In early June 2026 the Financial Conduct Authority confirmed it won’t draft new rules specifically for artificial intelligence. Firms are expected to apply the existing frameworks to their automated systems[1].

For the owner of an advice firm, that removes one of the more popular reasons for waiting. The rules aren’t going to change because the technology has. Under the Senior Managers and Certification Regime, accountability for an automated decision sits exactly where it sat for a manual one: with the named senior manager.

Financial services is now moving AI agents from pilot into production faster than any other sector, at a 21 per cent deployment rate this summer according to one engineering blog’s tracking[2]. As those systems move from internal testing to live client data, the compliance exposure grows with them.

Automation is moving from the back office to client-facing work

Firms are moving automation budgets towards the front office because that’s where the return shows up fastest. Until recently most of the spend went on internal jobs: pulling data out of provider statements, drafting meeting summaries for the file.

Now the systems touch the client. Digital onboarding is the obvious example. A 2023 University of Edinburgh study found automated onboarding cut KYC processing time by 34 per cent in regulated financial services[3], although the figure is three years old and the tools have moved on since. The saving comes from routing identity documents and flagging discrepancies before an administrator opens the file.

The moment a system handles client data or onboarding, the regulatory attention goes up with it. The European Central Bank has said customer-facing AI carries a real risk of discriminating against customers if it isn’t properly validated[4]. If a system decides how quickly a client gets through onboarding, or how their risk profile is assessed, that’s a Consumer Duty matter, not an IT project.

The bigger risk is how much access you give the system

The main security threat in automation today is giving an AI agent wide access to internal data with nobody checking what it does. VentureBeat reported in July that 54 per cent of enterprises have already had an AI agent security incident, usually because agents were allowed to share credentials across tools[5].

Agentic AI, where the system carries out a multi-step task across several applications rather than answering a question, is where most of the interest is. A vendor study by Digital Applied put interest at 75 per cent of businesses and found that 40 per cent of these projects get cancelled over deployment and integration problems[6]. In most of those cases the thing that failed was governance, not the model.

Risk does not disappear because a process is automated. It moves somewhere else.

Connect a model to your CRM, your document store and your email, and you’ve created a new way in. The FCA, the Bank of England and HM Treasury have jointly warned that frontier AI models now go beyond what baseline cyber resilience was designed for[7]. A workflow that can read a client file, write an email and send it without a person checking the output is carrying risk it doesn’t need to.

Where this leaves advice firms

Read together, these points say the regulator has handed the problem back to the firm. There’s no AI rulebook to wait for, so the test is the one that already applies to your advisers: can you show who approved the decision, what they saw, and why.

The firms that will get through the next round of supervision are the ones that can answer three plain questions about every tool: what can it see, what can it do without asking, and who signs off its output. A tool that only reads data and drafts is a small governance job. A tool that can send, approve or update client records on its own is a Senior Manager’s problem, whatever the vendor calls it.

My prediction is that the first enforcement cases involving AI in advice won’t be about the model getting something wrong. They’ll be about a firm that couldn’t show a human reviewed what it sent.

This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.

Sources

[1] Financial Conduct Authority, ‘Approach to AI in financial services’, Financial Conduct Authority, 9 June 2026. Available at: https://www.fca.org.uk/news/blogs/ai-financial-services-approach [2] Zen van Riel (2026) ‘Why 78% of AI Agent Pilots Never Reach Production’, AI Engineer Blog. Available at: https://zenvanriel.com/ai-engineer-blog/ai-agent-scaling-gap-pilot-production-2026 [3] University of Edinburgh (2023) ‘Digital onboarding reduces KYC processing time by 34% in regulated financial services’, University of Edinburgh. [4] European Central Bank (2024) ‘Customer-facing AI in financial services presents substantial risks of customer discrimination’, European Central Bank. [5] VentureBeat, ‘The agent security gap: 54% of enterprises have already had an AI agent incident’, VentureBeat, 17 July 2026. Available at: https://venturebeat.com/ai/the-agent-security-gap-54-of-enterprises-have-already-had-an-ai-agent-incident-and-most-still-let-agents-share-credentials [6] Digital Applied (2026) ‘Why Agentic AI Projects Get Canceled (and How to Ship)’, Digital Applied. Available at: https://www.digitalapplied.com/blog/agentic-ai-project-cancellations-gartner-40-percent-2026. Vendor-sourced. [7] Financial Conduct Authority, ‘FCA, PRA and Bank of England joint statement on frontier AI and cyber resilience’, Financial Conduct Authority, 25 May 2026. Available at: https://www.fca.org.uk/news/statements/fca-boe-treasury-joint-statement-frontier-ai-models-cyber-resilience