What supervision will do for you
In short: we will monitor each agent, you will receive a report every month and we will take charge of reported problems. You will remain the decision-maker.
- monitoring of each agent, by us: its status, what has been approved, its stops;
- alerts, for us and, depending on the case, for you;
- a clear monthly report, and every three months a review of levels (what each agent is allowed to do alone) and versions;
- support, with tickets (requests followed until they are resolved) that the agents will open themselves;
- updates, tested on your own cases before they are activated.
A few figures
- 1report every month
- 3 monthsbetween two reviews of levels and versions
- 4autonomy levels, from 0 to 3
- 5items in the subscription
Read moreHide details
Your agents will work continuously, and you will remain the decision-maker. Supervision will check that this rule holds, for each agent, over time.
The monthly subscription will include:
Price: €119 excl. VAT per month, plus €219 to €849 excl. VAT per month per AI agent depending on its tier. See our pricing.
A diagram to see what each agent does
In short: each agent will have a simple diagram of its work. You will see where it waits for your agreement and when it stops.
- Every working day, the agent reads the invoices that have fallen due in your invoicing software.
- It checks in your bank account whether payment has arrived.
- No payment: it prepares a reminder, in the tone you have chosen.
- Approval point: a person in your team reads, corrects or refuses.
- The reminder goes out. The agent's log will record who approved, when, and with which version.
Read moreHide details
Before going live, each agent will receive a diagram of its journey. It shows its steps, the information it reads, the moments when it waits for your agreement and the cases where it stops. No technical knowledge is needed to read it.
This diagram will also be used for supervision: it is what we will follow the agent against. If the journey changes, the diagram will change with it. Every version will be kept.
Example: an invoice reminders agent
Planned stops: customer in dispute, contested invoice, missing information. In these cases, the agent does not guess. It stops and hands over to you. A formal notice never goes out without human approval.
Every month, a clear report on your agents
In short: one report a month. You will know what your agents did, what got stuck and what we suggest.
- The work done: what each agent handled, and how many suggestions your team approved, corrected or refused.
- The stops: how many times the agent handed over, and why.
- The errors found, and what we fixed.
- Your indicators, chosen with you at the start. For example, the first-response time for a support agent, or the average payment time for a payment reminders agent.
- The tickets of the month, opened and resolved.
- The versions in service, and what has changed.
- Our advice on the autonomy level of each agent.
Read moreHide details
Every month, you will receive a report. It will answer three questions: what did your agents do, what got stuck, and what do we suggest?
Levels: what the agent does alone
A level says what the agent is allowed to do alone. It is set agent by agent, and action by action.
- Level 0, in the shadow: the agent suggests but does not act. Your team does the work, and we compare.
- Level 1: the agent prepares, your team approves every action.
- Level 2: the agent acts alone on safe cases, described in writing with you. The rest comes back to you, and a sample is checked.
- Level 3: the agent acts alone. You read the report and receive the alerts.
At which level, what does the agent do?
ExampleLevel 1, systematic approval: the agent prepares, your team approves every action.
- Chase a quote left unansweredIt prepares the reminder; a person approves, corrects or refuses it before it is sent.Suggests, you approve
- Send a quote to a customerA person approves the quote before it goes out.Suggests, you approve
- Pay an invoiceA person always approves a payment: the agent never pays.Suggests, you approve
- Delete a fileDeleting data always goes through your approval.Suggests, you approve
- A serious complaint arrivesIt stops and leaves the decision to you: this is normal operation, not a fault.Stops, you decide
A payment, sending a document that commits you and deleting data stay at level 1, whichever level is chosen. Example to illustrate the rules described on this page.
Example, for each level, of what the agent does:
Level 0, in the shadow. The agent suggests, but does not act. Your team does the work, and we compare.
- Chase a quote left unanswered: Suggests only. It writes the reminder; nobody sends it, and it is compared with your team's.
- Send a quote to a customer: Suggests only. It prepares the quote; nobody sends it.
- Pay an invoice: Suggests only. It may suggest; the agent never pays.
- Delete a file: Suggests only. It may suggest; deleting data stays in your hands.
- A serious complaint arrives: Stops, you decide. It stops and leaves the decision to you: this is normal operation, not a fault.
Level 1, systematic approval. The agent prepares, your team approves every action.
- Chase a quote left unanswered: Suggests, you approve. It prepares the reminder; a person approves, corrects or refuses it before it is sent.
- Send a quote to a customer: Suggests, you approve. A person approves the quote before it goes out.
- Pay an invoice: Suggests, you approve. A person always approves a payment: the agent never pays.
- Delete a file: Suggests, you approve. Deleting data always goes through your approval.
- A serious complaint arrives: Stops, you decide. It stops and leaves the decision to you: this is normal operation, not a fault.
Level 2, by exception. The agent acts alone on safe cases, described in writing with you. The rest comes back to you, and a sample is checked.
- Chase a quote left unanswered: Acts alone. This is a safe case written in its rulebook: it sends it alone, stating that the message comes from an AI system. You check a sample.
- Send a quote to a customer: Suggests, you approve. This action stays at level 1, whichever level is set: a person always approves.
- Pay an invoice: Suggests, you approve. This action stays at level 1, whichever level is set: a person always approves.
- Delete a file: Suggests, you approve. This action stays at level 1, whichever level is set: a person always approves.
- A serious complaint arrives: Stops, you decide. It stops and leaves the decision to you: this is normal operation, not a fault.
Level 3, monitored autonomy. The agent acts alone. You read the report and receive the alerts.
- Chase a quote left unanswered: Acts alone. It acts alone; you read the report and receive the alerts.
- Send a quote to a customer: Suggests, you approve. This action stays at level 1, whichever level is set: a person always approves.
- Pay an invoice: Suggests, you approve. This action stays at level 1, whichever level is set: a person always approves.
- Delete a file: Suggests, you approve. This action stays at level 1, whichever level is set: a person always approves.
- A serious complaint arrives: Stops, you decide. It stops and leaves the decision to you: this is normal operation, not a fault.
An agent only goes up a level if the measurements justify it, and with your written agreement. Some actions always stay at level 1: a payment, sending a third party a document that commits you, deleting data. Every three months, we will review levels and versions, and suggest the changes that are useful.
This principle matches the EU AI Act (Regulation 2024/1689). For so-called “high-risk” systems, it requires that a person can oversee the system, disregard or override its output, and stop it (Article 14). Whatever the classification of your agents, we will apply this principle to all of them.
A problem? Who is told, and when
In short: an agent that has doubts stops and lets you decide. In the event of a serious alert, we will be notified and the agent will go back down one level: it will do fewer things alone.
- The agent has doubts: it stops and leaves the decision to you. This is normal operation, not a fault.
- A serious alert: the agent will automatically go back down one level. We will be alerted, and you will follow the ticket in your future customer area.
- At level 3: you will also receive the alerts directly, since nobody approves every action any more.
- Personal data is affected: when we process personal data on your behalf, the GDPR makes us your “processor” (Article 4). In the event of a breach of this data, we must tell you without undue delay (Article 33). You must then notify the data protection authority, without undue delay and, where feasible, within 72 hours of becoming aware of it. The only exception: a breach that is unlikely to result in a risk to individuals.
Read moreHide details
Where does the information go? Your agents run where you have chosen: your cloud in a European region, your servers, or a dedicated server with us. The AI model itself may be called at a provider: see data and security. Our supervision tool will only receive indicators and alerts. Our support tool will receive the technical detail of tickets.
Your agents will report their own problems
In short: when an agent runs into a problem, it will open a support request itself. We will be alerted, and you will follow everything in one place.
- a ticket will be opened, linked to your company and the project concerned, with the useful technical detail;
- we will be alerted;
- you will see the ticket in your client area, with its status;
- duplicates will be grouped together: if the same problem comes back fifty times, there will be one ticket, not fifty messages.
Read moreHide details
A ticket is a tracked request, from the first report to the fix. It may come from you, through your future client area, or from your agents themselves.
When an agent runs into a bug (a defect in the program), a repeated error or a change of AI model, it will report it straight away, without waiting for anyone:
It will be the same support tool for all our services: agents, custom software and training. One place to follow everything.
Updates tested on your cases before they are switched on
In short: no new version will be switched on without testing. We will check it first on your own cases.
Key fact
- 60 daysof notice at least before a published model is retired, at Anthropic
- a frozen version: each agent will run on a specific version of its model, its tools and its connectors (the links with your software). No new version will be switched on without us.
- a test on your cases: before an update, we will replay the test cases prepared during set-up with the new version, using anonymised data.
- a comparison: we will compare the results with those of the version in service.
- switch-on only after this test: the new version will go live only once the test is passed. The monthly report will tell you what changed.
Read moreHide details
The AI model is the engine that reads and writes for the agent. Anthropic, which makes Claude, regularly retires its older models. It gives its customers at least 60 days' notice before a published model is retired. After that, requests sent to that model fail. Amazon Bedrock and Google Cloud, which also offer Claude, set their own schedules.
What we will do:
The same rule will apply when one of your software tools changes the way it exchanges data.
How quickly we take a problem in hand
The response time is the time between a ticket being opened and the moment we start working on it. It is not the time to resolution. It will be stated in your quote.
Whatever this time, an agent that has doubts stops. Your safety does not depend on the support response time.
Frequently asked questions
Do we need someone on our side to watch the agents?
We handle the technical monitoring. On your side, you need the people who approve the proposals. Their training is defined with you in your quote.
What happens if one of our software tools goes down?
The agent will no longer be able to read or write. It will stop, as it does with any missing data, and a repeated error will be reported to us.
Next step: the audit
It all starts with a short audit. We look at your tools, the way you work, your volumes and the personal data involved. You receive a one-page report: the AI agents that are useful for you, what they can save you, and the quote.
The audit costs €490 excl. VAT, and its price is deducted from the agents quote if you sign it within 30 days.
Sources
Accessed on 13 October 2026.
- Anthropic, “Model deprecations”: “at least 60 days' notice” before a published model is retired, “Requests to retired models will fail”, Amazon Bedrock and Google Cloud “set their own retirement schedules”
- GDPR (Regulation (EU) 2016/679), Chapter IV, Article 33, paragraphs 1 and 2 (text reproduced by the CNIL): the processor notifies the controller “without undue delay”; notification to the authority “where feasible, not later than 72 hours”
- GDPR (Regulation (EU) 2016/679), Chapter I, Article 4, point 8 (text reproduced by the CNIL): definition of the processor
- European Commission, AI Act Service Desk, Regulation (EU) 2024/1689, Article 14 “Human oversight”, French consolidated text as of 27/07/2026, with no mention of amendment by the omnibus: paragraph 1, paragraph 4 points (d) and (e)