Updated 12 min read

Data and AI governance for food manufacturers: a 7-step roadmap for SMEs

By Émile Gadenne, CEO & co-founder of Agrolytics

Data and AI governance for food manufacturers: a 7-step roadmap for SMEs

Data and AI governance, for a food manufacturer, is the set of rules that says which data AI can see, what it is allowed to do and who decides. It comes before choosing a tool, in seven steps, with no IT department and no eighteen-month project: a handful of clear decisions is enough to get started.

This guide is for managing directors who sense that AI has already entered the business without a framework, and who are not sure where to start. It sets out three principles to settle, walks through the seven steps in order, and offers a table showing what to do depending on where you start from. If you are new to the subject, start with our guide to AI in the food industry, which covers use cases and costs.

Key takeaway

governance is decided before the tool, because the tool will change and your data will stay. Three principles are enough to set the frame: read-only by default, data perimeters chosen block by block, agent autonomy limited to your instructions. Since February 2025, the EU AI Act already requires companies that use AI to train their teams. An SME of 20 to 200 people can lay the foundations in one quarter, with a named AI lead and a two-page AI usage policy.

Data and AI governance: what does it mean for food manufacturers?

The word governance is off-putting because it evokes the committees and procedures of large groups. In a food and drink SME, it actually covers three simple questions, which are not dealt with in the same place.

Data governanceAI governanceRegulatory compliance
The questionWhat data do we have, where is it, who is responsible for it?What can AI see and do, and who decides?What does the law require of us?
Food industry examplePrices negotiated with each retailer, product specifications, lotsAn agent that reads sales but never touches pricesGDPR for customer files, EU AI Act
Who handles itOne owner per data blockThe managing director and an AI leadThe managing director, with legal advice
What it producesA data map and quality rulesAn AI policy and access perimetersEvidence: records, training, information

The three hold together. Poorly framed AI exposes data nobody had listed. Poorly kept data produces wrong analyses, even with the best tool. And compliance can only be proven if the first two exist.

The food industry has a particularity that makes the subject more sensitive than elsewhere. Your most valuable data is not necessarily personal: it is your pricing terms with major retailers, your recipes, your yields, your supplier purchase prices. GDPR does not protect them. Only your governance does.

Why governance comes before the tool in an AI strategy for SMEs

AI is already in the business, often through the back door. A sales rep pastes sales figures into an online assistant to draft a follow-up. A sales administration assistant has a margin table summarised. These uses are helpful, but nobody knows where the data goes or what becomes of it.

Faced with this, the reflex is to think in terms of tools: which software to choose, which agent to deploy. That is the wrong order. AI models are replaced every quarter, and nobody knows what an agent will look like in three years. Your data, on the other hand, will remain your data.

A sound AI strategy for an SME is therefore built on a framework, not on a product. A company that has set its rules for data flow and control will be able to adopt any future tool without rebuilding everything. One that rushed in without a framework will back-track at the first incident. And once the framework is in place, governance speeds up adoption instead of slowing it down: you only truly delegate what you control.

The 3 principles of a company AI policy to settle from the start

Before the steps, three decisions of principle. They depend on no particular technology and will remain valid whatever progress comes next.

First principle: read-only by default

The first fear of a managing director, when told about connecting AI to their information system, is almost always the same: will it touch my ERP? It is a legitimate fear. The ERP holds the accounts, stock, orders and production.

The answer lies in a verifiable architecture rule: by default, information flows from the ERP to the analysis tool, never the other way. AI reads a copy of the data in its own environment. It writes nothing, changes nothing and deletes nothing in the source system.

When an agent genuinely has to record something, the write goes through human approval. At one of our clients, a charcuterie and ready-meals maker, the traceability agent prepares the record of each cutting lot, but nothing enters the ERP without a supervisor's approval. The exception is explicit, limited and controlled.

Second principle: data perimeters you choose

Connecting AI must never mean giving it vague access to everything. Good governance requires explicit data blocks: customers, orders and sales, products and prices, purchasing. Each block is opened or closed separately.

A company that wants its raw material purchasing analysed, but keeps its customer files outside the perimeter, must be able to do so. This clarity is exactly what GDPR expects of responsible processing: a defined purpose and no collection beyond what is necessary. Compliance then follows from the way the tool is built.

Third principle: autonomy that answers to your instructions

An AI agent can monitor dozens of indicators continuously and produce analyses every day without being asked again. But autonomy does not mean freedom. The right question is not how far the agent can go, but who decides how far it goes.

An agent that takes initiativesAn agent that follows instructions
Decides alone what deserves its attentionActs within the frame you have set
Extends its perimeter without consulting youStays strictly within its perimeter
Forces you to take back controlLeaves you in control by default

Example: you ask Sophie, our financial analysis agent, once to alert you whenever a purchase price moves away from the market beyond a threshold. She will do so every day for months. But she will never renegotiate with a supplier and never place an order. That is what separates a framed agent from a simple generative tool, a difference covered in our article on the difference between generative AI and AI agents.

How to implement AI governance in 7 steps

The principles say what to decide. The steps say in which order to go about it. They are designed for an SME that is just starting: each fits into a few weeks and produces a concrete deliverable.

1. Assess your data maturity

Before building, you need to know where you start from. Where is your sales, purchasing and production data? In the ERP, in Excel files, in mailboxes? Who already uses an AI tool, and for what? How long does it take to produce last month's margin by customer?

The answers place the company on a simple scale: beginner, intermediate or advanced. That level sets the ambition of the following steps. A beginner company is not looking for a governance committee, it is looking to know who holds what.

Deliverable: one page describing the current situation, without judgement, and the list of AI tools already used by the teams.

2. Map your data

The data map answers three questions for each data block: where the data is stored, who enters it, who uses it. In a food and drink SME, the critical blocks are almost always the same: orders and sales, prices and terms per retailer, items and product specifications, purchasing and suppliers, lots and traceability.

There is no need to list everything at once. Firms that advise SMEs see it every time: the most effective starting point is the data that already causes problems, such as figures that differ from one spreadsheet to another or files that anyone can open. Start with five blocks, not fifty.

Deliverable: a one-page table, one data block per row.

3. Name who decides

Data without an owner is data nobody corrects. For each block on the map, name the person responsible: the sales administration lead for orders, the purchasing manager for supplier prices, the financial controller for margins. In an SME these roles are often combined, and that is not a problem as long as everyone knows what is theirs.

Also name an AI lead. It is not a job, it is a remit of a few hours a month: keeping the list of tools in use, taking questions from the teams, organising training. The managing director remains the one who decides.

Deliverable: a name next to each block, and a designated AI lead.

4. Make your ERP data reliable

AI connected to wrong data produces wrong analyses, with great confidence. Before connecting anything, check the points that most often distort calculations in the food industry: customers duplicated under two codes, items attached to the wrong family, units that mix kilos and cases, year-end rebates kept outside the ERP.

You do not need a perfect database. You need to know where the defects are, fix those affecting the blocks you will open to AI, and set a simple rule so that new entries stay clean. It is often the step that pays off most, AI or not. The six flaws to check first and the method to fix them are covered in our guide to ERP data quality in the food industry.

Deliverable: the list of known defects and one data entry rule per opened block.

5. Write your usage rules: the AI usage policy

The AI usage policy tells the teams what they can do with AI. Which tools are allowed. Which data must never be pasted into an online assistant: retailer prices, recipes, customer files. Who reviews what AI produces before it leaves the company.

Two pages are enough, and our 8-clause AI usage policy template for food and drink SMEs can be downloaded as a Word file. The canton of Geneva provides a free AI policy template for businesses, which adapts well. Watch for local employment law: in France, for example, if the policy is appended to the company rules, it must go through the works council, as set out in article L1321-4 of the French Labour Code.

Deliverable: a policy signed by management and presented to the teams.

6. Check the legal framework: GDPR and the EU AI Act

For an SME that uses AI tools without developing them, two texts matter. GDPR, as soon as personal data is involved: customer contacts, employee data. The French data protection authority, the CNIL, offers a section dedicated to small businesses, with practical sheets on generative AI.

Since 2 February 2025, the EU AI Act has required companies that use AI to take measures so that their teams know how to use it (article 4). Relaxed in 2026, this obligation has not been removed. Most uses in a food and drink SME, such as sales analysis or preparing follow-ups, do not fall under so-called high-risk systems.

Deliverable: the list of personal data processing concerned, and a record of the training delivered.

7. Choose a tool that respects your framework

Only now does the question of the tool arise, and it becomes much simpler. You know which data to open, to whom, and under which rules. What remains is to check that the tool allows it: read-only access to the ERP, block-by-block opening, where the data is hosted, the ability to retrieve everything if you leave.

These questions quickly rule out part of the market. They complement the selection criteria we cover in our article on how to choose an AI vendor and in our comparison of AI solutions for the food industry.

Deliverable: a list of questions given to every supplier consulted.

Which governance for your maturity level

Not all companies start from the same place. The table below shows where to focus effort over the next three months.

BeginnerIntermediateAdvanced
Typical situationFigures come out of Excel, AI is used without a frameworkThe ERP is the reference, a few dashboards existData is reliable, a first AI tool is running
PrioritySteps 1 to 3: know what you have and who owns itSteps 4 and 5: make data reliable and write the policySteps 6 and 7: prove compliance, extend use
First deliverableThe list of AI tools in useThe AI usage policyThe supplier selection grid
Trap to avoidTrying to map everythingOpening uncleaned data to AILetting the agent widen its own perimeter

Moving from one level to the next is not measured by the number of tools installed. It is measured by the ability to answer, in one minute, the question: what data does our AI see today, and who decided?

What governance doesn't do

It does not clean your data for you. An entry rule does not erase five years of duplicates. Governance says where the defects are and who fixes them; the fixing still has to be done.

It is not an eighteen-month project. Frameworks designed for large groups, such as the ISO 42001 standard, are useful when a customer demands it. To get started, an SME needs decisions, not a certification.

It does not replace legal advice. This guide gives the landmarks. For sensitive personal data processing or a policy appended to the company rules, have it reviewed by a lawyer or your accountant.

Data governance: where to start in practice

This week, list the AI tools already in use. Ask each department, without looking for someone to blame. You will almost always discover uses you did not suspect.

This month, name an AI lead and an owner for each data block. Five blocks, five names, one page. It is the base on which everything else rests.

This quarter, write the policy and choose the blocks you will open to AI. With these two decisions, you can assess any tool without the risk of having its rules imposed on you.

Frequently asked questions about AI governance for food manufacturers

Do you need an IT director or a DPO to implement AI governance?

No. A food and drink SME can set up its governance without an IT department: it needs a managing director who decides, an AI lead who follows the subject a few hours a month, and one owner per data block. A data protection officer is only mandatory in specific cases set out by GDPR, such as large-scale processing of sensitive data, which rarely applies to an SME in the sector.

How long does it take to lay the foundations?

One quarter is enough for a company that is starting out: a few weeks for the inventory and the data map, a few weeks to name owners and make the priority blocks reliable, then the policy. Most of the time goes into discussions with the teams, not into writing. Governance then grows as new uses appear.

Is an AI usage policy mandatory?

No text requires an AI policy as such. But it is the simplest way to show that you frame usage, that you train your teams as the EU AI Act requires, and that your sensitive data does not end up in consumer tools. In France, if you append it to the company rules, the works council must give its opinion.

Does the EU AI Act apply to an SME that uses a tool without developing it?

Yes, in part. An SME that uses an AI tool is a deployer within the meaning of the regulation. It must take measures so that its teams know how to use it, and comply with transparency obligations where they apply. The heavy obligations mainly target providers and high-risk systems, which the sales or financial uses of a food SME rarely reach.

Does connecting AI to my ERP put my data at risk?

The risk depends on the architecture, not on AI itself. Read-only access cannot change anything in the ERP: the tool works on a copy. Check three points before signing: the type of access, the ability to open data block by block, and where it is hosted. At Agrolytics, data is hosted in France, encrypted, and never shared.

Which data should you start with?

Start with the data that already causes problems and weighs on margin: orders and sales, purchase prices, prices per customer. It is also where AI delivers fastest, for example to spot a customer who is drifting away or a purchase price that is creeping up. Keep sensitive files, such as recipes, out of the first perimeter.

Further reading

Governance is not an end in itself: it is what lets you hand AI increasingly useful tasks safely. To see what an agent actually does once the framework is in place, read our guide to what AI agents do in the food industry, then find out why a useful agent combines several data sources and what an integration with Euragro changes in practice. To lay the foundations on the management side, see also how to move to data-driven management and why ERP vendors lag behind on AI.

Want to know which data to open first and under which rules? In 30 minutes, we look together at your data blocks and the perimeter an agent could cover without ever touching your ERP. Book a call

Ready to take control of your data?

Book a 30-minute demo and see what Agrolytics can do for you.

Book a demo