Taking AI to Production: A Checklist for French Companies

Taking AI to production means moving a model or agent that worked in a demo into daily use, on real data, with real users, under real security and legal constraints. The technology is rarely what blocks you. What blocks you is everything around it: separated environments, logging, human oversight, GDPR, cost control and the people who must adopt the tool. This guide walks through each of those, with the French and European references you will be asked about.

TL;DR

• A pilot proves that an idea works. Production proves that it keeps working: monitored, secured, documented and owned by someone in your company.
• In France, the CNIL treats training and production as distinct processing phases, so your GDPR analysis cannot stop at the model.
• Separate pre-production from production, log inputs and outputs, filter both directions, and cap inference usage before launch.
• The EU AI Act now applies, with staggered deadlines. Chatbots and similar systems should check their transparency duties first.
• Name internal business owners and train users. Adoption, not code, decides whether a pilot becomes a lasting tool.

Table of Contents

• What changes between a pilot and production
• Step 1: Decide what “ready” means before you build
• Step 2: Separate your environments
• Step 3: Secure the system for how it actually fails
• Step 4: Settle the legal questions in France and the EU
• Step 5: Design for humans in the loop
• Step 6: Monitor quality, drift and cost
• Step 7: Prepare the people
• A realistic timeline
• Our view: production is a product, not a project
• Where we can help
• Pre-launch checklist
• FAQ
• Sources

What changes between a pilot and production

A pilot answers one question: can this work on our data? It usually runs on a small scope, with a few friendly users, a developer watching the logs and manual fixes behind the scenes. Nobody minds if it fails on Friday evening.

Production answers another question: can the business depend on this? That changes the requirements in five ways.

1. Availability. Someone must be accountable when the tool is down or answers badly.
2. Security. The system is now a target, and it touches real customer or employee data.
3. Compliance. Data protection, transparency and, depending on the use case, risk-classification duties apply.
4. Cost. Usage is no longer a few hundred test calls. It is open-ended.
5. People. Users must trust the tool, know its limits and know when to override it.

The typical French SME project we see fails on none of the technical points. It stalls because the pilot never had an owner, a budget line for run costs, or a decision rule for “good enough to launch”. Fix those three and half the work is done.

Step 1: Decide what “ready” means before you build

Write the go-live criteria while the pilot is still running. Without them, the launch date is decided by impatience.

Useful criteria are measurable and tied to the business case:

• A target accuracy or resolution rate on a fixed test set drawn from real cases, not invented ones.
• A maximum share of answers that need human correction.
• A response-time ceiling.
• A cost per transaction that fits your margin.

France Num advises expressing return on investment in hours saved or errors avoided, and reminds executives that legal and financial responsibility remains human: AI is an assistant, not the sole decision-maker. Build that principle into your criteria. For every automated decision, define who reviews it, when, and how they can reverse it.

If you have not yet priced the return, our piece on turning hours and errors into dollars gives a simple way to do it. Our guide to running an AI pilot covers the stage before this one.

Step 2: Separate your environments

The CNIL’s guide on using an AI system in production stresses a distinct operational phase, with logging of inputs and outputs and a clear split between pre-production and production. This is not bureaucracy. It is how you avoid testing a new prompt on live customers.

A workable minimum setup:

Two rules save most teams from trouble. First, no change goes live without passing through pre-production. Second, credentials and API keys differ per environment, so a test script can never touch live data.

Step 3: Secure the system for how it actually fails

Conventional application security still applies. AI adds attack types that are specific to it.

France’s national cybersecurity agency, ANSSI, published recommendations for generative AI. A summary of the 35 measures identifies three main threat families: manipulation (hijacking the system’s behavior through malicious requests), infection (poisoning of data) and exfiltration of data. Among the measures it highlights are banning retraining in production, systematically filtering inputs and outputs, and prohibiting consumer tools for processing sensitive professional data.

In practice, that translates to a short list:

• Filter in both directions. Check what goes into the model (injection attempts, personal data that should not be there) and what comes out (leaked data, off-topic or harmful content).
• Do not retrain live. Learn from production logs offline, validate the result, and ship a new version through your normal release path.
• Assume your system prompt can leak. The OWASP guidance on LLM applications warns against treating system instructions as secrets. Keep credentials and business rules out of the prompt and enforce permissions in code.
• Limit what the agent can do. An assistant that can read a CRM does not need permission to delete records. Give each agent the narrowest rights that let it work.
• Ban shadow tools. If staff paste customer files into a consumer chatbot because your official tool is slow, you have a data leak, not a productivity gain.

If you want a structure to hold these decisions, our AI governance framework guide lays out roles, policies and review cycles for decision-makers.

Step 4: Settle the legal questions in France and the EU

GDPR and the CNIL

The CNIL’s position is that development and deployment are separate processing operations, each with its own purpose, data and legal regime. A lawful training phase does not make the operational phase lawful. For production you should be able to answer:

• What is the purpose of this deployment, and is it the one users were told about?
• What is the legal basis, and who is the controller and who the processor?
• Which personal data enters prompts, logs and outputs, and for how long is it kept?
• Is a data protection impact assessment needed?
• Where is the data hosted, and what do your model and cloud providers do with it?

If any answer is “we will see”, the system is not ready. We go deeper in our CNIL checklist for GDPR and AI in France.

The EU AI Act

According to the European Commission’s AI Act Service Desk, general application began on 2 August 2026. Transparency obligations apply on 2 December 2026 for systems placed on the market before August 2026, high-risk systems follow on 2 December 2027, and AI embedded in regulated products on 2 August 2028. Dates in this area have been moving, so check the Service Desk before you commit to a compliance calendar.

For a company deploying a customer-facing chatbot, the practical question is whether users know they are talking to a machine, and whether the use case falls into a high-risk category such as recruitment or credit. Classify the system before launch, not after. Record the classification and the reasoning in your project file.

Step 5: Design for humans in the loop

The CNIL’s production guide includes a section on letting a human keep control: automation can outperform an operator on some tasks, but a person must be available when needed. Concretely:

• Give users an easy escalation path to a person from every automated flow.
• Show confidence levels or sources where it helps the reviewer.
• Define which cases are never automated, for example complaints with legal consequences.
• Measure how often humans override the AI, and treat a rising rate as an alert.

For a WhatsApp or website assistant, escalation is the feature customers remember. A bot that hands over cleanly with the conversation history attached is better than a bot that tries to answer everything.

Step 6: Monitor quality, drift and cost

Going live is the start of operations. Four things need dashboards from day one.

Quality. Sample real conversations or documents every week and score them against your criteria. Keep a regression test set and run it before every release.

Drift. Your catalog, prices, policies and customers change. An assistant trained on last season’s FAQ will answer confidently and wrongly. Schedule content refreshes and tie them to business events.

Security signals. Alert on unusual prompt patterns, repeated refusals, spikes in output length and any detection of personal data in outputs.

Cost. OWASP lists unbounded consumption among the risks of LLM applications: uncontrolled use of inference can produce unexpected bills and even denial of service. Set quotas per user and per day, alerts at 50 and 80 percent of your monthly budget, and a hard stop you have agreed with finance.

Our budget guide for decision makers explains why run costs over three years deserve more attention than the build price.

Step 7: Prepare the people

France Num’s guidance on deploying generative AI in small businesses makes a point we agree with from experience: adoption depends more on people than on technology. It recommends naming internal business referents and including organizational change work, such as training and updating your risk assessment document (DUERP).

A short adoption plan that works:

1. Name one business owner per use case, not only an IT contact.
2. Train users on what the tool does, what it does not do, and how to report a bad answer.
3. Tell employee representatives early if the tool changes how work is done.
4. Publish a one-page usage policy: allowed data, forbidden data, escalation.
5. Collect feedback weekly for the first month, then monthly.

A realistic timeline

How long does it take? It depends on integrations and the quality of your data, not on the model. A scoped proof of concept can be delivered quickly: on our side we often deliver one within 14 days. Hardening it for production is separate work. Our deployment calendar for decision makers details a 2 to 4 month path from kickoff to live use.

The controlled launch is the step teams skip when they are in a hurry. Do not skip it. Releasing to 10 percent of traffic for two weeks will reveal what no test set contains.

Our view: production is a product, not a project

We build chatbots, agents and workflow automations for businesses, and the lesson we repeat to clients is simple: a project ends, a product does not. A custom AI tool in production needs an owner, a roadmap, a budget for run costs and a way to change safely.

That is why we structure engagements around a short proof of concept followed by a monthly subscription for the production tool, rather than a handover and goodbye. We also start with a free initial audit and ask for no long-term contract, so you can judge the result before committing further. Our deployments are built to be GDPR-compliant, and we have worked with large enterprises, including CAC 40 companies.

Three convictions shape how we work:

• Start narrow. One process, one measurable gain, one accountable owner. Expand after launch, not before.
• Integrate, do not bolt on. An assistant that cannot read your CRM, ERP or order system is a demo. The value is in the connection to your backend.
• Make failure cheap. Logging, quotas, human escalation and rollback let you recover quickly when something goes wrong, and something will.

— Botiqueai

Where we can help

If you have a pilot that needs to survive contact with real users, talk to us about a free audit. We develop custom AI automations connected to your systems, and workflow automations built on n8n and Make through our automation service. For websites and online stores, Aria is our customer assistant. If you run a Shopify store, our Shopify app plans start at 19 USD per month (Plan Starter), with Plan Pro at 49 USD per month.

Pre-launch checklist

Print this and tick it off with your team before go-live.

• [ ] Go-live criteria written, measured on a real test set
• [ ] Business owner and technical owner named
• [ ] Pre-production and production separated, with distinct credentials
• [ ] Inputs and outputs logged, with a defined retention period
• [ ] Input and output filtering in place
• [ ] No retraining on live data
• [ ] Agent permissions reduced to the minimum
• [ ] GDPR purposes, legal basis and impact assessment documented
• [ ] AI Act classification and transparency duties checked
• [ ] Human escalation path tested
• [ ] Quotas, cost alerts and a hard budget stop configured
• [ ] Users trained, usage policy published
• [ ] Rollback plan tested

FAQ

How long does it take to put an AI system into production?

Plan for a few months from kickoff in most business cases. A proof of concept can be ready in weeks, but security, integration, legal review and a controlled launch take additional time. Our deployment calendar shows a 2 to 4 month path.

Do I need a data protection impact assessment?

It depends on the data and the use. Because the CNIL treats training and operational use as separate processing, assess the production phase on its own. If you are unsure, ask your data protection officer or a lawyer specialized in data protection before launch.

Can we keep improving the model on live data?

ANSSI’s recommendations, as summarized here, advise against retraining in production because of poisoning risks. Collect feedback, review it offline, test a new version in pre-production, then release it deliberately.

What does an AI system cost once it is live?

Expect hosting, model usage, monitoring, maintenance and content updates, not only the build. Usage-based billing can climb quickly without quotas, which is why OWASP flags unbounded consumption. Our budget guide explains how to estimate total cost over three years.

Sources

• CNIL: Determining the applicable legal regime for AI
• CNIL: Using an AI system in production
• AI Act Service Desk: Frequently Asked Questions
• Solutions Numériques: ANSSI’s 35 recommendations for securing generative AI
• France Num: How to deploy generative AI in your small business
• OWASP GenAI Security Project: LLM applications guidance

Recommended

• Before First Training Run: CNIL Checklist for GDPR and AI in France
• AI Transformation Governance: A Guide for Decision-Makers
• AI Integration with Snowflake: A Practical Guide for French Companies