Why Digital Transformation Automated Only 15% of Banking Workflows (and What Closes the Gap)

Banks have spent two decades and enormous budgets on digital transformation, and a large share of core workflows still run through manual review. Digital transformation automated the structured, rule-based slice of banking work. Everything that required judgment, unstructured data or multiple steps of reasoning stayed manual. This guide looks at the reasons for the gap between digital transformation and manual implementation, the approach to close it and the implications for five financial services verticals.

In the past ten years, most banks have focused on developing quicker apps, improving their portals, and making their dashboards cleaner. Yet behind that digital front end, key decisions, credit decisions, KYC reviews, and reconciliations have remained as they were fifteen years ago, still carried out through inboxes, PDFs and a person deciding what happens next.

Roughly 15 per cent of banking workflows, the structured, repeatable, rules-based parts, got automated. The remaining 85 per cent stayed manual, because the tools available could not handle work that requires judgment, unstructured data or multiple steps of reasoning. That split is Dailoqa's own reading of where the automation ceiling sits, not a published industry figure, but it lines up closely with what the research below shows about why the ceiling exists.

This guide looks at what got automated, why the rest resisted every tool thrown at it, and what closes that gap without another multi-year transformation programme.

What is Digital Transformation in Banking?

Digital transformation in banking is the shift from paper-based, manual processes to digital channels and tools, mobile apps, online portals, cleaner dashboards, along with back-office automation such as RPA. It has run for roughly the last decade and has reshaped how customers interact with a bank far more than how the bank actually makes decisions.

The majority of the investment in banking was directed towards the front end, that is, the parts which a customer sees and touches. A much smaller amount was allocated to the back-office judgment work involved in deciding whether a loan is approved, whether a transaction is flagged, or whether a claim is paid. This imbalance explains the 15/85 split in this guide.

The 15% Digital Transformation Actually Automated

Robotic process automation, which has been at the heart of the digital drive of the last decade, is skilled at one particular activity: carrying out structured, rule-based back-office jobs. This includes reconciliation, payment processing, regulatory report generation, routing KYC documents, and data entry into various core banking systems. These workflows have fixed steps, and the data arrives in the same form each time.

McKinsey's own framing of the problem is direct. Running a complex banking workflow, evaluating a commercial customer's loan application, for instance, involves highly variable steps and a mix of structured and unstructured data. Traditional automation cannot handle that. It was never built to [1].

This is part of why BCG's landmark study of 70 leading companies found that 70 per cent of digital transformation programmes fell short of their objectives, with only 30 per cent fully meeting their targets and delivering sustainable change [2]. The transformations that did succeed tended to target narrow, well-defined domains, exactly the kind of structured work automation was built for. The 15 per cent that got automated is not a failure of ambition. It is the ceiling of what rule-based tools were ever capable of reaching.

Why the Remaining 85% Stayed Manual

The manual work has three characteristics: it calls for judgment rather than merely involving the application of rules; it deals with unstructured data such as documents, conversations, and exceptions which do not come in a predictable format; and it usually involves multiple systems and steps instead of remaining in one place.

Backbase's 2026 analysis of RPA in banking shows the real cost involved. About one in five loan applications is abandoned during onboarding because of KYC and AML friction that a rules-based bot cannot deal with on its own [6]. A bot can copy data from a form into a core system. Still, it cannot read a handwritten note, assess whether a document appears forged, or determine how to proceed when a customer's situation does not fit any predefined route.

Preparing a credit memo is the clearest example of why this work resisted automation for so long. It draws on multiple conversations with a customer, several types of documents, and a judgment call about risk that a rules engine was never designed to make. That is not a gap in effort. It is a gap in what the underlying technology could do.

Why Surface-Level Digital Fixes Are Not Enough?

Adding another portal, another dashboard, or another RPA bot does not close the 85 per cent gap. Each of those fixes still runs on the same rule-based logic that hit its ceiling a decade ago, a faster interface sitting on top of the same manual judgment work underneath.

The issue was never that banks hadn't digitised enough; rather, the tools available were only capable of handling structured, repetitive tasks. A new customer portal won't help a credit officer interpret an unusual document, and a more streamlined dashboard won't determine whether an exception during onboarding represents a real risk or is just a false alarm. To close the remaining 85 per cent will require a different kind of technology, one that can reason across unstructured data and multiple steps, not just an improved version of the same rule-based method.

What “Computation Without Coding” Actually Means

While RPA tried to close the gap, agentic AI does so differently. Rather than requiring a developer to write and maintain rules for each possible route through a workflow, an agentic system reasons about the task, plans a sequence of steps, and adjusts when something does not fit the expected pattern. The logic is set up using natural language and structured inputs instead of hand-coded rules for each exception.

A 2025 survey by Google Cloud among 556 leaders in the financial services sector showed that organisations are now taking this type of system beyond the pilot phase and putting it into live production, particularly for the multi-step, judgment-adjacent tasks which RPA was never able to handle [4]. Similarly, Gartner's 2026 study on agentic AI, from the perspective of analysts, draws the same distinction: generative and rules-based tools produce or carry out actions within set limits, whereas agentic systems take action and alter their course in response to changing circumstances [5].

Computation without coding keeps the same rigour and the same governance. What changes is the number of custom-built rules that break the moment a document looks slightly different from the last one.

Legacy Automation vs Agentic AI vs Broccoli™: A Comparison 

Dimension	Legacy digital automation (RPA)	Agentic AI	Broccoli™ Agentic AI platform
Handles structured data	Yes	Yes	Yes
Handles unstructured data	No	Yes	Yes, with compliance rules encoded
Adapts to exceptions	No, breaks or escalates	Yes, reasons through most cases	Yes, escalates only genuine judgment calls to a human
Configuration approach	Custom code per workflow	Computation without coding	Computation without coding, bounded by financial services rules
Typical share of workflow reached	The structured 15%	Extends into the remaining 85%	Built specifically to reach the 85% in regulated environments

“The initial wave of automation dealt with the straightforward issues, whereas the 85 per cent which remained manual is where the real risks and the actual costs lie. The transformation that is important at this stage is in closing that gap, not simply a further project following on from the last one.”

Dr. Shuki, Partner, Dailoqa

Closing the Gap Across Five Financial Services Verticals

In every vertical the 85 per cent appears different, but the pattern is the same, the work which remains manual is that which requires judgment, involves unstructured input, or spans several systems.

Vertical	What stayed manual	How agentic AI closes it
Retail & Commercial Banking	Credit memo preparation, exception handling in onboarding	Drafts the memo from documents and conversations, escalates only the judgment call
Wealth Management	Portfolio reviews that draw on unstructured client history	Assembles the review across systems, flags what a relationship manager should decide
Risk & Compliance	Investigating flagged transactions and building case files	Runs the investigation steps, compiles the file, surfaces the decision point
Capital Markets	Reconciling trades across venues with mismatched formats	Reconciles across formats automatically, routes genuine breaks to a human
Insurance	First notice of loss processing from varied document formats	Reads varied formats, exceptions for review

Combined Intelligence in Practice

This is the operating principle Dailoqa builds around. Agents run the process across the 85 per centt that legacy automation could never reach: document intake, reconciliation,and first-pass review. Humans own the judgment on anything touching risk appetite, suitability or reputational exposure. Rules encode what compliance and policy require. Technology scales the output, and financial expertise decides what a good outcome looks like.

The extent to which an organisation is honest about the amount of its workflow that still goes through inboxes and PDFs, and the care it shows in deciding which decisions within that remaining work should remain with humans, is more important than any individual vendor's assertion.

Frequently Asked Questions on 

Why did digital transformation only automate a small share of banking workflows?

Most of the tools used in digital transformation, especially those based on RPA, were designed for structured, rule-driven tasks; tasks that involved judgment, unstructured data, or a number of steps spanning different systems were beyond the tools' capabilities and therefore had to remain manual no matter how much was invested.

What is the difference between RPA and agentic AI?

RPA operates according to fixed rules and fails when there is a change in a workflow or data format, whereas agentic AI plans and carries out a number of steps aimed at achieving a goal by reasoning through exceptions rather than stopping when it encounters the first one.

What does “computation without coding” mean in practice?

Computation without coding means setting up an agentic system's logic using natural language and structured inputs, rather than writing custom code for each workflow scenario, as traditional automation did for exceptions.

Which banking workflows are hardest to automate?

Processes that involve the combination of unstructured data with discretionary decisions, such as the preparation of credit memos, the processing of claims from a variety of document formats, and transaction investigations, have in the past been resistant to automation since they do not follow a single fixed route.

Is agentic AI a replacement for digital transformation investments already made?

No. Agentic AI extends what digital transformation started, working on top of existing digital infrastructure to reach the workflows that structured automation could not.

How much of a bank's operations could agentic AI realistically affect?

It is estimated by analysts that about half, this figure is frequently quoted, of the workforce in a typical bank carries out work which is of a type suitable for this kind of automation, although the exact number will differ from one institution and one function to another [3].

Does closing the automation gap remove the need for human review?

No, it alters the kind of cases that people review; rather than handling routine cases themselves, they concentrate on the exceptions and judgments that actually require a human decision.

What is the enterprise-grade Broccoli™ platform's role in closing this gap?

It is Dailoqa's platform for deploying agentic AI in regulated financial services, built to extend automation into unstructured, judgment-adjacent work while keeping a documented, auditable decision trail.

Conclusion

The decade was not a wasted one since it has automated the workflows which could actually be automated using the tools available back then; the fact that 85 per cent remained manual should not be seen as a failure of effort but rather as a result of the technological limitations, limitations which have since been overcome.

The  Broccoli™ agnetic AI platform is built to work on that remaining 85 per cent inside regulated financial services, not to replace the digital infrastructure already in place, but to finally reach the workflows it could never automate. If your institution is trying to work out how much of its own operation still runs through inboxes and PDFs, that is worth mapping before the next investment decision, not after it.

Talk to Dailoqa about an AI readiness assessment

References

[1] McKinsey & Company. "Banking and AI: When the Tech Starts Doing the Work, Not Just Assisting It." 2026. https://www.mckinsey.com/featured-insights/mckinsey-explainers/banking-and-ai-when-the-tech-starts-doing-the-work-not-just-assisting-it

[2] Boston Consulting Group. "Flipping the Odds of Digital Transformation Success." 2020. https://www.bcg.com/publications/2020/increasing-odds-of-success-in-digital-transformation

[3] McKinsey & Company. "The Paradigm Shift: How Agentic AI Is Redefining Banking Operations." 2026. https://www.mckinsey.com/capabilities/operations/our-insights/the-paradigm-shift-how-agentic-ai-is-redefining-banking-operations

[4] Google Cloud. "New Research Shows How AI Agents Are Driving Value for Financial Services." 2025. https://cloud.google.com/transform/new-research-shows-how-ai-agents-are-driving-value-for-financial-services

[5] Gartner. "Hype Cycle for Agentic AI, 2026." 2026. https://www.gartner.com/en/articles/hype-cycle-for-agentic-ai

[6] Backbase. "RPA in Banking Limitations: What Comes After Automation in 2026." 2026. https://www.backbase.com/blog/intelligent-automation-vs-rpa