TL;DR: AI can now complete months’ worth of food formulation work in just a few days. This doesn’t make product decisions any less important; it means poor judgments can now be repeated at a lower cost. Before you accept an AI-generated answer, consider which constraint is real, what the model optimized, who owns the tradeoff, and whether the result still sounds like your company. Ultimately, this raises the question: What Is the Role of the Product Manager When AI Can Come Up With the Product?
On Monday morning, a funding announcement landed in my editorial brief. A food technology company had raised $6 million to help large food brands use AI to formulate products before conducting physical tests.
The money was interesting, but the claim beneath it was even more interesting.
Proxy Foods AI, a platform focused on food research and development, states that its customers have reduced the number of development iterations by up to 80 percent and shortened prototype timelines from months to days. These figures come from the company itself, so they should not be treated as an independent benchmark. But the general trend seems plausible. Food formulation presents a major search problem, and machines are good at searching.
Give the system a target cost, nutrition requirements, a clean-label constraint, an allergen restriction, and a shelf-life goal. It can evaluate combinations more quickly than a team working through samples one at a time on the bench.
At first, I thought this seemed useful, but my second thought made me uncomfortable.
If AI can propose the formula, predict how it will perform, and reduce the number of physical trials, what is left for the product manager to do?
It turns out to be a great deal. AI reduces the cost of producing answers but doesn’t make the question any clearer. If you’ve been using Claude Opus lately, you can attest that an answer can be long without being clear.
Faster experiments create a new problem
For much of the history of product development, new ideas carried a natural cost. A new formulation required ingredients, bench time, packaging, testing, and someone’s attention. That cost forced a team to narrow the options before development began.
AI reduces that cost, which is mostly positive. It also means a team can produce twenty possible answers before agreeing on what makes an answer good.
I have come at this problem from the software side before. In The Human Touch in AI-Driven Product Development, I argued that AI can accelerate the work without replacing the judgment around it. Food makes this distinction impossible to ignore. A spreadsheet can tell you that a blend meets its cost target; it cannot tell you whether saving twelve cents has made the jar less deserving of the name on the label.
And that is where product management moves when experimentation gets cheap. The work shifts away from producing options and toward deciding which constraints deserve to survive.
We usually refer to that as prioritization. I believe the word is too mild. Prioritization seems to imply merely putting items in a list. What’s going on here is more about taste.
Taste is your ability to say no to everything and yes to a few things.
The Product Judgment Test
Now, before I agree to an AI-generated product recommendation, I want four questions to be answered. I refer to this as the Product Judgment Test.
1. What constraint is real?
All models require a target. The error lies in thinking that every target should carry equal weight.
A food product may need to be more affordable, last longer, use fewer ingredients, come from a specific region, meet a nutrition goal, and taste better. Put all of those into the system without ranking them, and the output will hide the decision you refused to make.
Some constraints are laws. Others come from the customer. Some protect the economics, while others are simply habits carried over from the previous version of the product.
The product manager’s first job is to separate them.
This is the same inside-out discipline I use in the Product Onion Framework. Start with the problem and the promise at the center. Let those determine which outer-layer requirements matter. If you begin with a list of available ingredients or features, the machine will efficiently build from the outside in.
2. What did the model optimize?
An answer can be mathematically good and commercially wrong.
If the system reduced cost, ask what it spent to get there. Maybe the texture changed. Maybe the ingredient list grew. Maybe a regional supplier disappeared. Maybe the product became easier to manufacture and harder to explain.
Optimization always comes with a cost. Sometimes it lands in the budget; sometimes it lands in the product.
It’s easy to overlook that bill when the output looks like a clear recommendation. The model shows a winning formula, and the team sees certainty. But a recommendation is only the visible side of a tradeoff. Product judgment means turning it over and reading the cost on the back.
I mentioned a similar mistake in Your Moat Is the Part You Can’t Demo. The most obvious feature is usually the simplest one to copy. This also applies to formulation. A cheaper recipe shows up in the margin. The supplier relationship, the trust built with customers, and the standard you refused to lower are harder to show. They might be the elements worth protecting.
3. Who owns the tradeoff?
AI should never become the unnamed person who made the decision.
Someone has to say, out loud, “We accepted this change because the margin improvement matters more than the difference customers noticed in testing.” Or, “We rejected it because the product stopped delivering the experience we promised.”
That sentence needs a human name beside it.
That is why I have maintained that an AI agent needs a contract, rather than another well-tuned prompt. The contract should specify the decision the system may make, the evidence it must provide, and the point at which a person assumes responsibility.
The model can recommend. It cannot be accountable. If no one owns the tradeoff, the team defaults to automated avoidance rather than product development.
4. Does the answer still sound like us?
This is the hardest question because it does not fit neatly into a score.
At Trevean Spice, a blend is not a collection of ingredients that happens to pass a test. The product carries a point of view about freshness, provenance, the grower, and the experience when someone opens the jar. A recommendation can improve one number while sanding away that point of view.
I am not saying founders should reject data every time it contradicts the story they want to tell; that would be vanity disguised as something more respectable. Evidence should change our minds.
But a company also needs a line it will not optimize past. Otherwise, the product slowly becomes whatever the spreadsheet finds easiest to defend.
The final question is not whether the machine found a valid answer. It is whether we would still be proud to explain that answer to the customer without hiding the tradeoff.
Why the big companies may struggle first
Deloitte recently surveyed 200 retail and CPG executives and found a wide gap between AI enthusiasm and results. Seventy-five percent called AI a top strategic priority, but only 16.5 percent said they could quantify a return. CPG respondents reported more impact in product development than retailers did, yet most AI programs across the sector remain stuck between pilots and broad use. Deloitte’s 2026 retail and CPG survey describes the pattern as high conviction and low execution.
It would be easy for a small founder to read that and assume the large companies will win once they finish installing the technology. I am not sure.
Large companies have more data and more money. They also have more people who can avoid making a tradeoff. R&D can blame the model. Marketing can blame the brief. Finance can point to the margin target. The product can travel through the whole system without one person saying, “I chose this compromise.”
A small team has fewer places to hide.
A wrong decision hurts, but on a small team, it is visible enough to learn from. The founder can put the sample on the table, consider the cost, listen to the customer, and close the decision in one meeting. AI speeds up the search while the team keeps its judgment close to the product.
That is also why building Trevean Spice in public through the Product Onion has mattered to me. The framework forces the outer decisions to trace back to the promise at the center. AI can help explore the outside layers. It should not quietly rewrite the center.
What this changes for Trevean
The practical change is small. Before we use AI to recommend a blend, supplier, package, price, or product requirement, I want a one-page judgment record beside the output.
It will name the real constraint. It will show what the system optimized and what got worse. It will name the person accepting the tradeoff. And it will include one plain sentence explaining why the answer still belongs inside Trevean.
If we cannot fill in those four lines, the recommendation is not ready. It may be clever. It may save money. It may even be correct on paper. But it is still an option, not a product decision.
AI lets us test more formulations than a small team could have managed a few years ago. I want to use that advantage. But I also know what happens when more options arrive without stronger judgment. The backlog grows, the conversation gets louder, and nobody closes the door.
A future product manager will not earn a place by producing one more answer. The work is choosing which answer the company is willing to become.
What to steal
Use this after an AI tool recommends a formulation, feature, supplier, price, or product change:
Review the recommendation below as a product decision, not as an optimization result.
1. REAL CONSTRAINT
Identify which requirements come from law, customer evidence, unit economics,
brand promise, or internal habit. Flag any requirement that has not been ranked.
2. OPTIMIZATION BILL
State exactly what improved and what became worse. Include cost, quality,
customer experience, explainability, sourcing, and long-term defensibility.
3. HUMAN OWNER
Name the role that must accept the tradeoff. Do not assign accountability
to "the team," the model, or a department.
4. IDENTITY CHECK
Write one plain sentence explaining why this recommendation still fits the
company's promise. If that sentence requires hiding a tradeoff, mark it FAIL.
Finish with one verdict: TEST, ACCEPT, REJECT, or NEEDS A HUMAN DECISION.
RECOMMENDATION:
[paste the AI recommendation here]The one input to add before running it is your actual product promise. Without that context, the model can audit the numbers but cannot test whether the answer still sounds like you.
Related reading from PMJ
- The Human Touch in AI-Driven Product Development: Why faster product work still needs human judgment.
- Stop Tuning the Prompt. Your AI Agent Needs a Contract.: How to define evidence, limits, and accountability before delegating work to AI.
- The Product Onion Framework: the inside-out method for keeping features and requirements tied to the real customer problem.
- Building Trevean Spice From Scratch: The Product Onion applied to a physical CPG product in real time.
FAQ
Can AI actually formulate a food product without physical testing?
It can narrow down the search and predict how formulations are likely to perform. Physical trials, sensory testing, manufacturing validation, and regulatory review still have to be carried out. The useful change is a reduction in the number of blind iterations, not the elimination of bench testing.
Is the 80 percent reduction in iterations reliable?
Since Proxy Foods presents the figures relating to its own customers, it should be regarded as a statement by the company rather than as an objective benchmark for the industry. The argument also doesn’t rely on the precise percentage; even a small decrease in iteration costs alters the way that teams generate and assess options.
Is the Product Judgment Test anti-AI?
On the contrary, it is a method of using AI without mistaking a recommendation for a decision. The test increases the system’s usefulness since it compels the team to provide the context, evidence, and boundaries that the model cannot responsibly invent.
Who should own an AI-assisted product decision?
The person responsible for both the business and customer outcomes should own it. This could be a product manager, the founder, the R&D lead, or the category owner. It must never be an unnamed committee or the AI system itself.
What is the one step I should take this week?
Choose one recommendation your team generated with AI and write down what improved and what became worse. If you cannot name both, you have not yet reviewed the tradeoff.
If you want the frameworks I use to make decisions like this, grab the free Startup PM Toolkit, then reply and tell me which tradeoff your AI recommendation was hiding.
Dan
Leave a Reply