Everyone in the food industry, from the boardroom to the operations staff, is getting excited about the advent of AI and how it can impact their business. There is so much possibility in what an AI agent can handle to speed up delivery, streamline efficiencies, and flag issues before they become real problems. However, after being mesmerized by the razzle-dazzle of the demo, food companies are forgetting to slow down and ask the most critical questions of the vendors looking to serve them.

AI agents move more swiftly than a person ever could, which means their mistakes can be much more costly than we imagined. Especially for food safety teams, where an agent operating without guardrails in lot traceability or QC hold management can contribute to massive compliance failure.

I have spent most of my career in operations and finance, CFO and COO roles in manufacturing, and before that leading information security, and I have seen what happens when controls are not in place, well before AI exploded into the stratosphere.  It is very similar to the advent of ERP implementation in the industry.  Food manufacturers found that the generic systems they were putting into place could not understand the nuances of the operational needs they were trying to serve.  These systems had to be rebuilt with a specific purpose in food.  Now, it is a similar issue, but with AI, and the implications are much greater.

AI is being sold to the food industry the same way ERP systems were years ago. It is the latest shiny tech in the room and, like ERP sales meetings of the past, it is a pitch of capabilities assessments and a review of interactive demonstrations.  However, there is something missing…the question of real-world compatibility and, frankly, the elephant in the room: what happens if the system goes sideways?

That said, before engaging with an AI provider who says they can offer your company a bevy of undeniable operational benefits, you must ask the right questions.  These are three of the most important to note and inquire about…please, do not leave the room without asking them!

1) This is a question that should stop a demo in its tracks if a vendor fumbles the answer. “Will the agent guess or stop when it cannot come to a conclusion on its own?”  Agents should not be left to their own judgment. They must have guardrails around a defined set of operations specific to the task they are seeking to accomplish. If the agent finds it needs assistance, it should be programmed to flag a human for interaction, rather than the other way around.  Ask the vendor what the agent will do if it is not sure of its conclusion. If they cannot answer the question, it is indeed a cause for alarm and an indication that their service may not be an asset, but a liability.

For food safety professionals, this is imperative to FSMA 204.  The rule demands companies produce their Key Data Elements at Critical Tracking events within 24 hours[1]. If the agent is guessing at the decisions it makes in the data layer, that means your ability to meet the FSMA requirement depends on if the agent guessed correctly.

2) Many companies regret not asking this one. “Who is in control of the agent?”  This is a very important question because you need to understand the difference between an agent programmed to operate at the system level and one with autonomy to make decisions for itself while executing actions within the guidelines it has been given. Yes, they essentially become extended members of your team and can take on tasks they know they can handle, such as monitoring inbox communications and processing new invoices, but they should never bypass a human’s permission in what they are not allowed to handle. While agents make things quicker, faster, and easier, a human touch remains important for monitoring their interactions.

In the context of food safety, this matters most for QC hold management. Agents should have the capability to flag holds, provide the test data, and notify the right people. They should not be allowed to release any hold without human permission.

3) Finally, if a vendor cannot answer this question, walk. “If something goes wrong, will the team be able to find out what happened and why?” If there is an error, the agent must be able to produce a report to show what happened, where it happened, how it happened, and the potential implications of its mistake on operations.  It should also be able to provide you with troubleshooting recommendations so that changes can be made directly into the system and the program relaunched. If the agent cannot operate on a fail-diagnose-inform-and-fix cycle, then it is not a technology you can trust.

[1] U.S. Food & Drug Administration. Requirements for Additional Traceability Records for Certain Foods (Food Traceability Rule), 21 CFR Part 1, Subpart S. Effective January 20, 2023; compliance date July 20, 2028. https://www.fda.gov/food/food-safety-modernization-act-fsma/fsma-final-rule-requirements-additional-traceability-records-certain-foods.

This is not optional per FSMA 204. An auditor is going to ask for a complete record of every action taken by an agent in your traceability data.  Make it a priority practice to ensure you can provide one.

Most food professionals in operations have been managing this kind of oversight throughout their careers.  The only change is that they are using a different tool to enhance efficiency. Just like a generic ERP in the food industry will never function as efficiently as a purpose-built ERP, the same holds true here for AI.  Agents must be purpose-built to understand the industry they are working in and the components that make up the engine, including lot records, catch weight, FEFO inventory, and quality control.

We encourage professionals in the food industry to ask hard questions of any vendor:

1. Will the agent guess or stop?
2. Who is really in control of the agent?
3. Will my team be able to understand an agent’s failure and fix it?

Vendors with real expertise in the food industry can confidently answer these questions. Those who can’t will struggle to explain whether the technology they are offering you was really built for the job.

Photo by Mohamed Nohassi on Unsplash