Pay by Bank and Account-to-Account (A2A) Payments Architecture: Open Banking PIS, FedNow, RTP and Pix Compared for Merchants (2026)

Pay by Bank and Account-to-Account (A2A) Payments Architecture: Open Banking PIS, FedNow, RTP and Pix Compared for Merchants (2026)

Pay by Bank and Account-to-Account (A2A) Payments Architecture: Open Banking PIS, FedNow, RTP and Pix Compared for Merchants (2026)

Brazil’s Pix processed 319.7 million transactions on a single Friday this November, and UK open banking passed one billion payments in July. Yet most merchant engineering teams still bolt pay by bank onto checkout as if it were a cheaper card. It is not. A card payment is a pull: you request authorisation, capture later, and the network gives the shopper a chargeback. Account-to-account payments are pushes. The payer’s bank sends money that is final within seconds, refunds are brand-new payments in the other direction, and fraud liability depends on which country’s rules apply. In 2026, FedNow and RTP both carry $10 million limits, the EU has a verification-of-payee mandate and a political deal on PSD3, and the UK has a commercial VRP scheme. The rails are ready. The hard part is the merchant-side architecture.

What this covers: a reference architecture for multi-rail pay by bank, the UK open banking payment initiation flow at API level, a payment state machine with webhook code, FedNow vs RTP vs Pix vs UPI on limits and messages, refunds and fraud liability, reconciliation, an illustrative fee comparison with cards, and the failure modes that cause real incidents.

This article is systems-architecture analysis only. It is not financial, legal or investment advice.

Context and Background

“Pay by bank” is a marketing label, not a protocol. Underneath it sit two quite different things. The first is an initiation layer: an API that lets a third party ask a shopper’s bank to send a payment on the shopper’s behalf. In the UK and EU this is the payment initiation service (PIS) defined by PSD2, Directive (EU) 2015/2366, and implemented in the UK through the Open Banking Standard that the CMA order made mandatory for the nine largest banks from 2018. The second is a settlement rail: the interbank system that actually moves the money, such as UK Faster Payments, SEPA Instant Credit Transfer, FedNow, The Clearing House RTP network, Brazil’s Pix or India’s UPI.

Some markets bundle both layers into one scheme. Pix and UPI define the initiation experience (QR codes, aliases, payment requests) and the rail together, run by the central bank or a national operator. The UK and EU split them. A licensed payment initiation service provider (PISP) calls the bank’s API, and the bank then submits an ordinary instant credit transfer. The US sits in between. FedNow and RTP are rails with a built-in request-for-payment message, but no regulated open banking API mandate for third-party payment initiation equivalent to PSD2.

That structural difference drives almost every design choice in an a2a payments architecture. It decides who owns the consent screen, which status codes you see, how long the “unknown” window lasts, and who pays when a scam succeeds. Our earlier deep dive on variable recurring payments and the payment initiation API in the UK and EU covers the recurring-mandate side. This post focuses on one-off checkout payments and the merchant platform that has to support several rails at once.

Scale is no longer the question. Open Banking Limited reported 40.16 million open banking payments in June 2026, of which 32.43 million were single domestic payments and 7.73 million were sweeping VRPs, alongside the one-billion cumulative milestone. India’s NPCI processed 25.47 billion UPI transactions in October 2026 according to press reports of NPCI’s monthly data. FedNow reported more than 1,800 enrolled banks and credit unions at its third anniversary in July 2026. The open question is whether your checkout, ledger and support processes are built for push payments.

Pay by Bank Reference Architecture for Merchants

Pay by bank is a checkout method in which the shopper authorises a credit transfer from their own bank account to the merchant’s account, usually via a bank app redirect, QR code or request-for-payment message. Settlement is instant and final. The merchant must therefore manage consent, status uncertainty, refunds and reconciliation itself.

Pay by bank reference architecture routing account to account payments through PISP, FedNow or RTP and Pix into a reconciliation ledger
Figure 1: A multi-rail pay by bank architecture. One orchestrator hides rail differences behind a single payment intent, and the ledger is fed by status events, not by the checkout response.

Figure 1 shows the components most merchants converge on. The checkout creates a payment intent. An A2A orchestrator chooses the rail and provider based on the shopper’s country, currency and amount. For UK and EU shoppers it calls a PISP. For US shoppers it uses a bank or processor that can send a FedNow or RTP request for payment. For Brazilian shoppers it asks a Pix PSP for a dynamic QR code. In every case the payer’s bank performs strong customer authentication and pushes funds over an instant rail into the merchant’s collection account. Status events flow back through webhooks into a reconciliation ledger, which is the only component allowed to say “paid”.

Separate the intent from the instruction

In pay by bank systems, the single most useful abstraction is a rail-neutral payment intent with its own identifier, amount, currency, merchant reference and expiry. Each attempt to collect it creates a rail-specific instruction: a UK consent ID, a FedNow request-for-payment message ID, or a Pix transaction ID (txid) bound to a QR code. One intent can have several instructions. A shopper may abandon a bank redirect, scan a QR code instead, or retry with a different bank.

This matters because A2A payments have no authorisation hold. With cards, an authorisation reserves funds and a failed capture simply releases them. With pay by bank, two successful instructions mean two irrevocable transfers and a refund to send. The orchestrator must therefore enforce “at most one instruction can reach a submitted state per intent”, using a database constraint rather than application logic. The patterns in our guide to idempotency keys and double-charge prevention apply directly. UK PISPs must also send an x-idempotency-key header when creating a consent, so idempotency exists at two layers.

Treat the checkout response as a hint, not a fact

In card flows, the synchronous authorisation response is close to authoritative. In account to account payments the synchronous response usually says only “received”. In the UK Open Banking Read/Write API v4.0, a newly created domestic payment can only return RCVD or RJCT. Settlement information arrives later as ISO 20022 status codes such as ACSP (accepted, settlement in process), ACSC (settlement completed on the debtor side) or ACCC (settlement completed on the creditor side).

The redirect back to your site after bank approval is weaker still. It proves that the shopper finished a journey in the bank app. It does not prove that money moved. Shoppers close tabs, mobile operating systems kill app switches, and redirects can be replayed. A robust design ships goods on a ledger event derived from a confirmed status, or from a credit seen on the collection account, never on the redirect alone.

Put the ledger in charge of truth

Because push payments arrive as credits to an account, the strongest evidence of payment is the credit itself. Mature merchants run a reconciliation service that consumes three streams. The first is provider webhooks. The second is account statements or real-time credit notifications from the collecting bank, such as ISO 20022 camt.054 debit/credit notifications or camt.053 statements where offered. The third is the internal order system. The ledger links each credit to an intent through the end-to-end reference carried in the payment message.

That reference is a design decision you make once. In ISO 20022 credit transfers (pacs.008), the end-to-end identification is limited to 35 characters and must survive every hop. In UK Faster Payments, the legacy payment reference field that a payer sees is only 18 characters. Pix carries the txid you set when you create a dynamic QR charge. Choose a compact, unique, rail-safe reference format early. Retrofitting reconciliation keys after launch is one of the most expensive mistakes in A2A projects. Our post on ISO 8583 vs ISO 20022 message formats explains why ISO 20022’s structured remittance fields make this easier than card messaging ever did.

Route by rail capability, not by price alone

The orchestrator is a routing engine. The logic resembles the smart routing described in our payment orchestration architecture guide, with different inputs. Instead of BIN ranges and acquirer approval rates, the router looks at the shopper’s bank, the amount relative to the rail limit and the bank’s own limit, historical conversion by bank, and whether the provider supports refunds through the same API. A £4,000 order may exceed what a given bank lets customers send to a new payee over Faster Payments, even though the scheme limit is £1 million. Routing it to pay by bank and failing at the bank’s limit is worse than offering cards up front.

Walk-through: Rails, Status Flows and the State Machine

This section walks through one rail end to end, then generalises the status model and compares the main rails. The UK is the best-documented case, so it comes first.

UK open banking payment initiation, step by step

Open banking payment initiation sequence for pay by bank showing consent creation, SCA in the bank app, payment submission and status updates
Figure 2: The UK open banking payment initiation flow for a single domestic payment, from consent to settlement status.

Figure 2 follows a single domestic payment under the Open Banking Read/Write API. The PISP first calls POST /domestic-payment-consents with a client-credentials token, a signed payload and an idempotency key. The bank (the account servicing payment service provider, or ASPSP) returns a ConsentId with status AWAU, awaiting authorisation. The PISP then sends the shopper to the bank by redirect or app-to-app link. The shopper authenticates with strong customer authentication and approves the exact amount and payee.

The bank redirects back with an authorisation code. The PISP exchanges it for an access token and calls POST /domestic-payments, referencing the consent. The bank responds with RCVD or RJCT and submits the credit transfer to Faster Payments. Later the PISP polls the payment, or receives an event notification, and sees a status such as ACSC. The v4.0 specification also defines an optional payment-details endpoint that can return a richer history including scheme-level statuses. Once a payment is created, the consent moves to COND, meaning the consented action is complete.

Version 4.0 of the standard matters for two reasons. It mandated FAPI 1.0 Advanced Final, replacing the older draft profile, and it aligned payment messaging with ISO 20022, including fields needed for Bank of England CHAPS requirements from May 2025. OBL also flagged “timely, definitive payment statuses” as an optional improvement that the FCA may later make mandatory. In practice, status quality still varies by bank, which is why the state machine below has an explicit unknown state. The security layer is covered in depth in our FAPI 2.0 security profile explainer.

How the US rails differ: FedNow vs RTP

The FedNow vs RTP comparison is mostly about reach and operator, not about mechanics. Both are 24x7x365 credit-push systems using ISO 20022 messages. Both settle each payment individually and irrevocably. Both support a request for payment. The Clearing House launched RTP in November 2017. The Federal Reserve launched FedNow on 20 July 2023.

Limits converged in 2025. TCH raised the RTP per-transaction limit from $1 million to $10 million in February 2025. FedNow raised its customer credit transfer limit from $1 million to $10 million on 12 November 2025, and the Federal Reserve’s notice states that participants may use the maximum or set lower limits. FedNow had launched at $500,000. Those are network ceilings. Your shopper’s bank will usually apply much lower consumer limits.

Volumes differ by an order of magnitude, which matters for pay by bank reach. Digital Transactions, citing operator figures, reported 128 million RTP transactions worth $480 billion in Q1 2026, against 2.7 million FedNow transactions worth $271.3 billion. Average FedNow value is therefore far higher, which suggests heavy use for account funding and business payments rather than retail checkout. RTP reports more than 1,200 participating institutions. FedNow reports more than 1,800 enrolled, though enrolment for receiving is more common than sending.

For checkout, the relevant US mechanism is request for payment. The merchant’s bank sends a creditor payment activation request, ISO 20022 pain.013, to the shopper’s bank. The shopper approves it in their banking app, and the shopper’s bank sends a pacs.008 credit transfer. If the shopper declines, the response is a pain.014. Oracle’s RTP implementation guide documents exactly that pairing. Adoption of request-for-payment for retail checkout remains limited in the US as of late 2026. Most US “pay by bank” checkout products still rely on ACH debits with account-linking, which are pulls with return windows. Our ACH and Nacha architecture guide covers that model.

Pix and UPI: the scheme-integrated model

Pix, launched by Banco Central do Brasil in November 2020, is the clearest example of a rail designed for checkout. Merchants generate a dynamic QR code tied to a charge with a txid, amount and expiry. The shopper scans it in any bank app and approves. The merchant’s PSP receives a settlement notification within seconds. On 6 November 2026, Pix set a record of 319.7 million transactions worth R$180.5 billion in a single day, according to Central Bank figures reported by Agência O Globo. The previous record, on 4 September 2026, was just over 318 million.

Pix also standardised recurring payments. Pix Automático went live across most institutions on 16 June 2025. The payer authorises a mandate once in their bank app, setting frequency, fixed or variable amount and a maximum per transaction, and can cancel at any time. That is functionally close to a UK VRP but delivered as a central-bank scheme rule.

India’s UPI uses virtual payment addresses and QR codes, with NPCI as operator. Press reports of NPCI’s October 2026 data show 25.47 billion transactions worth ₹31.53 lakh crore, about 822 million per day. From 15 September 2025, NPCI raised person-to-merchant limits for verified merchants in categories such as insurance, capital markets and travel to ₹5 lakh per transaction and ₹10 lakh per day. The person-to-person limit stayed at ₹1 lakh per day. For most merchants outside India, UPI is reached through a local acquirer or PSP rather than integrated directly.

Rail comparison matrix

Dimension UK FPS via PIS EU SCT Inst via PIS FedNow RTP Pix UPI
Operator Pay.UK EPC scheme, CSMs Federal Reserve The Clearing House Banco Central do Brasil NPCI
Initiation for checkout PISP API, bank redirect PISP API, bank redirect Request for payment via bank Request for payment via bank Dynamic QR, Pix key QR, VPA, intent
Scheme ceiling £1m since 2022 No scheme cap mandated for SCT Inst by IPR $10m since Nov 2025 $10m since Feb 2025 Set by PSP and BCB limit rules ₹1 lakh P2P, up to ₹5 lakh some P2M
Message standard ISO 20022 aligned in OB v4.0 ISO 20022 ISO 20022 ISO 20022 Pix spec, ISO 20022 based NPCI UPI spec
Refund path New push, OB refund account data New push New push or return request New push or return request Pix devolução linked to original UPI refund API
Scam reimbursement regime PSR mandatory, up to £85,000 PSR text agreed, impersonation refunds No federal mandate No federal mandate MED 2.0 trace and return Bank and NPCI dispute process

Treat the matrix as a starting point for pay by bank rail selection. The “scheme ceiling” cell for the EU deserves a caveat. The Instant Payments Regulation lets payment service users set their own instant-payment limits, and the SCT Inst scheme historically carried a €100,000 maximum. I could not verify the current EPC rulebook figure for this post (unverified), so check the current SCT Inst rulebook before relying on any number.

A rail-neutral payment state machine

Pay by bank payment state machine from created through authorised, submitted, unknown, settled, credited and refunded
Figure 3: A rail-neutral state machine. The UNKNOWN state is mandatory, because every rail has a window where the merchant cannot yet tell whether money moved.

Figure 3 shows the states every rail mapping should land in. CREATED and AWAITING_AUTH cover consent creation and the bank redirect or QR display. ABANDONED is a timeout, not a failure. SUBMITTED means the bank or rail has accepted an instruction. SETTLED_DEBTOR corresponds to ACSC in UK v4.0 or a FedNow pacs.002 accepted status. CREDITED means the money is on the merchant’s account and matched in the ledger. Refund states hang off CREDITED because a refund is a separate outbound payment.

The mapping tables are where bugs live. UK ACSP is not settlement; it means “accepted, settlement in process”. UK PDNG means pending further checks and can last minutes. ACWP, accepted without posting, is rare but appears in the code list. A Pix charge can expire while the shopper is still in the app. A US request for payment can be ignored rather than rejected, leaving it open until expiry. Each of these should land in a defined state with a timeout policy.

Below is a minimal, runnable Python sketch of the transition guard and webhook handler. It is illustrative, not provider-specific.

from enum import Enum
import hmac, hashlib, json

class S(str, Enum):
    CREATED="CREATED"; AWAITING_AUTH="AWAITING_AUTH"; AUTHORISED="AUTHORISED"
    SUBMITTED="SUBMITTED"; UNKNOWN="UNKNOWN"; SETTLED_DEBTOR="SETTLED_DEBTOR"
    CREDITED="CREDITED"; REJECTED="REJECTED"; ABANDONED="ABANDONED"
    REFUND_PENDING="REFUND_PENDING"; REFUNDED="REFUNDED"

ALLOWED = {
    S.CREATED: {S.AWAITING_AUTH, S.REJECTED},
    S.AWAITING_AUTH: {S.AUTHORISED, S.REJECTED, S.ABANDONED},
    S.AUTHORISED: {S.SUBMITTED, S.REJECTED},
    S.SUBMITTED: {S.SETTLED_DEBTOR, S.REJECTED, S.UNKNOWN},
    S.UNKNOWN: {S.SETTLED_DEBTOR, S.REJECTED},
    S.SETTLED_DEBTOR: {S.CREDITED},
    S.CREDITED: {S.REFUND_PENDING},
    S.REFUND_PENDING: {S.REFUNDED, S.CREDITED},
}

# ISO 20022 ExternalPaymentTransactionStatus1Code -> internal state
UK_V4_MAP = {
    "RCVD": S.SUBMITTED, "PDNG": S.SUBMITTED, "ACTC": S.SUBMITTED,
    "ACCP": S.SUBMITTED, "ACFC": S.SUBMITTED, "ACSP": S.SUBMITTED,
    "ACWC": S.SUBMITTED, "ACWP": S.SUBMITTED,
    "ACSC": S.SETTLED_DEBTOR, "ACCC": S.SETTLED_DEBTOR,
    "RJCT": S.REJECTED, "BLCK": S.REJECTED,
}

def verify(sig_header: str, body: bytes, secret: bytes) -> bool:
    expected = hmac.new(secret, body, hashlib.sha256).hexdigest()
    return hmac.compare_digest(expected, sig_header)

def apply_event(db, body: bytes, sig: str, secret: bytes):
    if not verify(sig, body, secret):
        return 401
    evt = json.loads(body)
    if db.seen_event(evt["event_id"]):          # dedupe redelivery
        return 200
    target = UK_V4_MAP.get(evt["status"])
    if target is None:
        db.log_unmapped(evt); return 200        # never crash on new codes
    with db.tx() as tx:
        p = tx.lock_payment(evt["payment_id"])  # SELECT ... FOR UPDATE
        if target == p.state:
            pass                                # duplicate status, no-op
        elif target in ALLOWED.get(p.state, set()):
            tx.set_state(p.id, target, evt["status"], evt["occurred_at"])
        else:
            tx.flag_out_of_order(p.id, evt)     # e.g. ACSC before RCVD
        tx.mark_seen(evt["event_id"])
    return 200

For any pay by bank integration, three details matter more than the code. Webhooks arrive out of order, so the handler records out-of-order events instead of rejecting them. Statuses the mapping does not know are logged, not thrown, because banks add codes. And CREDITED is never set by a provider webhook. Only the reconciliation job sets it, after matching a credit on the collection account. A separate sweeper moves payments stuck in SUBMITTED past a rail-specific timeout into UNKNOWN and issues a status query, for example a GET on the UK payment resource or a status request to the US bank.

Capacity and cost, worked through

Consider a merchant doing 200,000 pay by bank checkouts a month at an average basket of £80. If 3% of attempts hit the UNKNOWN path for more than five minutes, that is 6,000 orders a month, about 200 a day, needing an automated status query and possibly a support touch. At a fully loaded support cost of, say, £4 per contact (illustrative), unresolved unknowns alone could cost £24,000 a month if every one generated a ticket. The engineering lesson is that unknown-state automation pays for itself faster than any fee saving.

Now the fee side. The following figures are illustrative assumptions, not quotes. Suppose card acceptance costs a blended 1.5% plus £0.20 per transaction, and an A2A provider charges a flat £0.20 per payment. On an £80 basket, card cost is £1.40 against £0.20 for pay by bank, a saving of £1.20. Across 200,000 payments that is £240,000 a month before conversion effects. If pay by bank conversion is five percentage points lower than cards for a given segment and margin is 20%, each lost sale costs £16 of margin. Losing just 1.5% of 200,000 orders, 3,000 sales, would cost £48,000. The real decision hinges on conversion by bank and segment, not on the headline fee. For card-side cost structure, see our guide to interchange fees and least-cost routing.

Refunds, Disputes and Fraud Liability

Card schemes give shoppers chargeback rights and merchants a structured dispute process. Pay by bank, like other account to account payments, gives neither by default. That is the single biggest product difference, and it shapes support, legal terms and risk models.

Pay by bank dispute and refund paths compared across UK APP reimbursement, Pix MED 2.0 and US instant payment return requests
Figure 4: What happens when a pay by bank payment is disputed. Merchant-initiated refunds are new pushes. Fraud claims follow jurisdiction-specific rules that do not resemble chargebacks.

Refunds are outbound payments

A refund on a push rail is a new credit transfer from the merchant to the shopper. That needs the shopper’s account details, a funded outbound account and a payout capability. The UK Open Banking Standard lets a PISP request the payer’s account details in the payment response for refund purposes, provided the bank supports it. In Brazil, Pix defines a devolução (return) linked to the original transaction, so the refund carries a reference to the payment it reverses. In the US, a merchant can send a new FedNow or RTP credit, and both networks also support a return-request flow that the receiving bank may honour or decline.

Architecturally, store the refund as a child of the original payment with its own state machine and idempotency key. Partial refunds must sum to no more than the original credit. If the merchant’s payout account and collection account differ, the ledger must show the transfer between them.

APP fraud: the UK reimbursement regime

Authorised push payment (APP) fraud happens when a scammer tricks someone into sending money themselves. Since 7 October 2024, the UK Payment Systems Regulator has required PSPs to reimburse APP scam victims for Faster Payments and CHAPS payments. The maximum reimbursement is £85,000 per claim, which the PSR says covers over 99% of claims. Firms may apply an optional excess of up to £100, which cannot be applied to vulnerable consumers. Claims must be made within 13 months of the payment. Under the PSR’s policy, the cost is shared between the sending and receiving PSPs, generally 50:50.

For a legitimate merchant receiving pay by bank payments, the receiving-side liability matters. Your collecting bank now shares reimbursement costs if it banks a fraudster. Expect it to apply more scrutiny to merchant onboarding, payee-name matching and unusual inbound patterns. Our dedicated post on APP fraud detection and Confirmation of Payee covers the detection side.

EU: verification of payee and the PSD3/PSR deal

In the euro area, the Instant Payments Regulation required PSPs to offer verification of payee (VoP) from 9 October 2025. Payment institutions and e-money institutions in the euro area face the same VoP deadline, and non-euro-area PSPs follow on 9 July 2027. VoP applies to all credit transfers, not only instant ones. For pay by bank merchants, VoP means the legal name on your collection account must match what shoppers’ banks expect, or shoppers will see a mismatch warning at the worst possible moment.

The broader reform is PSD3 and the Payment Services Regulation (PSR). The European Parliament and Council reached a provisional political agreement on 27 November 2025. Final compromise texts followed in April 2026. As of the latest status this post could verify, a tracker last reviewed in August 2026, formal adoption and Official Journal publication were still pending and expected in the second half of 2026, with most rules applying about 21 months after entry into force. Verify the current legislative stage before relying on dates. Agreed content includes liability for PSPs whose anti-fraud mechanisms are inadequate, full refunds for impersonation fraud where the scammer poses as the PSP and the victim reports to police, fraud-data sharing between PSPs, a customer permissions dashboard, and stronger obligations on banks’ open banking interfaces.

Brazil: MED 2.0

Pix’s Special Return Mechanism (Mecanismo Especial de Devolução, MED) was upgraded to MED 2.0, effective 2 February 2026. It can now trace funds beyond the first receiving account into intermediary accounts, lets victims contest through a self-service button in their bank app, and allows suspicious funds to be blocked for up to 11 days during analysis. MED is limited to fraud, suspected fraud and operational errors by institutions. It does not cover a payer who simply typed the wrong key. A merchant whose account receives funds later traced as fraud proceeds can see them blocked and returned, so ledger designs need a “clawback” state as well as a refund state.

United States: no federal scam-reimbursement mandate

FedNow and RTP do not currently carry a federal APP-style reimbursement mandate comparable to the UK’s. Both networks let a sending bank request a return, and the receiving bank decides whether to honour it. Regulation E protections cover unauthorised electronic fund transfers for consumers, but how they apply to authorised scam payments on instant rails has been debated rather than settled. Merchants should treat US instant pay by bank receipts as final, with bank-to-bank return requests as an exception path.

Trade-offs, Gotchas, and What Goes Wrong

Pay by bank conversion depends on the bank, not the rail. Bank-app redirects fail in specific ways: app not installed, biometric prompt timeouts, desktop-to-mobile handoffs, or a bank that forces web login. Measure conversion per bank and per device. Route around the worst performers by offering cards first for those shoppers.

Bank limits are lower than scheme limits. A $10 million FedNow ceiling does not mean a consumer can send $10,000 to a new merchant. Banks set per-payment, per-day and new-payee limits, and they change them without notice. Your orchestrator should learn these from rejections and stop offering pay by bank above the observed threshold.

The unknown window is real. Bank APIs time out after submitting to the rail. Webhooks are delayed. Statement feeds lag. If your system treats “no answer” as “failed” and lets the shopper retry, you will collect twice. Lock the intent, query status, and only release it after a definitive rejection or expiry.

Reference truncation breaks reconciliation. A 36-character UUID will not fit the 35-character ISO 20022 end-to-end ID, let alone an 18-character Faster Payments reference. Design a short reference with a checksum and enforce it in the orchestrator.

Name mismatch warnings kill pay by bank conversion. With UK Confirmation of Payee and EU VoP, a shopper may see “name does not match” if your trading name differs from the account’s legal name. Align account naming with your bank before launch, and test with real consumer accounts.

Refunds need payout capability. Teams often discover at the first refund that their provider supports collection but not payouts, or that the refund account details were never stored. Make refundability a launch criterion.

Sweeps and settlement timing. Money landing in a collection account at 02:00 on a Sunday is real money, but your finance team may sweep weekly. Instant rails remove float, which changes treasury reporting. The large-value counterpart is covered in our post on Fedwire, CHIPS and RTGS settlement finality.

Anti-patterns to avoid. Shipping on redirect. Retrying failed payments without locking the intent. Mapping every non-final status to “failed”. Using one global timeout across rails. Treating pay by bank as a card in your risk engine, which then scores chargeback risk that does not exist and misses mule-account risk that does.

Practical Recommendations

Start with one market where the rail and initiation layer are mature and your customers already bank digitally. For UK merchants that is Faster Payments via a PISP. For Brazil it is Pix with dynamic QR codes. In the US, assess whether your customer base is reachable by request for payment before committing, because ACH-based account linking may still be the practical option there. Build the orchestrator, state machine and ledger as rail-neutral from day one, even with one rail live, because the second rail always arrives sooner than planned.

Invest in pay by bank observability before marketing. You need per-bank conversion funnels, a dashboard of payments in UNKNOWN by age, webhook lag percentiles, and reconciliation match rates. Treat an unmatched credit older than one business day as an incident, not a backlog item.

Finally, rewrite customer terms and support scripts. Shoppers used to chargebacks need to know how refunds work for account to account payments and what to do if they suspect a scam. Support must know which jurisdiction’s protections apply.

Checklist:

  • Rail-neutral payment intent with a database-enforced single submitted instruction.
  • Short, unique, checksummed reference that fits every rail’s field limits.
  • State machine with an explicit UNKNOWN state, per-rail timeouts and status queries.
  • Webhook handler that verifies signatures, dedupes, tolerates out-of-order and unmapped codes.
  • Ledger sets CREDITED only after matching a real account credit.
  • Refund and payout path tested before launch, including partial refunds.
  • Collection account legal name verified against CoP and VoP checks.
  • Per-bank conversion and limit learning feeding the router.

Frequently Asked Questions

What is pay by bank and how is it different from paying by card?

Pay by bank lets a shopper send money directly from their bank account to a merchant, usually by approving the payment in their banking app after a redirect or QR scan. A card payment is a pull: the merchant requests authorisation and the card network routes funds later, with chargeback rights for the shopper. Pay by bank is a push over an instant rail such as Faster Payments, FedNow, RTP or Pix. It settles in seconds, is final, and has no chargeback. Refunds are separate outbound payments sent by the merchant.

Is pay by bank safe for merchants and shoppers?

For merchants, pay by bank removes card-not-present chargeback fraud because the payer authenticates with their own bank using strong customer authentication. The main risk shifts to scams, where a shopper is tricked into paying a fraudster, and to merchants unknowingly receiving money from mule accounts. Shoppers’ protections depend on location. The UK requires APP scam reimbursement up to £85,000. Brazil’s MED 2.0 traces and returns fraudulent Pix funds. US instant rails have no equivalent federal mandate. Merchant refunds are never automatic.

What is the difference between FedNow and RTP?

FedNow is operated by the Federal Reserve and launched in July 2023. RTP is operated by The Clearing House, a bank-owned company, and launched in November 2017. Both are 24×7 instant credit-push rails using ISO 20022 messages, both support request for payment, and both now allow up to $10 million per transaction. RTP processes far more transactions, 128 million in Q1 2026 versus 2.7 million for FedNow. FedNow reports more enrolled institutions. Many banks connect to both, so merchants usually access them through a bank or processor.

Can a pay by bank payment be reversed or charged back?

There is no chargeback in account-to-account payments. Once an instant credit transfer settles it is final. A merchant can refund by sending a new payment back to the shopper. If the payment was fraudulent, the shopper claims from their bank under local rules, such as the UK’s APP reimbursement regime or Brazil’s MED 2.0. US banks can send return requests over FedNow and RTP, but the receiving bank decides whether to return funds. Ledgers should model refunds and clawbacks as separate linked transactions.

Do I need an open banking licence to offer pay by bank in the UK or EU?

To initiate payments on a shopper’s behalf you must be authorised as a payment initiation service provider or work with one. Most merchants use a licensed PISP or a payment provider that holds the permission, and integrate its API. The PISP calls the bank’s open banking payment initiation API, handles consent and security, and passes status events back. Merchants still own checkout design, reconciliation, refunds and customer communication. Collection accounts also need correct legal names to pass Confirmation of Payee or EU verification-of-payee checks.

How long does it take for a pay by bank payment to settle?

On instant rails, interbank settlement normally completes in seconds. Faster Payments, SEPA Instant, FedNow, RTP, Pix and UPI all target near-real-time crediting. The merchant-visible delay is usually longer, because it includes the shopper’s time in the bank app, the provider’s status update, webhook delivery and reconciliation against the collection account. Some payments enter pending states for extra fraud checks. Design pay by bank flows for most payments confirming within a minute and a small tail that needs an automated status query.

Further Reading

By Riju — about

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *