A "real-time bank feed" is not a product US banks sell. What arrives is a prior-day BAI2 file at SFTP before dawn, intraday balance reporting on a fixed cadence, or a token-brokered pull from an aggregator against a snapshot the bank refreshed minutes or hours ago.
Plaid, direct bank APIs and BAI2 over SFTP each solve a different problem at a different price. Picking the wrong channel for AP posting, cash application or liquidity forecasting shows up later as reversals, duplicate entries and an unreconciled balance a controller has to explain to audit.
What Each Feed Delivers
BAI2 is the oldest of the three. The original BAI format was released in 1980 and replaced by BAI2 in 1987, with copyright transferred to the Accredited Standards Committee X9 in 2008; BTRS, the X9.121 successor, is where new work is specified. In practice, US cash-management banks still cut flat BAI2 files after end-of-day posting and push them to a corporate SFTP endpoint overnight, with a second intraday file delivered on a fixed cadence during the business day.
Direct bank APIs are vendor-specific REST surfaces maintained by the top cash management providers: JPMorgan Access, CitiConnect, BofA CashPro, Wells Fargo CEO, HSBCnet. Authentication is mutual TLS or OAuth client credentials, and payloads are increasingly ISO 20022 XML on the payment side with JSON reporting on the statement side. The Federal Reserve migrated Fedwire to ISO 20022 in a single-day cutover on July 14, 2025, and SWIFT ended the MT coexistence period for cross-border payments and cash reporting on November 22, 2025, so any new direct integration is being built against MX structures by default.
Plaid sits on top of the first two. Its Auth and Transactions endpoints abstract away the per-bank wiring and return a normalized JSON object regardless of which institution the user linked. Coverage is deep on US consumer and small-business accounts and thin on enterprise cash-management rails, which matters more than it first appears.
| Metric | Dec 2025 total | YoY growth (%) |
|---|---|---|
| Active users (millions) | 16.5 | 36 |
| Payment transactions (millions) | 351 | 57 |
| API calls (billions) | 24 | 27 |
Latency and Posting Cadence
"Intraday" is the word that misleads here. A BAI2 intraday file is typically cut four to six times a business day, which means the balance an agent sees can be up to two hours stale even on the most generous schedule, and the prior-day file arrives in a window between 3 and 6 a.m. local time. Direct bank APIs can be interrogated on demand and return the most recent posted transaction within minutes, which is the right latency for intraday sweep decisions or same-day payment approvals.
Plaid refreshes on its own cadence per item, and the Balance endpoint can force a live pull against the underlying bank. The catch is that a forced refresh is slow and metered, so running one before every AP posting is not a workable pattern, and the cached snapshot you get by default may lag the bank's own books by material time.
Data Fidelity and Which Fields Survive
BAI2 type codes are rich for cash reporting. Summary, detail and continuation records cover opening and closing ledger, available balances, holds, and float. They are narrow for payment metadata. Remittance information, originator-to-beneficiary references and ultimate debtor or creditor party data were never first-class in BAI2, which is the point of moving to ISO 20022 for anything AP posts or AR applies.
Direct bank APIs vary by vendor, but the ones that deliver camt.053 end-of-day statements and camt.054 debit/credit notifications carry the structured remittance that an agent needs to match an invoice to a payment without a human reading an email. The orchestration layer in an AI-native ERP is where that matters: the AP agent and the treasury agent can bind to different feeds without the ledger ever seeing the difference. Plaid's normalized schema is a lossy translation on purpose: fields the aggregator cannot populate consistently across banks are left off rather than guessed at, which is honest but limits what a cash-application agent can do with the stream.

Cost Structure and Operational Overhead
BAI2 over SFTP is cheap to run and expensive to build well. Most US cash-management banks either include the file in a treasury package or charge $50 to $500 per account per month, and the integration itself is a parser plus a keypair. Direct bank APIs cost more in commercial terms and require a per-bank build, per-bank certification and per-bank ongoing maintenance. That is the kind of work the same team learns to hate after the third bank.
Plaid is priced per linked item and per product call, and the all-in cost scales with consumer or merchant counterparty count rather than with the enterprise's own bank footprint. A treasury team with four banks and 200,000 customers using Plaid to link payouts pays very different money than a controller with 40 operating accounts across the same banks. The integration trade-offs between cost, quality and speed show up here: no single feed is cheapest across all three.
Auditability and Controls for an Agent
An AI agent posting to the ledger needs a record of what it saw and when, and the three feeds leave different evidence behind. A BAI2 file is an immutable artifact with a filename, a timestamp and a hash. It is the easiest of the three to retain and replay through a reconciliation pipeline. Direct bank API responses can be captured as signed JSON at the moment of the call, which is also auditable but puts the burden on the integration layer to persist the exact payload rather than the parsed fields.
Plaid is where audit complexity rises. A refreshed snapshot that does not match a later bank statement is not a Plaid defect; it is a property of the normalized abstraction, and it is a problem for an agent that has to prove why it posted. Regulatory drift is in the same direction: the UK's HMRC confirmed in May 2026 that screen scraping and browser automation for Government Gateway data access breach its terms, and US open banking is moving the same way under Section 1033. Agents built on screen-scraped feeds should be architected on the assumption that path closes. That is the kind of control secure agent access patterns are built against, and why the deterministic ledger sits behind an API the agent never writes to directly.
Matching Feed to Agent Workflow
An AP agent that extracts, codes, approves and posts a bill needs structured remittance on the way back in, which points to a direct bank API returning camt.054 for the matching payment and to ISO 20022 on any wire or ACH leg. A cash-application agent working consumer receipts can run on Plaid where the counterparty population is the public, and on direct APIs where it is enterprise. A treasury agent producing an intraday cash position across 20 operating accounts is better served by BAI2 intraday for scheduled refreshes and direct API calls on demand at cut-off.
The alternative pattern, wiring every agent to the same aggregator for simplicity, is how teams discover six months in that the aggregator was never the right channel for the AP workflow. Migrations from NetSuite tend to inherit a Plaid-first posture by default, which is one of the things an agentic ERP platform rebuilds on the way through. The right answer is usually a mixed stack: BAI2 for the accounts where it is free and sufficient, direct APIs where remittance fidelity or intraday latency earns the build cost, and an aggregator only where the counterparty breadth cannot be covered any other way.
The Feed Is the Floor the Agent Stands On
Ask any ERP or automation vendor claiming an AI agent can own AP posting or daily cash close which bank feed the agent reads from, and whether the feed carries the structured remittance fields the agent needs to match without a human in the loop. A good answer names the specific endpoint, camt.053, camt.054 or a bank's REST statement API, states the posting latency in minutes or hours rather than "real time," and shows the audit record the agent persists against every posting. A weak answer treats Plaid as if the aggregator layer solves the fidelity problem, or BAI2 as if a prior-day file can carry remittance it was never designed to carry. Model quality does not compensate for a channel that does not carry what the workflow needs.
