AI does not need another dashboard if nobody can say who owns the decision.
1. Regulators are asking for evidence, not slogans.
ESMA will make AI and tokenisation an EU-wide supervisory priority from 2027. Ireland's Data Protection Commission has published lessons from its AI supervision: lawful basis for training data, transparency, data minimisation and children's protection. India's central bank has also called for validation, monitoring, human oversight and named accountability. None of these announcements is a new blanket AI law. The practical starting point is an inventory: which model, which data, which decision, which owner?
https://www.esma.europa.eu/press-news/esma-news/esma-sets-new-supervisory-priority-digital-innovation-2027
https://dataprotection.ie/en/dpc-guidance/publications/dpc-AI-insights-report
https://rbi.org.in/scripts/BS_SpeechesView.aspx?Id=1581
2. The supervisor is using AI too.
The UK FCA says it processes more than 56 million records daily to flag high-risk firms and is exploring "agentic supervision" as it prepares to take on AML oversight of legal and accounting firms from late 2028. Its speech is clear that AI will not replace human judgment. Firms should expect inconsistencies in returns, reports and risk data to become easier to spot.
https://www.fca.org.uk/news/speeches/financial-crime-protecting-hive
3. When agents spend, consent becomes a control.
Six banks have published voluntary principles for AI agents that choose and pay for things: transparency, safety, privacy and data, customer choice, and interoperability. They are principles, not regulation. The compliance questions are already here: did the customer authorise this purchase, what could the agent access, and who handles a disputed payment?
https://newsroom.bankofamerica.com/content/newsroom/press-releases/2026/09/global-banks-collaborate-on-principles-for-trusted-agentic-comme.html
4. The criminal use case is not theoretical.
Microsoft says it disrupted EvilTokens, a service that used access to compromised inboxes and AI to map payment relationships and help write impersonation messages. A convincing email from a real account is no longer enough proof for a payment change. Verify through a known channel, especially when bank details or approval routes change.
https://blogs.microsoft.com/on-the-issues/2026/09/22/disrupting-eviltokens-the-ai-chatbot-built-for-cybercrime/
5. The agent's reading list can become its attack path.
Mandiant's AI risk report describes indirect prompt injection through material an agent reads, including pages, emails and documents. Treat external content as data, not an instruction to act. Inventory agents and connectors, limit credentials and require human approval for high-risk actions.
https://cloud.google.com/security/resources/ai-risk-and-resilience-2026
The test for any AI deployment is simple: can you show what it saw, what it did, and who could stop it?
More at the RAW training hub: https://www.rawcompliance.ai
AI does not need another dashboard if nobody can say who owns the decision.
1. Regulators are asking for evidence, not slogans.
ESMA will make AI and tokenisation an EU-wide supervisory priority from 2027. Ireland's Data Protection Commission has published lessons from its AI supervision: lawful basis for training data, transparency, data minimisation and children's protection. India's central bank has also called for validation, monitoring, human oversight and named accountability. None of these announcements is a new blanket AI law. The practical starting point is an inventory: which model, which data, which decision, which owner?
2. The supervisor is using AI too.
The UK FCA says it processes more than 56 million records daily to flag high-risk firms and is exploring "agentic supervision" as it prepares to take on AML oversight of legal and accounting firms from late 2028. Its speech is clear that AI will not replace human judgment. Firms should expect inconsistencies in returns, reports and risk data to become easier to spot.
3. When agents spend, consent becomes a control.
Six banks have published voluntary principles for AI agents that choose and pay for things: transparency, safety, privacy and data, customer choice, and interoperability. They are principles, not regulation. The compliance questions are already here: did the customer authorise this purchase, what could the agent access, and who handles a disputed payment?
4. The criminal use case is not theoretical.
Microsoft says it disrupted EvilTokens, a service that used access to compromised inboxes and AI to map payment relationships and help write impersonation messages. A convincing email from a real account is no longer enough proof for a payment change. Verify through a known channel, especially when bank details or approval routes change.
5. The agent's reading list can become its attack path.
Mandiant's AI risk report describes indirect prompt injection through material an agent reads, including pages, emails and documents. Treat external content as data, not an instruction to act. Inventory agents and connectors, limit credentials and require human approval for high-risk actions.
The test for any AI deployment is simple: can you show what it saw, what it did, and who could stop it?
More at the RAW training hub: https://www.rawcompliance.ai
AI does not need another dashboard if nobody can say who owns the decision.
1. Regulators are asking for evidence, not slogans.
ESMA will make AI and tokenisation an EU-wide supervisory priority from 2027. Ireland's Data Protection Commission has published lessons from its AI supervision: lawful basis for training data, transparency, data minimisation and children's protection. India's central bank has also called for validation, monitoring, human oversight and named accountability. None of these announcements is a new blanket AI law. The practical starting point is an inventory: which model, which data, which decision, which owner?
2. The supervisor is using AI too.
The UK FCA says it processes more than 56 million records daily to flag high-risk firms and is exploring "agentic supervision" as it prepares to take on AML oversight of legal and accounting firms from late 2028. Its speech is clear that AI will not replace human judgment. Firms should expect inconsistencies in returns, reports and risk data to become easier to spot.
3. When agents spend, consent becomes a control.
Six banks have published voluntary principles for AI agents that choose and pay for things: transparency, safety, privacy and data, customer choice, and interoperability. They are principles, not regulation. The compliance questions are already here: did the customer authorise this purchase, what could the agent access, and who handles a disputed payment?
4. The criminal use case is not theoretical.
Microsoft says it disrupted EvilTokens, a service that used access to compromised inboxes and AI to map payment relationships and help write impersonation messages. A convincing email from a real account is no longer enough proof for a payment change. Verify through a known channel, especially when bank details or approval routes change.
5. The agent's reading list can become its attack path.
Mandiant's AI risk report describes indirect prompt injection through material an agent reads, including pages, emails and documents. Treat external content as data, not an instruction to act. Inventory agents and connectors, limit credentials and require human approval for high-risk actions.
The test for any AI deployment is simple: can you show what it saw, what it did, and who could stop it?
More at the RAW training hub: https://www.rawcompliance.ai