with a business problem, such as “win back share from private label in pasta sauce.” Turing Labs’ agents “work that problem end to end, identify where the gap actually is (product, price, pack, regulations, claims etc.), what to make, the optimal formulation, what it costs at scale, where the regulatory risks sit, and which cost savings across the portfolio can fund the launch.
“Formulation optimization [an area where the firm originally focused] is still in there but it’s now one agent out of several.”
The system combines public data, expert knowledge, proprietary data generated by Turing Labs, and customer data, which is kept segregated, and is not used to train models for other customers.
Users can interact with the system through a browser-based workflow as well as chat-based interfaces, says Shrimali. However, chatbots alone are not enough for high-stakes R&D decisions, as the quality of the output “depends heavily on how questions are asked.” The company has therefore tried to balance LLM-style interaction with more structured workflows that reflect how product developers actually work.
Where agents can help—and where they still fall short
As for AI agents, they have improved dramatically, but they are not ready for prime time in every instance. They still fail, misunderstand context, require human intervention and need the right orchestration, context, guardrails and human collaboration, rather than being treated as autonomous problem-solvers that can be handed a task and left completely alone, says McCarty.
One example where they can perform well, however, is on regulatory work, he says. Today, a product developer may send a formula to a regulatory team and wait days for feedback. A more proactive AI workflow could flag regulatory risks earlier, translate rules into formulation constraints and suggest alternatives.
In this model, the AI is not merely checking compliance, but can help teams think about “how a regulatory change might create offensive as well as defensive innovation opportunities.”
‘Structurally defensive’ CPG innovation
The fundamental problems facing R&D execs in the survey are not new, acknowledges Shrimali. Had this survey been conducted 5-10 years ago, respondents would likely express the same general frustrations: they have to do more with less, waste time on unproductive tasks, deal with incompatible legacy IT systems, and expend blood, sweat and tears on innovations that ultimately fail, as most new CPG products do.
However, what is notable in 2026 is how many large food companies are “asking R&D teams to create the future” while budgets, staffing and incentives are still massively tilted toward defending existing products, he claims.
“The deeper issue is that CPG innovation has become structurally defensive and risk averse. Just look at how pharma is beating [the food industry] on GLP-1s. While organizations rely on innovation to drive growth, the vast majority of R&D capacity is directed toward protecting the existing portfolio. Breakthrough innovation is not just difficult to execute; it’s barely resourced.”
Meanwhile, the gap between brands and private label is narrowing, and smaller, more agile companies are beating Big Food in the innovation stakes, he says. In that environment, simply using AI to shave time from reformulation or come up with a packaging refresh may not be enough.
For Shrimali, the more relevant metrics for AI in food R&D are not just faster time to market but whether R&D is helping to drive revenue, margin, shelf space and genuinely differentiated products.
Survey highlights
According to Turing Labs’ survey:
- R&D effort over the next 12 months is overwhelmingly focused on maintaining and optimizing existing products rather than creating new ones.
- A large share of R&D effort is invested in concepts that never reach the shelf. And when products do launch, over half require reformulation within 12 months.
- 52% of respondents report that changes are required within the first year of launch, including 49% who make adjustments between 6 and 12 months.
- 60% of respondents say the quality gap between branded and private label products has reduced over the past three years, and 3% say it’s disappeared entirely.
*At companies with 1,000+ employees that formulate and manufacture food and beverages.
The post AI is everywhere in CPG R&D. But is it moving the needle? Turing Labs surveys the biggest players appeared first on AgFunderNews.
