Skip to content
AI Automation

The Banks Stopped Piloting: At Sibos, BNY, BNP Paribas, HSBC And Deutsche Showed Agents Repairing Payments And Triaging Trades In Production, And Barclays Put Claude On 120,000 Emails A Day

The second day of Sibos in Miami was the first time a room full of the world's largest banks described AI agents as things that already run, rather than things they are testing. BNY's agents read payment fields for meaning and check them against ISO 20022 before a human sees them. BNP Paribas cut a securities-services workflow from ten steps to six and had agents doing 80-85% of the work within weeks, with staff teaching exceptions instead of processing them. HSBC's Smart Checking digitises trade documents and verifies rules, keeping humans for edge cases and high-value transactions. Deutsche Bank's Ada framework reuses agents across divisions and has taken a lending decision from a month to a day. The same day Barclays said Claude now routes about 120,000 Global Markets emails daily, serves 16,000 colleagues through a knowledge assistant, and will be in half its developers' hands by year-end. Swift's ledger is live with 17 banks in five currencies. The judgement, every bank insisted, stays with people - and the customer research explains why.

AlchmAI Editorial12 min read

10 → 6

Steps in a BNP Paribas securities-services workflow after agents took over, doing 80-85% of the work within weeks

1 month → 1 day

Deutsche Bank lending decision time with its Ada agent framework and knowledge graphs

120,000

Global Markets emails a day that Barclays now classifies, enriches and routes with Claude; 16,000 colleagues use its knowledge assistant

<50%

Of banking customers who would let AI execute a transaction, even though 68% accept conversational AI (Temenos research)

Sibos has hosted a decade of AI panels. This year's day two was different in one word: production. BNY described agents that read the fields of an incoming payment for meaning - not just format - check the syntax against ISO 20022 and flag the issues for a person to resolve, which is the unglamorous work of payment repair that eats operations headcount at every bank. BNP Paribas said agents had cut a securities-services workflow from ten steps to six and were doing 80 to 85 per cent of the work within weeks of rollout, with staff teaching the system how to handle exceptions rather than handling them. HSBC's Smart Checking digitises trade-finance documents and verifies them against rules, with humans retained for edge cases and high-value transactions. Deutsche Bank's Ada framework lets agents be reused across divisions, and with knowledge graphs has taken a lending decision from a month to a day.

Across the Atlantic the same day, Barclays put numbers on its own rollout. Its Colleague Knowledge Assistant, running on Claude since 2025, has more than 16,000 users and has handled over a million searches. In Global Markets, the models classify, enrich and choose a processing path for incoming client emails - about 120,000 a day. Claude Code is to be in use by half of Barclays' developers by the end of 2026 and a majority of its software engineers during 2027, aimed squarely at modernising ageing technology. Group co-COO Craig Bright described AI as 'an increasingly agentic capability' reshaping engineering and cybersecurity; the bank emphasised governance, security controls and human oversight throughout.

Why The Judgement Stays With People

Temenos research presented at the conference puts the customer constraint in numbers: 68% of banking customers would accept conversational AI, but fewer than half would authorise AI to execute a transaction, with privacy (47%) and the risk of errors (36%) the reasons. That maps almost exactly onto what the banks have built - agents everywhere in the back office, people on every decision a customer would notice. It also maps onto this week's wider news: OpenAI shelving a model that acted beyond its authorisation, and Nvidia shipping a hardware watchdog for agents. The banks' architecture is the regulated-industry answer to the same problem.

Swift: Rails For The Agents To Run On

  • Swift's shared ledger is live, with 17 first-mover banks across five currencies and at least 19 expected by year-end, for 24/7 tokenised-deposit payments; delivery-versus-payment and payment-versus-payment are next.
  • Citi became the first bank to launch multi-market instant payments across Swift's scheme, in Australia, the UK, India and the US; Swift reports more than 100 institutions live or launching on consumer payments.
  • Chainlink, Oracle and IBM each announced middleware to connect banks to the ledger within a week - the first vendor race for access to Swift's rails.
  • The IMF's Dan Katz argued the biggest near-term gains from AI will come from agents improving the payment system we already have: optimising fees, automating mechanics, cutting compliance costs. The BNY and BNP examples are exactly that.

What Smaller Institutions Should Take From It

The temptation is to read Sibos as a story about banks with ten-figure technology budgets. The transferable part is cheaper than that. Every example above began with one workflow, measurable in steps and hours, where exceptions are the cost and a reviewer already exists. Payment repair, KYC refresh, claims document checks, broker email triage and credit-memo preparation all qualify, and all can be built on the same pattern in a quarter: agents on the routine majority, people on the exceptions, an audit trail of both. Barclays' email-routing number is the clearest version - 120,000 messages a day is a volume no human team should be triaging, and the model does not need to make a single decision to be worth it.

“The banks did not announce that AI would change everything. They announced which steps it had already taken over, and which they had deliberately kept.”

The Bottom Line

Sibos 2026 was the week banks stopped describing AI agents in the future tense: BNY repairing payments against ISO 20022, BNP Paribas cutting a workflow from ten steps to six with agents doing most of the work, HSBC checking trade documents, Deutsche Bank reusing agents across divisions to take lending from a month to a day, and Barclays routing 120,000 emails a day and serving 16,000 colleagues with Claude while putting it in half its developers' hands. Swift's live ledger and instant-payment scheme give those agents rails to run on. Every bank kept consequential judgement with people, and customer research says they are right to. For any financial institution, the model is now proven and repeatable: one high-volume workflow, agents on the routine, humans on the exceptions, an audit trail of both. That is the AI and workflow automation we deliver in London, and this week the largest banks in the world published the template.

References & Further Reading

AI AutomationAgentic AIWorkflow Automation AgencyBanking Portals & InterfacesClient-Facing & Internal ToolsSibos 2026AI Agency London
Share Email
AI

AlchmAI Editorial

Research and analysis, London

The AlchmAI team writes about the markets, technology and regulation we work with every day. We build trading platforms, real-time charts and AI analysis tools for brokers, prop firms and fintech teams from our office in Mayfair, London. Every article lists its sources. Nothing we publish is investment advice.

This article is general information and commentary. It is not investment advice or a recommendation to buy or sell any investment. Important information

Keep reading

All articles