What problem does it solve?
Sending money in a banking app means finding the right menu, choosing a payee, typing an amount and a reference, and passing a confirmation step. Each step is a place where a customer can stall, and an assistant that only explains where the transfer screen is leaves the customer to do the work.
A new group of assistants takes the instruction itself. A customer says or types "send 10,000 yen to Tanaka", sends a voice message with a name and an amount, or photographs an invoice, and the assistant takes it from there. Nubank in Brazil, Bank ABB in Azerbaijan and Boost in Malaysia have published versions of this, SBI Sumishin Net Bank in Japan began a beta of one in February 2026, and Itaú Unibanco is rolling one out in its app. The releases differ on approval: Itaú describes an explicit confirmation before the Pix closes, Bank ABB describes the transfer being completed right after the customer names the contact, and the others do not say.
This page covers the payment instruction a customer gives in the moment. The customer asks for each payment, and whether the bank adds a confirmation step is a design choice that the recommendations below settle in favor of one. An agent that buys from merchants on its own within a spending mandate is covered by the agentic payment initiation page. Routine servicing such as card blocks and statements is covered by the account and card servicing page, and stopping a payment that looks like a scam is covered by the scam payment interception page.
How does it work?
- Understand the instruction. The assistant reads text, a voice message or an image and extracts the intent (send money, pay a bill, move money between own accounts), the amount, the currency and a reference to the payee. Some assistants accept several recipients in one message, and some read an invoice or a screenshot.
- Resolve the payee from known data only. The assistant matches the name against saved contacts, earlier payees or the bank's own directory. Bank ABB, for example, requires the mobile number to be saved in the sender's phone contacts. An unknown name or an ambiguous match goes back to the customer as a question.
- Show the payment before it moves. The assistant presents the resolved payee, the amount and the reference on screen. Itaú Unibanco's assistant builds a Pix from a voice command and closes it only after an explicit confirmation from the customer. Bank ABB's assistant completes the transfer right after the customer names the contact, so each bank decides this step for itself.
- Authenticate and apply limits. The payment goes through the same strong authentication, per transaction and daily limits and fraud analysis as a payment made by hand. Nubank applies its app risk analysis to Pix sent on WhatsApp, with a cap per transaction and per day during the test phase.
- Execute through the bank's payment API. A service call to the payment API creates the payment, with an idempotency key so a retry cannot pay twice.
- Confirm, record and hand over. The assistant reports the result, keeps the instruction, the confirmation and the payment together as one record, and passes anything it cannot resolve to a person or to the normal app screen.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Mobile app, WhatsApp, Phone and voice, Web chat
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Users served | Not pooled | 30,000 to 2 million | 2 | 2 organization |
| Interactions handled | Not pooled | at least 1 billion | 1 | 1 organization |
| Cycle time reduction | Too few to pool | Not pooled: up to 60% | 0plus 1 up to | 1 organization |
Value drivers: Customer experience, Speed and cycle time, Inclusion and access, Lower cost to serve.
Indicative value
A retail bank with 2 million digitally active customers
USD 80,000 to USD 1.2 million
Annual servicing cost avoided on payment requests handled by phone or branch per year
How this is calculated
Formula: customers * assistedRequestRate * assistantShare * costPerAssistedRequest. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Digitally active customers customers, customers | 2,000,000 | 2,000,000 | The reference bank. |
| Payment requests per customer per year that today reach the phone or a branch assistedRequestRate, requests per customer per year | 0.05 | 0.15 | Editorial assumption, replace with your own. Take the figure from your contact center and branch logs for payment related contacts. |
| Share of those requests the assistant completes assistantShare, fraction of requests | 0.2 | 0.5 | Editorial assumption, replace with your own. No deployment on this page has published a completion rate, so no benchmark is available. |
| Cost of a payment request handled by a person costPerAssistedRequest, USD per request | 4 | 8 | Editorial assumption, replace with your own fully loaded cost per phone or branch contact. |
What it leaves out: Counts only payment requests that today go to people. It leaves out the time customers save on payments they already make in the app, which is where Nubank reports a reduction in transaction time of up to 60%, and it leaves out build and run cost, misdirected payment losses and the cost of disputes. Customers who switch from the app screens add no saving.
Who already uses it?
8 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Bank ABB
Azerbaijan · Banking · 2026
Bank ABB in Azerbaijan opened its conversational assistants AI-khan and AI-nur to all customers in March 2026 after a beta stage. Customers can tell the assistant to move money between their own cards and accounts or to send money to another person by saying the contact's first and last name, and the assistant completes the transfer to the mobile number saved under that name. The assistants can also block or reactivate a card and answer product questions. The bank says more banking operations will follow. No usage or outcome figures are disclosed.
No outcome disclosed.
Boost
Malaysia · Banking · 2026
Boost, the Malaysian fintech and digital bank, announced in July 2026 an upgrade of Boba AI, its agentic assistant in the Boost app and the Boost Bank app. Customers can pay bills in the same conversation, after the assistant retrieves the real time outstanding balance, and can complete onboarding with eKYC and open a Boost Bank account. A separate voice service, Boba Voice, answers the Boost Bank hotline. Boost reports more than 30,000 customers and close to 100,000 requests in Boba AI since January 2026. The figures are Boost's own, and the release gives no split between bill payments and other requests.
- Users served: at least 30,000, since January 2026, all Boba AI use in the Boost app
"Since January 2026, more than 30,000 customers have interacted with Boba AI across the Boost app, generating close to 100,000 requests as customers increasingly use AI to manage everyday financial needs."
Claimed by: organization
SBI Sumishin Net Bank
Japan · Banking · 2026
In February 2026 SBI Sumishin Net Bank began a phased beta of its NEOBANK ai agent inside its mobile banking app. A customer can type or say a request such as sending 10,000 yen to a named person, name several recipients in one message, or upload a photo or screenshot of an invoice, and the agent guides the transfer and generates the matching screens. The bank says its own research as of 14 February 2026 found this to be the first AI agent offered by a Japanese bank that can execute banking transactions such as fund transfers directly in a mobile app. It is a beta, and no usage or outcome figures are disclosed.
No outcome disclosed.
Starling Bank
United Kingdom · Banking · 2026
Starling Assistant is an agentic AI assistant in the Starling app that responds to text and voice, analyzes spending patterns, creates savings Spaces and sets up transfers on the customer's behalf. In August 2026 Starling added "smart tools" to it and said new ones would follow every week for the rest of 2026 and at least monthly after that.
The launch set includes a tax saver that sweeps a share of the transactions a small business picks into a Space, a Making Tax Digital guide, a spending quiz and a student budget planner. A rainy day saver, which works out with the customer how much they can realistically save and sets up transfers into a savings Space, was announced as a follow up tool. The assistant is built on Google's Gemini models. No outcome figures are disclosed.
No outcome disclosed.
Nubank
Brazil · Banking · 2024
In December 2024 Nubank expanded Pix with generative AI in its app and on WhatsApp. A customer sends a voice message, a short text or an image that contains the recipient's name and the amount, and the AI builds and completes the Pix transfer. On WhatsApp the transfer runs entirely in the chat after a one time activation in the app, with a limit of R$200 per transaction and R$2,000 a day during the test phase. Nubank reports that two million customers had tried the WhatsApp version since October 2024 and that the AI version in the app cuts transaction time by up to 60%. The same fraud analysis that protects the app applies on WhatsApp.
- Cycle time reduction: up to 60%
"Pix with AI in the Nubank app uses technology to speed up processing, reducing transaction time by up to 60%, without compromising the security ensured by the app's defense mechanisms."
Claimed by: organization - Users served: 2 million, since October 2024 (WhatsApp test phase)
"On WhatsApp, the feature was tested by 2 million customers since October."
Claimed by: organization
Wells Fargo
United States · Banking · 2023
Wells Fargo's Fargo virtual assistant, built into the Wells Fargo mobile app since 2023, helps customers send money with Zelle, pay bills, locate routing numbers and gain insights into their spending and account balances. The bank reported that Fargo has now supported customers through more than 1 billion interactions, reached in under three years since launch, and that more than 3 million Spanish speaking customers have used it, engaging with the assistant over 160 million times.
- Interactions handled: at least 1 billion, since Fargo's 2023 launch, reached in under three years
"Fargo has now supported customers through more than 1 billion interactions – achieved in less than three years since its launch."
Claimed by: organization
Itaú Unibanco
Brazil · Banking · 2026
Itaú Unibanco began releasing ia.i in July 2026, a floating assistant on the home screen of its app that accepts text, voice and images. It answers questions about the customer's finances, explains products and executes transactions, including a Pix assembled from a voice command that closes only after the customer gives an explicit confirmation. The release started with 300,000 customers, with the bank planning 1 million by September and the whole customer base by the end of 2026. No usage or outcome figures are disclosed.
No outcome disclosed.
Warba Bank
Kuwait · Banking · 2026
Warba Bank in Kuwait launched Bdr AI in July 2026, an agentic assistant inside its mobile banking app. Retail customers give voice or text instructions in English or Arabic, including conversational Kuwaiti Arabic, to make instant transfers through WAMD, move money between their own Warba accounts, pay saved beneficiaries and generate PayMe links. Every transaction is shown on screen and needs the customer's explicit confirmation before it runs, and the assistant is available only after secure login with biometric protection. The bank describes it as the first agentic AI system to execute banking transactions directly by voice. No usage or outcome figure is disclosed.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Saved contacts, earlier payees and beneficiary data per customer, with names in the forms customers actually say
- Per customer limits, daily totals and the authentication level each payment type needs
- Transaction history and fraud signals for risk checks on the new channel
- Labeled utterances for transfers and bill payments in each language, including local names and amounts spoken in words
Systems to integrate
- Payment initiation API of the core banking or instant payment platform
- Strong customer authentication (biometrics, passkeys, PIN or app confirmation)
- Fraud engine and payee name check or confirmation of payee service
- Messaging channel such as WhatsApp Business, or the bank's mobile app with voice input
- Speech recognition and speech synthesis for voice, and image reading for invoices or screenshots
- Notification service for receipts and payment alerts
Complexity: Medium
The work sits in the bank's payment API, the payee data and the authentication step, which already exist for the app; the assistant adds speech or text understanding, a payee resolver and a confirmation screen on top. Voice and WhatsApp add their own identity and limit questions, as Nubank's separate caps for WhatsApp show.
- 1
Start with payments the customer already makes
Begin with transfers to saved contacts and own accounts, plus bill payments to billers the customer has paid before. A short allow list of payment types keeps the first release reviewable and gives a clean baseline for error and completion rates.
- 2
Resolve the payee with a deterministic resolver
Match names against contacts and payee history with a deterministic resolver that returns candidates with a confidence value. Below the threshold, ask the customer to pick. Never let the model supply an account number, phone number or payee identifier of its own.
- 3
Make the confirmation screen the only way a payment starts
Show payee, amount, currency and reference as structured fields, and require the customer's explicit approval with the authentication level the amount calls for. A voice command counts as an instruction to prepare a payment, and the approval is a separate step.
- 4
Reuse the bank's limits and fraud checks on the new channel
Send every assistant payment through the same payment API, limits and fraud scoring as the app, and set tighter caps for the first weeks on messaging and voice. Raise them as the error and fraud data comes in.
- 5
Run a test set of awkward instructions before each release
Include similar names, amounts spoken in words, mixed languages, instructions that change mid sentence and requests to exceed a limit, and replay them on every change to the model or the resolver.
Guardrails
- Payee identifiers come only from the customer's contacts, payee history or the bank's directory, never from model output
- Explicit customer confirmation of payee, amount and currency before every payment, with strong authentication at the level the amount requires
- The same limits, authentication and fraud scoring as the app, with lower caps on new channels in the first weeks
- Idempotency keys and velocity limits so a retry or a repeated voice command cannot pay twice
- Prompt injection and PII screening on text, images and documents the customer sends, and no card numbers or full account numbers in logs
KPIs to instrument
- Share of instructions that end in a confirmed payment, and the share abandoned at the confirmation screen
- Misdirected payment rate (payee corrected or payment reported as wrong) compared with payments made by hand
- Payee match rate on first attempt, split by language and channel
- Time from instruction to confirmed payment, compared with the app flow
- Fraud and scam losses and customer reported scams on assistant payments compared with the app
- Handover rate and customer satisfaction on payment conversations
Human in the loop
In the recommended design the customer approves every payment. Payments and fraud teams own the allow list of payment types, the caps per channel and the confidence threshold for payee matching, review a weekly sample of conversations with corrections, and approve each new payment type before the assistant can prepare it. Anything the assistant cannot resolve goes to the standard app screens or a person.
Common failure modes
- The wrong person is paid
- A similar name, a nickname or a speech recognition error leads to the wrong payee. Resolve against the customer's own contacts, show the full payee name on the confirmation screen and track corrections by language.
- A voice command is treated as approval
- Speech can be replayed, overheard or misheard. Keep approval as a separate authenticated step, so that the instruction alone cannot execute a payment.
- Scam victims pay faster
- A fluent, quick channel removes the pauses that let a customer reconsider a coached payment. Apply the bank's scam checks and warnings to assistant payments, and add friction on new payees and unusual amounts.
- Limits differ by channel and are exploited
- Caps on messaging or voice that are looser than the app invite abuse, and caps that are too tight make the assistant useless. Set them per channel from the first weeks of fraud and error data.
- Language coverage lags the launch language
- A resolver tuned for one language mismatches names and amounts in others. Test each language and dialect separately before it goes live and keep a fallback to the app screen.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
A customer facing assistant must make clear that people are dealing with AI, unless that is obvious from the context (Article 50). Preparing a payment the customer approves is not listed in Annex III. It becomes high risk if the same assistant evaluates creditworthiness, for example by deciding on an overdraft or credit line at the moment of payment (Annex III point 5(b)).
Rules that apply
Guidance
- Commission Delegated Regulation (EU) 2018/389 on strong customer authentication and common and secure open standards of communication (European Union, Europe). Sets the strong customer authentication requirements, including dynamic linking of the authentication to the amount and the payee, which a confirmation screen has to satisfy whether the instruction came by voice, text or image.
- APP scams (Payment Systems Regulator, Europe). UK reimbursement requirement for APP scam victims paying by Faster Payments or CHAPS, which applies to payments a customer approves after an assistant prepared them.
- Guidelines on Shared Responsibility Framework (Monetary Authority of Singapore, Asia Pacific). Covers phishing scams, where it assigns duties to financial institutions and telcos and requires payouts to victims where those duties are breached. A payment the customer authorized may fall outside it, so it applies to this use case only in part.
- Regulation (EU) 2024/1689 (AI Act), Article 50: transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system unless this is obvious from the context.
Controls to put in place
- AI disclosure at the start of the conversation and on voice calls
- A record that links each instruction, the resolved payee, the confirmation and the payment
- Caps, payment type allow list and payee confidence threshold approved by the payments and fraud risk owners
- Monitoring of misdirected payments, corrections and scam reports on assistant payments against the app baseline
- Regression tests for payee resolution, limit enforcement and injection resistance before every release
Frequently asked questions
- How is this different from a banking chatbot that explains how to transfer money?
A chatbot of the older kind tells the customer which screen to open. A conversational transfer assistant takes the instruction, finds the payee, prepares the payment and sends it after the customer confirms. Bank ABB's release describes assistants that carry out banking operations through voice commands, and Nubank's customers send a Pix by audio, text or image.
- Does the customer still confirm each payment?
The public records differ. Itaú Unibanco's assistant builds a Pix from a voice command and closes it only after an explicit confirmation. Bank ABB says that after the customer names the contact, the AI immediately completes the transaction. Nubank, Boost and SBI Sumishin Net Bank do not describe a confirmation step in their releases. We recommend treating the voice or text instruction as a request to prepare a payment, with the approval as a separate authenticated step.
- Who is running this live?
Nubank offered Pix by AI in its app and on WhatsApp from December 2024 and reports that two million customers tried the WhatsApp version. Bank ABB opened its voice assistants to all customers in March 2026, Boost customers pay bills in the Boba AI chat, SBI Sumishin Net Bank (now DOCOMO SMTB Net Bank) began a beta of its NEOBANK ai agent in February 2026, and Itaú Unibanco began a staged rollout of ia.i in July 2026. None of them has published a misdirected payment rate or a completion rate.
- Does a faster payment channel increase scam losses?
It can, because a fluent channel removes pauses in which a customer might rethink a coached payment. Nubank states that the risk analyses of its app also apply on WhatsApp. Apply your scam checks, new payee friction and warnings to assistant payments, and compare scam reports against the app.
How to cite this page
Blits.ai AI Use Case Library, "AI assistant for money transfers and bill payments by chat or voice", last verified 14 October 2026, https://www.blits.ai/ai-use-cases/conversational-money-transfer-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 14 October 2026: First published