AI in Banking and Finance: HSBC Automates Trade Document Checks, Opens Bank Data to Clients' AI Tools, and a Cloned Voice Helps Fraudsters Take €95m From Fideuram
Title: AI in Banking and Finance: HSBC Automates Trade Document Checks, Opens Bank Data to Clients' AI Tools, and a Cloned Voice Helps Fraudsters Take €95m From Fideuram
Date: 1 October 2026
Type: Paper
Author: SAASiQ (contact@saasiq.ai)
Word count: 2988 words
Reading time: 12 min
Published: 01-10-2026
HSBC announced Smart Checking on 28 September, an AI system that reads and checks trade finance documents. Smart Checking is already in use in the UK, Hong Kong and the UAE. On 29 September HSBC launched HSBCnio, which lets corporate clients' own AI tools query permitted account and transaction data. The same day OvationCXM published a survey of 520 US finance leaders. In it, 54 per cent said they would find a workaround if their bank could not support their AI agents. Reports from 25 September described how the Italian private bank Fideuram sent about €95m to fraudsters who used a cloned voice. This paper sets out how each use works, what is known about cost, and where it breaks.
What happened this week
This is the first of SAASiQ's Thursday papers on how AI is used in practice, sector by sector. Finance comes first. The week from 24 to 30 September produced three working examples from banks and one expensive failure.
HSBC announced Smart Checking on Monday 28 September, according to Disruption Banking and Global Trade Review. On Tuesday 29 September it launched HSBCnio, a digital service for corporate and institutional clients, as FinTech Global, Payment Expert and FStech reported. Also on 29 September, OvationCXM, a US company that sells customer service software to banks, released a survey of finance leaders about AI agents. Fortune published a report the same day on how Bank of America and Morgan Stanley measure the return on their AI tools.
The failure came from Italy. Reuters reported on 25 September that Fideuram, the private banking arm of Intesa Sanpaolo, had transferred about €95m to fraudsters in February. The fraudsters impersonated senior executives and a lawyer, and Reuters sources said they used AI to copy the lawyer's voice. The OECD's AI incident monitor logged the case the same day.
Use case one: checking trade documents
Trade finance still runs on paper and PDFs. Under a letter of credit, a bank promises to pay an exporter once the exporter presents documents that match the credit's terms. These are typically an invoice, a bill of lading, an insurance certificate and a certificate of origin. Staff check each set against the credit and against each other. If the documents comply, the bank pays. If they do not, the bank can refuse.
The rules most letters of credit follow are the International Chamber of Commerce's UCP 600. They give a bank a maximum of five banking days after presentation to decide whether the documents comply. So the checking work has a fixed clock, and an error in either direction costs money. Paying against bad documents exposes the bank. Refusing good ones delays the exporter's cash.
Smart Checking splits the job in two, according to HSBC's description reported by Global Trade Review. First, AI models extract and classify information from the documents and turn "complex and inconsistent documentation into structured data". Second, an AI reasoning engine supported by a knowledge graph interprets that data and performs the documentary credit checks. A knowledge graph is a structured map of facts and the links between them, here the rules and the relationships between documents. HSBC's trade specialists stay in the process and apply judgement to the cases the system cannot settle.
HSBC built the system itself. Global Trade Review notes that this goes against the usual pattern of banks buying document-checking platforms from fintech suppliers. The bank says the design gives "explainable and auditable outcomes", meaning each decision can be traced back to the documents and rules behind it. It is live in three markets and HSBC says it will roll it out globally over time. For clients, HSBC's case is speed: faster checking can bring forward payment and free up working capital sooner.
HSBC's announcement gave no accuracy figures, volumes or costs for Smart Checking. A bank or corporate considering anything similar would want three numbers from a pilot: the share of presentations the system clears without a person, the rate at which specialists overturn its findings, and the time from presentation to decision against the five-day limit. The first two together show whether the system is saving work or just moving it.
Where document checking breaks
In our experience, extraction is the weak step in document automation. Trade documents arrive as scans, photographs and PDFs made by many different systems, with stamps, handwriting and abbreviations. A field read wrongly at the start passes a wrong value to every later check. HSBC's description keeps extraction and reasoning as separate steps, with specialists on the exceptions.
The second risk is drift. The rules in UCP 600 change rarely, but each letter of credit has its own terms, and the ICC publishes separate guidance, known as ISBP, on how to examine documents. A knowledge graph has to be kept up to date by people who know the practice. If that upkeep stops, the system goes on applying old readings with full confidence.
The third risk is in the people. If specialists only see the hard cases, they see fewer routine ones, and routine cases are where junior staff learn the work. In our view, a bank running this kind of system should keep a sample of cleared presentations going to people for review. That checks the system and keeps the skill alive.
Use case two: letting clients' AI tools read bank data
HSBCnio is HSBC's new digital service for corporate and institutional clients. It brings web and mobile access, embedded connectivity and AI tools into one service, according to the reports of 29 September. Clients can view cash positions and transactions, manage trade loans and foreign exchange, and track payments.
It has three parts. The first is the web and mobile front end, giving one view across cash, transactions, trade loans, foreign exchange and payment tracking. The second is a set of APIs, the interfaces that let one system call another, with documentation, software development kits and a sandbox for testing. The third is AI access through what HSBC calls its Model Context Protocol connection. The Model Context Protocol is an open standard that lets AI tools connect to outside data sources. Through it, a client's own AI tools can query permitted account and transaction data.
The practical change is who asks the question. A treasury analyst used to log in to the bank portal, run a report and paste it into a spreadsheet. With an AI connection, the analyst can ask the company's own assistant for yesterday's cash position across accounts, and the assistant fetches the answer from the bank. The bank decides which data the client's tools may see. The client decides which of its tools may ask.
The OvationCXM survey shows the demand behind this. OvationCXM surveyed 520 US finance leaders. Of them, 75 per cent said they already use or test AI agents across finance and banking work. Three of the six most common finance tasks they give to agents involve the bank: tracking transactions, approving payments and starting payments or transfers. And 68 per cent expect their agents to make transactions, with 28 per cent expecting to use agents for ACH payments and 23 per cent for wire transfers in the next 12 months. ACH is the US system for batch payments such as payroll.
The figure in the headline of the release is the 54 per cent who would find a workaround if their bank could not support their agents. OvationCXM describes the workarounds as an internal or third-party fix, or moving banking business elsewhere. OvationCXM sells software to banks, so it has an interest in this result, and the survey covers the US only.
Where client AI access breaks
HSBCnio's AI access is described as querying permitted data. That is read access. The survey's respondents want more: two thirds expect their agents to make payments. Reading a cash balance and sending a wire carry very different risks, and the step from one to the other is where controls matter most.
The first control is scope. A client's AI tool should see only the accounts and data it needs, under its own credentials, separate from any person's login. BNP Paribas set out the same principle on 24 September for its own agents built with Google Cloud: each agent must authenticate and can reach only the resources assigned to its task. SAASiQ covered that agreement on 28 September.
The second control is the record. When a person runs a report, the bank's logs show who ran it. When an assistant runs it on a person's behalf, both the client and the bank need logs that show which tool asked, for whom, and what came back. Without that, an audit trail stops at the tool.
The third is the workaround itself. If a bank does not support agents, the survey says many finance teams will find another route. In our view, the likely routes are a third-party tool holding bank credentials, or an agent driving the bank's own web portal as if it were a person. Both are harder to control than a supported connection. For a bank, that is an argument for offering a supported, logged route. For a finance team, it is a reason to ask IT what agents already touch the bank before a supplier asks.
Use case three: help for staff on calls
The most widely deployed use in banks so far is help for staff, and Bank of America has published the clearest numbers. Its tool EricaAssist works alongside more than 18,000 customer service representatives during calls, according to the bank's announcement of July 2026. It summarises why the client is calling, pulls together the relevant information and suggests next steps based on the employee's role and the client's relationship with the bank. The bank says it delivers these insights in under three seconds and has cut average call times by nearly one minute per call.
Fortune's report of 29 September adds that an internal version, Erica for Employees, is used by 90 per cent of Bank of America's staff. The bank also uses AI to help developers write code, to draft preparation documents before client meetings and to help investment bankers build pitchbooks. Morgan Stanley's internal tool, AI@MS, dates from September 2023. Fortune reports that the bank works with several AI companies, including OpenAI, Anthropic and xAI, and is trying to hand the processing tasks to language models while advisers keep the client relationships.
A minute per call can be turned into money. Multiply it by call volume and the cost of an hour of an agent's time, and the saving can be set against the licence and running cost of the tool. Bank of America has not published those costs. Lloyds Banking Group is one of the few banks to put a value on its AI work as a whole: in early 2026 it said generative AI had produced about £50m of value in 2025, and it expects more than £100m in 2026.
An assistant that summarises a call wrongly, or suggests the wrong next step, relies on the person to notice. The control is the same as in document checking: sample the outputs, measure how often staff override them, and treat a falling override rate with some care. It can mean the tool is improving. It can also mean staff have stopped checking.
Where it breaks hardest: payment fraud with cloned voices
The Fideuram case shows the other side of AI in finance. According to Reuters and later reports, the then chairman of Fideuram, Paolo Molesini, received a WhatsApp message in February that appeared to come from Carlo Messina, chief executive of the parent company Intesa Sanpaolo. It asked for urgent help with an overseas transaction. Molesini then received a call from someone posing as Paolo Nastasi, managing partner of A&O Shearman in Italy, who confirmed the transaction was genuine. Nastasi was not involved. Reuters sources said the fraudsters used AI to copy his voice.
The bank sent about €95m, mainly to accounts in China and Hong Kong. Reports put the amount recovered at about €53m and the amount still missing at about €36m; the reports do not explain the gap between those figures. Italian prosecutors are investigating a foreign national, and other similar scams aimed at companies.
Fraud teams know the method as CEO fraud or business email compromise: a message from someone senior, urgency, secrecy, and a payment to a new account. AI changed one step. The second call, from a trusted outside lawyer, is the check many organisations rely on to confirm an unusual request. A cloned voice turned that check into part of the fraud.
Other figures from the week point the same way. BioCatch said on 30 September that attempted banking scams across more than 370 financial institutions rose 35 per cent over 12 months, and that 90 per cent of the attempts happened on mobile devices. Thomson Reuters published a report on 24 September arguing that transaction monitoring struggles when a coerced victim authorises the payment themselves.
Agents add a new version of the same problem. On 22 September a group of five banks, NatWest, Bank of America, ING, ASB Bank and Capital One, set out principles for AI agents that shop and pay for consumers. They warned that agents could ask for card details and type them into websites, or steer users towards payment methods with weaker protection. They want disclosure when an agent is involved in a transaction, more transparency on how agents decide, and safeguards for customer data.
What it costs and who pays
Most of the costs in this week's examples are not public. HSBC's announcements gave no price for HSBCnio and no build cost for Smart Checking. Bank of America gives a time saving but no spend. The figures that do exist come from the banks themselves, and SAASiQ has found no independent check of them.
The cost structure is still clear enough to plan around. Document checking and call support are internal tools: the bank pays to build or license them, pays to run the models, and pays specialists to handle exceptions and keep the rules current. The saving shows up as staff time and faster turnaround. AI access for clients is different. The bank carries the cost of the connection, the permissions and the logs, and the client's own AI tools run on the client's budget. The OvationCXM survey suggests some clients would pay for that access, which would make it a priced service rather than a cost of doing business.
Fraud sits outside both. In our view, any plan that adds AI to payment work should cost the controls alongside the tool, because the controls are what stop the large losses.
What UK regulators are doing
UK regulators are handling AI under their existing rules rather than writing new AI-specific ones. The FCA's AI Live Testing programme lets firms try AI systems in the market with the regulator watching. Its second cohort, announced in April, has eight firms: Aereve, Barclays, Coadjute, Experian, GoCardless, Lloyds Banking Group (through Scottish Widows), Palindrome and UBS. Their uses include agentic payments, anti-money laundering detection, Know Your Customer checks and credit score insights for consumers. Testing began in April and ends by the end of 2026, with the FCA's evaluation due in the first quarter of 2027.
On 2 September the FCA published findings from a review of frontier AI and cyber resilience across several firms, aimed mainly at small and medium-sized firms. The Bank of England said in its April response to the Treasury Committee that it plans to study what happens if AI agents across firms behave in the same way at the same time. That is a financial stability question: many firms acting on the same signals can move a market further than any one of them would.
What a finance team can do on Monday
Make a list of where AI already touches money. That means any tool that reads bank data, drafts payments, approves invoices or matches transactions, including features that arrived in a software update. For each one, record what it can read, what it can do, and under whose credentials.
Ask the bank what it supports. If the bank offers APIs or a connection for AI tools, ask what data is in scope, how permissions are set, and what the bank logs. If it offers nothing, find out whether anyone in the team is already using a workaround, such as a third-party tool holding bank logins.
Keep read and pay apart. An AI tool that reports cash positions needs read access only. Any step that moves money should need a separate approval by a named person, through a channel the tool cannot use.
Change the call-back rule. Many payment controls say to confirm an unusual request by phone. After Fideuram, that call should go to a number already held on file, never to a number or contact given in the request, and a voice alone should not count as proof. A code word agreed in advance between senior staff and the payments team is a cheap addition.
Measure before scaling. For any AI tool in finance, set the measures before the pilot: how much work it clears, how often people overturn it, and how long the whole process takes. HSBC and Bank of America have published only some of these figures, and a buyer will need all three to judge the case.
Ask suppliers the same questions. Any supplier offering an AI feature in finance should say which model runs underneath, where the data goes, what is logged, and how quickly it will report an incident involving the customer's data.
What happens next
HSBC says Smart Checking will be rolled out to more markets over time, and has given no dates. The FCA's second AI Live Testing cohort ends by the close of 2026, with the evaluation due in the first quarter of 2027. Italian prosecutors' investigation into the Fideuram fraud continues.
SAASiQ - Intelligent Solutions for SaaS ©
