F8 / IT department

Open AI to the business without losing control.

Teams already use AI tools that IT neither chose nor listed, and every agent connected to your systems adds permissions nobody reviews. AI becomes governable when every use has an owner, every agent limited rights and every action a trace.

Submit a brief

Three concrete uses

01Mapping AI uses and reducing shadow AI
AI Act level · Minimal risk
F8 · Mapping AI uses and reducing shadow AI
1. 01Detect usesLogs, licences, declarations
2. 02Classify by riskData, purpose, supplier
3. CheckSign-off by the AI committee
4. 03Decide the statusAllow, replace or ban
5. 04Publish the catalogueWith the approved alternative
The problem
Staff paste contracts or customer files into consumer tools for want of an approved alternative. IT finds out about these uses by chance, often after the fact.
What the AI does
AI helps list uses from proxy logs, licence purchases and team declarations, and proposes a risk classification in a catalogue. It blocks nothing and sanctions no one: IT and the business decide what is allowed, replaced or banned.
Data needed
Proxy or cloud access broker logs, licence purchases and a questionnaire to teams.
Who signs off
The AI committee, with IT, the DPO and the business, approves each catalogue entry and the alternative offered.
Minimal risk (EU · AI Act); but if the analysis targets people rather than tools, it becomes monitoring of employees’ activity, to be presented beforehand to the works council and staff (France · Labour Code).
02Limiting the permissions of AI agents
AI Act level · Minimal risk
F8 · Limiting the permissions of AI agents
1. 01Create an identityOne dedicated account per agent
2. 02Limit the rightsRead by default, write when justified
3. 03List the toolsApproved MCP servers and APIs
4. CheckApproval by an authorised person
5. 04Approve the irreversibleSend, payment, deletion
The problem
An agent connected to email, the CRM or the ERP acts with whatever rights it is given, often administrator rights to save time. A malicious instruction hidden in a document can then turn it against you.
What the AI does
Each agent gets its own identity, rights limited to its task, a list of authorised tools and human approval before any irreversible action. These guardrails are enforced by the infrastructure, not by the prompt: the agent performs nothing irreversible on its own.
Data needed
The identity directory, the rights matrix per application and the list of tools exposed to agents.
Who signs off
The agent’s business owner approves its rights; a human approves every send, payment or deletion.
The level depends on the use, not the technology: the same agent screening job applications or assessing individuals’ creditworthiness would be high-risk (EU · AI Act, Annex III).
03Logging every AI action so you can prove it
AI Act level · Minimal risk
F8 · Logging every AI action so you can prove it
1. 01Trace every callRequest, tool, response, version
2. 02Link to identityAgent, user, business owner
3. 03Retain and protectWritten retention, restricted access
4. CheckSign-off by the CISO
5. 04Investigate an incidentAssisted search, human conclusion
The problem
After an incident or a complaint, nobody can say which agent did what, with which model version and on which data. Logs sometimes exist, but at the supplier’s and for a period it alone decides.
What the AI does
Every request, every tool call and every human approval is recorded in a log you control, linked to the agent’s identity and the model version. AI then helps search and summarise these traces; it does not alter or delete them, nor conclude the investigation.
Data needed
Logs from the AI gateway, agents and applications, with a written retention policy.
Who signs off
The CISO sets retention and access; any incident analysis is led and signed off by a person.
Logging is not a risky use but a form of evidence: for a high-risk system, the deployer will have to keep its logs for at least six months (EU · AI Act, Art. 26), and these logs, rich in personal data, must be secured (EU · GDPR, Art. 32).

To define before launching AI

This is meaning debt: if these words lack a single definition and an owner, no agent will give a reliable answer, and no control will be possible.

“AI use”A tool, an AI feature built into software you already own, or an agent: decide what goes into the catalogue, or the inventory stays partial.
“Agent”An assistant that suggests or a program that acts: the boundary determines the rights, human approval and logging required.
“Sensitive data”Under the GDPR, trade secrets or your internal classification: a single scale must say what may go to which tool.
“Owner”The person accountable to management for the use, not the one who installed it: every catalogue entry has a named owner.

What applies to you

EU · AI ActSince 2 February 2025, providers and deployers must take measures to build their staff’s AI literacy (Article 4); Regulation (EU) 2026/1744 turned this into a best-efforts obligation without removing it.
EU · NIS2If your organisation is an essential or important entity, management approves the security measures (Article 20), which cover the supply chain and access control among others (Article 21); a significant incident requires an early warning within 24 hours and a notification within 72 hours (Article 23). In France, the transposing law had not been finally adopted as at 29 October 2026.
EU · GDPRAny AI processing of personal data goes into the record of processing activities (Article 30), must be secured (Article 32) and, where the risk is high, must be preceded by an impact assessment (Article 35).
EU · Data ActCloud contracts, including SaaS, must allow you to switch provider; from 12 January 2027, switching charges are banned (Article 29).

What to measure

Share of AI uses listed in the catalogue with a named owner
Share of agents with their own identity and rights reviewed in the last six months
Share of agent actions traceable end to end in your own logs

What Clairancy does

A · Framing and decision
Build the catalogue of AI uses, the classification rule and the decision path between IT, the business and the DPO.
C · Sovereign AI and data
Design a sovereign AI architecture: a single gateway, logs under your control, a choice between trusted cloud, open model and local AI.
D · Governance and compliance
Write the usage policy, the agent rights matrix and the evidence file expected under the AI Act, the GDPR and, where relevant, NIS2.

Concepts to understand

Read on the blog

AI governanceAI governance: five proofs of control

Frequently asked questions

Other functions

Should we ban consumer AI tools?

A ban without an alternative pushes use onto personal phones. It is better to offer an approved tool, explicitly prohibit certain data and measure actual use.

Is hosting in Europe enough?

Not always. A provider subject to US law can be compelled to hand over data it controls, even when stored in the EU (United States · CLOUD Act). For the most sensitive data, compare trusted cloud, an open model hosted in-house and local AI.

Who is responsible for an AI agent: IT or the business?

Both, for different things. The business answers for the use and its decisions; IT answers for the agent’s identity, rights, logs and security.

Independent analysis, not legal advice. Up to date as of 29 October 2026.

Three concrete uses

01Mapping AI uses and reducing shadow AI

AI Act level · Minimal risk

F8 · Mapping AI uses and reducing shadow AI

01Detect usesLogs, licences, declarations

02Classify by riskData, purpose, supplier

CheckSign-off by the AI committee

03Decide the statusAllow, replace or ban

04Publish the catalogueWith the approved alternative

The problem: Staff paste contracts or customer files into consumer tools for want of an approved alternative. IT finds out about these uses by chance, often after the fact.
What the AI does: AI helps list uses from proxy logs, licence purchases and team declarations, and proposes a risk classification in a catalogue. It blocks nothing and sanctions no one: IT and the business decide what is allowed, replaced or banned.
Data needed: Proxy or cloud access broker logs, licence purchases and a questionnaire to teams.
Who signs off: The AI committee, with IT, the DPO and the business, approves each catalogue entry and the alternative offered.

Minimal risk (EU · AI Act); but if the analysis targets people rather than tools, it becomes monitoring of employees’ activity, to be presented beforehand to the works council and staff (France · Labour Code).

02Limiting the permissions of AI agents

AI Act level · Minimal risk

F8 · Limiting the permissions of AI agents

01Create an identityOne dedicated account per agent

02Limit the rightsRead by default, write when justified

03List the toolsApproved MCP servers and APIs

CheckApproval by an authorised person

04Approve the irreversibleSend, payment, deletion

The problem: An agent connected to email, the CRM or the ERP acts with whatever rights it is given, often administrator rights to save time. A malicious instruction hidden in a document can then turn it against you.
What the AI does: Each agent gets its own identity, rights limited to its task, a list of authorised tools and human approval before any irreversible action. These guardrails are enforced by the infrastructure, not by the prompt: the agent performs nothing irreversible on its own.
Data needed: The identity directory, the rights matrix per application and the list of tools exposed to agents.
Who signs off: The agent’s business owner approves its rights; a human approves every send, payment or deletion.

The level depends on the use, not the technology: the same agent screening job applications or assessing individuals’ creditworthiness would be high-risk (EU · AI Act, Annex III).

03Logging every AI action so you can prove it

AI Act level · Minimal risk

F8 · Logging every AI action so you can prove it

01Trace every callRequest, tool, response, version

02Link to identityAgent, user, business owner

03Retain and protectWritten retention, restricted access

CheckSign-off by the CISO

04Investigate an incidentAssisted search, human conclusion

The problem: After an incident or a complaint, nobody can say which agent did what, with which model version and on which data. Logs sometimes exist, but at the supplier’s and for a period it alone decides.
What the AI does: Every request, every tool call and every human approval is recorded in a log you control, linked to the agent’s identity and the model version. AI then helps search and summarise these traces; it does not alter or delete them, nor conclude the investigation.
Data needed: Logs from the AI gateway, agents and applications, with a written retention policy.
Who signs off: The CISO sets retention and access; any incident analysis is led and signed off by a person.

Logging is not a risky use but a form of evidence: for a high-risk system, the deployer will have to keep its logs for at least six months (EU · AI Act, Art. 26), and these logs, rich in personal data, must be secured (EU · GDPR, Art. 32).

What applies to you

EU · AI ActSince 2 February 2025, providers and deployers must take measures to build their staff’s AI literacy (Article 4); Regulation (EU) 2026/1744 turned this into a best-efforts obligation without removing it.

EU · NIS2If your organisation is an essential or important entity, management approves the security measures (Article 20), which cover the supply chain and access control among others (Article 21); a significant incident requires an early warning within 24 hours and a notification within 72 hours (Article 23). In France, the transposing law had not been finally adopted as at 29 October 2026.

EU · GDPRAny AI processing of personal data goes into the record of processing activities (Article 30), must be secured (Article 32) and, where the risk is high, must be preceded by an impact assessment (Article 35).

EU · Data ActCloud contracts, including SaaS, must allow you to switch provider; from 12 January 2027, switching charges are banned (Article 29).

What Clairancy does

A · Framing and decision

Build the catalogue of AI uses, the classification rule and the decision path between IT, the business and the DPO.

C · Sovereign AI and data

Design a sovereign AI architecture: a single gateway, logs under your control, a choice between trusted cloud, open model and local AI.

D · Governance and compliance

Write the usage policy, the agent rights matrix and the evidence file expected under the AI Act, the GDPR and, where relevant, NIS2.