AI is on every desk in the bank. So is your customer data.
Your people already use AI to draft emails, summarise cases and debug code. The risk is what they paste into it: customer records, IBANs, card numbers, credentials, source code. Tracehold checks what is sent to AI tools and, according to your policy, observes, warns, redacts or blocks. If nothing sensitive is found, the prompt goes through.
The fastest way for customer data to leave the bank is a text box
Generative AI arrived through the browser: helpful, fast, and one paste away from sending regulated data to a service outside your controls. The intent is almost always innocent. The exposure isn't.
Customer & account data
A support agent pastes a customer record (name, address, IBAN, balance) into an AI tool to draft a reply. A relationship manager summarises a client's portfolio. Each one is regulated personal and financial data leaving your control.
Transactions & statements
Analysts paste transaction logs, payment files and statement extracts to "explain this anomaly" or "format this table." Card numbers, counterparties and amounts go straight into a prompt.
Code, keys & internal docs
Developers paste core-banking code with embedded credentials into AI assistants to debug it. Risk and compliance teams drop confidential policies, audit findings and board memos in to summarise them.
Built to recognise the data a bank can't afford to lose
The extension's engine runs in the browser with more than 1,600 rules. Rules match data that has a recognisable shape; free-form text with no recognisable shape, such as a customer's name inside a sentence, needs the optional LLM classifier (Governance plan, beta; text in transit, not stored by Tracehold).
Credentials & keys
API keys, access tokens, passwords, private keys and connection strings hidden inside pasted code or config, caught before they reach an AI assistant and become someone else's.
Personal & financial data
IBANs, card numbers and identity numbers, validated with their check digit, so a number with a wrong digit is not flagged. Identity documents from other countries have their own rules.
Confidential documents
Documents are recognised by their markings (STRICTLY CONFIDENTIAL, TOP SECRET, DO NOT DISTRIBUTE) and by your own keywords. The engine does not read a memo and decide that it is confidential.
Checks it before the send
Detection runs in the browser, before the prompt is sent to an AI tool. The browser extension needs no proxy. The console receives event metadata plus a masked sample of the value; saving the content itself is off by default.
Observe, warn, redact or block: your policy decides
Observe logs the finding without interrupting. Warn tells the person. Redact replaces the sensitive value with a marker so the rest goes through. Block stops the send and says why, so people learn, not just get stopped.
Beyond the browser tab
The Windows agent covers USB drives, printing and screen captures of protected Office documents. The Linux agent covers terminal AI tools such as Claude Code and Codex, when the policy turns it on. Neither intercepts the network by default.
Not just pasted text: attached files too
The same check applies to files attached or uploaded on AI sites.
Uncontrolled AI use touches every framework on your desk
Banking is one of the most heavily regulated places AI can go wrong. The same paste can trip more than one rulebook at once.
DORA
Digital operational resilience means knowing and controlling the third-party digital tools your staff feed data into. AI tools are exactly that. Tracehold gives you visibility and control over that flow.
GDPR
Pasting a customer's personal or financial data into a public AI tool can be an unlawful transfer to a third party. Checking it before the send helps keep that data from leaving your control.
NIS2
Stronger cyber-risk management and accountability for essential entities. Demonstrable control over how data leaves to AI tools, backed by an audit trail, is part of showing you manage the risk.
EU AI Act
As you adopt AI, you need to know where and how it's used, and govern it. The Governance plan maps findings to 8 controls, including Art. 26 (AI systems in use by the deployer), Art. 12 (record keeping) and Art. 14 (human oversight).
Compliance reporting you can put in front of the board
In the Governance plan, Tracehold maps what it finds to 39 controls across six frameworks (GDPR, the EU AI Act, DORA, NIS2, ISO 27001 and SOC 2) and exports a report for each, in PDF or CSV. Each finding is kept as a record of its type, severity, destination and action, with a masked sample of the value rather than the full text, so your team can show how AI use is controlled.
Risk, not theatre
The mapping is built from what your people are actually doing with AI: what is caught, how often and how severe, rather than a point-in-time checklist.
Evidence on demand
Every finding is recorded with its type, severity, destination and action, plus a masked sample of the value rather than the full text.
Support for your audits, not a certification
The mapping is a product judgement about which detections evidence which obligation. It supports your audits; it is not legal advice or a certification. Review it with your compliance team.
From "we think we're fine" to "here's the number"
Most banks can't answer a simple board question: how much regulated data is going to AI right now? Compliance reporting helps answer it.
Governance reports cover GDPR, the EU AI Act, DORA, NIS2, ISO 27001 and SOC 2, in PDF or CSV. See compliance in depth →
No proxy, no network surgery.
A bank can't bolt a new inline proxy into its network on a whim. The browser extension needs no proxy and no network change, and you push it with the device management you already run.
Pushed via your MDM
The console generates the managed-policy files (.reg, .ps1, .json) and the MSI, and there is a guide for Microsoft Intune. Packages and policy come from the console; your MDM or GPO does the rollout.
Start observe-only
Begin in a pilot that watches and reports without blocking anyone. You see what would have been caught and where your real exposure is before you turn on warnings, redaction or blocking.
Plugs into your stack
Sign-in with OIDC or LDAP/AD. Findings from the browser extension go to your SIEM over HTTPS (Splunk HEC or CEF), and alerts go to Slack, Microsoft Teams or your own webhook URL.
What banking buyers ask first
Does customer data reach Tracehold?
Do we need a proxy or network changes?
How does it help with DORA, GDPR, NIS2 and the EU AI Act?
Does Tracehold hold any certification?
Does it recognise IBANs and card numbers?
Which plan do we need?
See what your bank is already sending to AI
Start an observe-only pilot and get a clear, honest picture of where regulated data is leaking to AI tools, with zero disruption to your people.