Eleven controls before a single byte reaches a model provider
Generative AI is already inside the bank
One breached organization in five found shadow AI in the incident. That figure comes from IBM's Cost of a Data Breach Report 2025, which also puts the extra cost at 670,000 dollars per breach where shadow AI is widespread. In the same report, 97% of the organizations hit by an AI-related security incident had no AI access control in place.
So the useful question for a CISO is no longer whether to allow generative AI. It is which door the traffic uses today, and who can prove it.
A ban does not answer that question. People move to a personal account and the traffic leaves the field of view. A ban is a memo. It is not a control.
One door, eleven controls
Central IA Souveraine is the product we built in 2026 around that question. Every AI use in the institution goes through one internal address. No employee holds a provider key. Between the employee and the provider, eleven controls run in order. Eight of them run before anything leaves the perimeter.
What the controls decide
Sensitive data. National ID numbers, bank account identifiers such as RIB and IBAN, company identifiers such as ICE and IF, passports, card numbers and secrets are detected in the message and in its attachments: PDF, Word, Excel, PowerPoint, images. A face in a document is treated as biometric data. The institution's policy then chooses to block, to mask, or to replace the value with a reversible token. With a token, the provider never sees the real value and the employee still reads it in clear.
Destination. Each request gets a class from D0, public, to D3, regulated. Text and documents are classified together and the highest level wins. The class decides where the request may go: allowed, anonymized first, restricted to a residency such as on-premises only, or rejected. The routing matrix is data that the institution edits. It is not a frozen configuration.
Authorization. No use case becomes active without a reviewed request, an owner who attests, and a reasoned decision. The committee quorum grows with the data class. Nobody approves their own request, and administration and governance roles are held by different people.
What the auditor sees
Every request and every response is captured byte for byte. Every decision is written to a hash-chained ledger that can be verified end to end and exported to the SIEM.
The right to erasure is executable. Erasing means destroying the key: the data becomes unrecoverable, the operation itself is audited, and the cost line stays for financial integrity.
Cost gets the same treatment. Budgets per team and per use case, a cap that really blocks, a cost per request, scheduled reports. The question "how much does AI cost us this month, and for what" gets a number.
The knowledge vault
The second piece is the Coffre de connaissances, the knowledge vault. Today employees feed AI by pasting whole documents, again and again. Each copy costs tokens, travels, and escapes control.
In the vault a document exists once, inside the institution's perimeter. The model receives only the passages that answer the question, with a citation. A regulated document needs an approval before any use. Someone who leaves loses access the same day. And when nothing in the vault answers, it says so instead of inventing.
What it costs
A gateway is not free. It adds latency, and it becomes a single point of passage that must be operated like any critical component. That is the price of a complete audit trail, and we prefer to say it before anyone asks.
It cannot be switched on in one go either. The pilot runs twelve weeks, in three phases, with written exit criteria. We observe before we constrain: each control is measured on real traffic before it is enforced, with a rollback window.
The product is delivered signed and deployed in the institution's own cluster. Data, keys and logs stay there. The administration console has 21 screens, in French, Arabic and English.
To see the eleven controls on your own traffic, write to us.