One in ten financial firms can orchestrate the AI agents they deploy.
Capgemini: 59% are piloting or deploying multi-agent AI. Eight in ten list reducing tech debt as a key objective. Roughly half cannot show a return.
The order looks backwards. That is the bill.
Bank of America told customers not to give AI agents their password.
The same week, Schwab let 17,000+ RIAs reach its data from Claude “through an authenticated connection”.
How the agent gets to the data is now part of the product.
The government's AI survey found insurance gaps holding back some finance firms.
One finance business: “very few, if any, PI policies out there that’s happy to provide cover for AI.”
Guidance says what a firm may do. Cover decides what it will do.
In financial services, every AI tool now comes with a compliance question.
FS marketers see AI’s gains at the top of the funnel. In 6sense’s 2025 research, 58% of B2B buyers engaged vendors sooner to ask how AI was implemented.
I think the AI checks are where the cycle slows.
Apollo’s chief economist sees zero AI payoff in corporate America outside tech.
Torsten Slok: “AI may be a great calculator, but in business, intelligence is mostly judgment.”
The model takes the calculation. People keep the judgement.
24% of responding firms report deep knowledge of the AI they buy.
For systems built in-house it is 77%, in the Bank of England and FCA’s 2026 AI survey. For bought AI, 81% rely on vendor assurance and contracts.
The evidence pack is the pitch now.
Clay’s first self-learning outbound loop made its own emails worse.
It ranked rep edits by how often they happened, not by their impact. The fix began with writing down what good copy is.
Judgement first. Then the loop.
AI tools influence an estimated £47.4bn of UK savings and investments.
By the lang cat’s reckoning, that would put AI among the UK’s five largest financial advice firms.
The advice gap did not close. It was filled.
Everyone clicks approve on AI now, and the FCA has noticed.
Its 19 October publication lists “excessive reliance on human-in-the-loop controls” as an area for improvement.
Auto-approve is a setting. Accountability has a name on it.
Bank of America needed 1,000 consultants to act on what AI found.
Inside most firms, I think, the gaps AI surfaces are ones someone has quietly managed for years.
An AI stack cannot use what nobody wrote down. Document first. Then expect the efficiency.
JPMorgan says customers who use outside AI agents carry the risk.
Jamie Dimon: “There’s a liability transfer.” The same week, Amex had the customer sign off the agent’s intent before it buys.
Same question, opposite answers. Trust is moving to the permission slip.
I read Anthropic's new Usage Policy. One line for advisers.
A personalised investment recommendation needs a qualified person who can change it, and the client told AI was used. Explaining how investing works does not.
The reviewer is part of the product.
Tactics and tools alone do not deliver results.
89% of organisations use AI in at least one function (McKinsey). High performers held flat at 6% both years.
Among the earliest adopters, those that redesigned workflows were 5.3x as likely to report enterprise-level value.
MAS's new AI guidelines have one line for every vendor.
"FIs remain accountable for AI used in the services they deliver, including AI developed, operated or provided by third parties."
The sale no longer ends at "the model works". It ends when the bank can evidence assurance.
12% against 37%. HSBC's own numbers.
HSBC's June survey: at the decision, 12% of investors named AI the most influential source. 37%, professional guidance.
The FT reports UK wealth adviser cuts could reach around 70%.
The sequence stays. The person at the last step goes.
Vanguard asked 6,686 investors about AI. Advisers came out ahead.
AI is 11 times as likely to raise the value they place on human advice as to lower it.
Its one edge: four in 10 users think it beats advisers on avoiding conflicts of interest.
70% of enterprises will abandon vendor-built agentic AI by 2028.
Gartner's diagnosis: these engagements "often fail structurally before they fail technically".
The capability that does not transfer is the cost that does not end.
Six banks wrote five principles for AI agents. None binding.
Their own diagnosis: "Consumers are unclear if AI agents will act in their interests."
Non-binding on paper. In practice, it is the list a bank partner's risk team will hold an agentic product against.
56% trust it. 44% think it's regulated.
The FCA's own survey: over half of 18 to 40 year old investors now trust AI with their money more than TV or the press. 44% wrongly think that information is already regulated.
The trust problem is solved. The understanding one is not.
I read the FCA's money-mule review. The second account matters.
35 firms closed the accounts of 238,396 suspected mules in 2025. Cash-out mainly came between the second and fifth mule accounts.
Closing an account ends the mule's time at one firm. The chain carries on.