Skip to main content
Viithiisys
Back to blog
AI Engineering7 min readThu, Nov 19, 2026

RPA in Banking: Three Workflows Worth Automating

RPA in banking, applied to three workflows: KYC checks, AML alert packets and ACH returns. Where bots break, what stays with people, and the cost drivers.

Gaurav Saini

Founder, Viithiisys

RPA in Banking: Three Workflows Worth Automating

What robotic process automation in banking means

RPA in banking means software bots that use the same screens, files and inboxes as a clerk, following fixed rules. It suits steps that never change. It is a poor fit for anything that needs reading, judgement or a stable interface it does not have.

Banks have a lot of that first kind of work. Data is copied from a core system into a case tool. A document is checked against a list. A return code is logged in three places.

This post covers robotic process automation banking teams actually run into, through three workflows: onboarding document checks, anti-money-laundering (AML) alert packets and ACH payment returns. Each is written as a design, with the systems named as examples, not as a delivery record.

Provenance: Viithiisys sells AI workflow automation, so we have an interest in this topic. Nobody has handed us a confirmed banking delivery record for this post, so each workflow is a [Reference build], a proposed design, and the public sources are cited in place.

Workflow one: KYC document checks at onboarding

An analyst receives an application, opens the documents, and types the details into the core system. A second person screens the name. The jam is the gap between "documents received" and "documents usable".

Typically the applicant uploads to a portal such as nCino or a similar onboarding tool. An analyst opens each file and compares it with a checklist. They then re-key the fields into the core, for example Fiserv, FIS or Jack Henry. A sanctions screen runs afterwards.

It breaks on blurry scans, missing pages and names spelled three ways. Each one returns to the applicant by email, and the file sits idle until they reply. The cost is days of elapsed time on files where nobody is working.

This is a [Reference build], a design for KYC intake, not something we delivered for a bank.

What to automate and what stays with a person

Automate the re-keying, the checklist comparison and the "missing page" email. Use a model to read the documents into fields, with a confidence score on each. Anything below the threshold goes to an analyst with the field highlighted.

The decision to accept an identity stays with a person. That is our recommended default, not a regulatory rule. Your compliance team and counsel decide where the line sits in your jurisdiction.

Workflow two: AML alert packets for investigators

Most of an investigator's time on an alert goes to gathering, not deciding. The bot's job is to assemble the packet. The investigator's job is to judge it.

Today a monitoring system such as NICE Actimize or Verafin raises an alert. An investigator then opens the core system for the account history, the case tool for prior alerts, and the KYC file for the customer profile. They paste it all into a case document.

The failure is inconsistency. Two investigators build two different packets for similar alerts, and a reviewer cannot compare them. Time pressure also matters. FinCEN's guidance says a suspicious activity report is generally due within 30 calendar days of initial detection of facts that may constitute a basis for filing, with a longer window in some cases (FinCEN, suspicious activity report filing information).

Diagram of an alert packet being assembled from three systems before an investigator reviews it

Packet assembly and the filing decision

Automate the pull: account history, prior alerts, KYC profile, counterparty list, all in one fixed layout. Log what was pulled and when, so a reviewer can see the evidence.

Leave the narrative and the file-or-close decision with the investigator. A model can draft a summary for them to edit, but it should be labelled as a draft. This is a [Reference build] for packet assembly, a design only.

Workflow three: ACH returns and payment exceptions

Returned payments are rule-bound, which makes them a good RPA target. The trouble is the long tail of odd cases around the common ones.

An ACH return arrives with a reason code. An operations clerk looks up the original payment, finds the customer in the core system, reverses or reposts the entry, and notifies the customer or the originator. Many banks do some of this in a payments hub and the rest in spreadsheets.

It breaks when the return does not match any payment cleanly. Duplicates, partial matches and closed accounts get parked in a shared inbox, and the same item gets handled twice or not at all.

This is a [Reference build] for return handling. The Nacha operating rules govern the return codes and timing, so check the current rules for your case.

Where a bot helps and where a person decides

Automate the match, the lookup and the posting for returns that match one payment with one reason code. Send every partial match or duplicate to a queue with the candidates already listed.

A person decides on closed accounts, disputes and anything that touches the customer relationship. Queues in tools such as UiPath Orchestrator give that hand-off a place to live (UiPath Orchestrator documentation).

Why scripted bots and generic tools fall short

A scripted bot works until a screen changes, and then it fails quietly. A generic workflow tool works until a case does not fit its form. Both are fair choices for simple steps.

The honest case for plain RPA is real. If the screens are stable and the volume is high, a bot from UiPath, or desktop flows in Microsoft Power Automate (Microsoft Learn, Power Automate desktop flows), is quick to stand up.

The failure mode is upkeep. Each core-system upgrade can break a selector, and the bot may post a wrong value instead of stopping. Build a stop-on-doubt rule into every bot, and an owner for each one.

A bot hitting a changed screen and routing the item to a person instead of posting

Where AI helps and where plain rules are enough

Use rules for anything a clerk could write down as a checklist. Use a model only for reading messy input or drafting text. Use a person for the decision.

StepPlain rulesModelPerson
Re-key data between systemsYesNoNo
Read a scanned ID into fieldsNoYes, with a confidence scoreReviews low scores
Match an ACH return to a paymentYesOnly for fuzzy casesResolves ambiguity
Draft an alert summaryNoYes, as a draftEdits and signs
Accept an identity or file a reportNoNoYes

In the US, the Federal Reserve's SR 11-7 guidance on model risk management applies to models in use at supervised banks (Federal Reserve, SR 11-7, 2011). A model in your pipeline brings validation work. Ask your model risk team early.

Automate the gathering and the re-keying, and keep the decision with a person who can be held to account.

Cost drivers, and what gets underestimated

Cost follows four things: how many systems the bot touches, how many exception types exist, what controls your regulator expects, and the upkeep after go-live. We do not publish a price for our own work.

Viithiisys scopes this per project after a discovery call, on a fixed scope or a fractional basis. Any figure you see in a blog post for "an RPA project" is not a quote.

What gets underestimated is the exception queue. [Illustrative example] Suppose a bot handles most returns cleanly and the remainder lands with staff. If that remainder is large, the people-hours do not disappear. They move into a queue that nobody sized.

Also underestimated: audit logging, access control for bot credentials, and a named owner when a core system is upgraded. Price the upkeep separately from the build.

Send us the workflow that gets stuck

If one of these three looks like yours, or yours is different, we will look at it with you. Tell us the systems involved, the volume, and where the work waits.

A broken workflow assessment is how we start: you describe the queue, and we map where it jams and what a build would need. For wider context, see our AI workflow automation work, how document AI reads messy inputs, and our approach to system modernisation when the core has no usable interface.

If you would rather talk first, use our contact page.

Sources

FAQ

What is RPA in banking?
RPA in banking is software that operates the same screens and files a clerk would, to move data between core banking, case and document systems. It follows fixed rules. It suits stable, high-volume steps and struggles when inputs vary or when a screen changes without warning.
Which banking processes suit RPA best?
Steps with stable rules, structured inputs and a clear owner suit RPA best. Examples are re-keying data between a core system and a case tool, assembling alert packets, and logging payment returns. Steps that need judgement, such as a coverage-style decision or a filing call, should stay with a person.
Does RPA replace a core banking integration?
No. A bot working through screens is a workaround for a missing API, and it is more fragile than an API. If the core system, such as Fiserv, FIS or Jack Henry, exposes a supported interface for the task, use that first and keep the bot for the gaps.
What does an RPA project in a bank cost?
It depends on the number of systems touched, how many exception types exist, the controls your regulator expects and the ongoing upkeep. Viithiisys scopes this per project after a discovery call, so no figure is published here. Ask any vendor to price maintenance separately from the build.