# From RPA to Intelligent Automation
In 2019, a global consumer goods company celebrated deploying 500 software robots across its finance shared-service center. Two years later, an internal audit found that roughly 40% of those bots were either dormant or breaking weekly. The culprit wasn't the technology. It was the sequencing: they had automated broken processes, scattered bots across every squeaky wheel, and built no capacity to maintain what they'd created. The robots worked. The economics didn't.
This is the trap most finance automation programs fall into. The question is not "Can we automate this?" Almost everything can be automated to *some* degree. The question is "What do we automate, in what order, and how do we capture the payback before the maintenance cost eats it?" That is a CFO's problem, not a technologist's.
Finance automation is not binary. It is a ladder, and each rung has a different cost structure, risk profile, and payback curve. Confusing the rungs is where money dies.
Robotic Process Automation mimics human keystrokes across systems. It logs in, copies a field from an invoice, pastes it into the ERP, clicks submit. It is deterministic: given the same input, it always does the same thing. It sees no meaning—it sees screen coordinates and data fields.
RPA earns its keep on high-volume, stable, rules-based, structured-data tasks: moving intercompany balances, triggering payment runs on approved invoices, pulling bank statements into a reconciliation template. The payback is fast (often under 12 months) and easy to model: fully loaded FTE cost times hours saved, minus license and build cost.
Its fatal flaw is brittleness. RPA has no judgement. Change a vendor portal's layout, add a column to a report, and the bot fails silently or floods a queue with exceptions. The 40% dormancy rate in our opening example was almost entirely a Rung 1 problem—bots built against processes that kept changing.
The next rung adds *perception*. Optical character recognition combined with machine learning lets the system read semi-structured documents—invoices in a hundred different formats, remittance advices, bank confirmations—and extract fields with confidence scores. Unlike RPA, IDP tolerates variation. A new invoice layout doesn't break it; the model generalizes.
This is where accounts payable transforms. Instead of RPA rigidly reading invoices that all look identical (they never do), IDP handles the messy reality of supplier documents and hands clean, structured data down to an RPA bot for posting. The two rungs work together: IDP for perception, RPA for execution.
Here automation starts making probabilistic decisions. In reconciliations, an ML model learns which transactions historically matched—even when reference numbers are missing or amounts differ by a rounding cent—and auto-clears them with a confidence threshold. In the close, anomaly-detection models flag journal entries that deviate from historical patterns before they hit the ledger.
The economics shift here. Rung 3 doesn't just save hours; it changes the *shape* of the work. Reconciliation moves from "match everything manually" to "review only the 5% the model can't confidently clear." That is a step-change in throughput, but it introduces a new cost: model governance. Someone must monitor drift, retrain, and own the false-positive/false-negative tradeoff.
The top rung uses large language models and agents that can interpret unstructured context, draft narratives, and orchestrate multi-step workflows with reasoning. A close agent can read variance data, draft the flux commentary, cite the drivers, and route it for review. A dispute agent can read a customer's email, pull the relevant invoice and contract terms, and propose a resolution.
This rung is powerful and immature. It hallucinates. It requires human-in-the-loop controls, especially anywhere near the general ledger or external reporting. The right posture in 2024–2025 is to deploy Rung 4 on *draft-and-review* tasks (commentary, first-pass analysis, exception explanation)—never on unsupervised posting of financial entries.
The instinct is to start at the top of the ladder because it's the most impressive. That is exactly wrong. The correct sequence is driven by three questions, applied in order.
Before any technology decision, apply the eliminate–simplify–standardize–automate discipline. Automating a process you should have killed just makes waste run faster. If three approval steps exist because of a control designed for a risk that no longer exists, remove them. If four regional teams reconcile the same account four different ways, standardize *first*. The consumer goods company's dormant bots were often automating variant processes that should have been standardized into one.
The rule: never automate a process you wouldn't be willing to freeze for 18 months. If it's still changing, RPA will break and the ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → evaporates in maintenance.
MapMapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → every candidate on two axes: volume/frequency and process stability. High-volume, high-stability tasks are Rung 1 gold—fast, cheap, durable payback. High-volume but variable-input tasks (AP invoice capture) need Rung 2. Judgement-heavy, pattern-rich tasks (reconciliation matching, anomaly detection) need Rung 3.
Crucially, model the total cost of ownership, not the build cost. A useful discipline: assume maintenance runs 15–30% of the initial build cost *annually*. A bot that saves $80k in labor but costs $25k a year to maintain has a very different NPVNPVNet Present Value is the sum of an investment's future cash flows discounted to today, minus the initial outlay. A positive NPV signals value creation.View full definition → than the business case that ignored maintenance. This single adjustment would have flagged most of the dormant robots before they were built.
This is where CFOs consistently underinvest. Automation is not a project; it's a capability. You need:
Without this, you get the sprawl: hundreds of unowned bots, no one accountable, and a maintenance burden that quietly exceeds the savings. The scoreboard should track not just bots built but bots retired, exception rates, and net FTE hours actually redeployed—not theoretically saved.
In practice, the highest-returning programs follow a consistent arc across the finance close, AP, and reconciliations:
1. Standardize and consolidate the process (no technology yet).
2. Deploy Rung 1 RPA on the stable, structured core to bank quick, credible wins and fund the program.
3. Add Rung 2 IDP where input variation is the bottleneck (AP is the classic first target).
4. Layer Rung 3 ML on the high-judgement volume—reconciliation matching, anomaly detection in the close—once you have clean data and a governance function.
5. Pilot Rung 4 agents on draft-and-review tasks, tightly scoped, with humans owning every decision that touches the ledger.
Each step funds the next. The quick RPA wins buy political capital and budget for the harder, higher-value ML and agentic work. Skip the early rungs and you have no track record when you ask the board for the expensive stuff.
Knowledge check
1. According to the lesson, what was the primary root cause of the 40% bot dormancy rate in the opening example?
2. The lesson argues that the right question for a CFO is not 'Can we automate this?' Why is that framing considered inadequate?
3. Why does the lesson describe finance automation as a 'ladder' rather than a single leap?
4. Select ALL correct answers. Which task characteristics make a process well-suited to Rung 1 rules-based RPA?
Select all the correct answers.
5. Select ALL correct answers. Which statements accurately describe the nature and limits of Rung 1 RPA as presented in the lesson?
Select all the correct answers.
Frameworks are useless without a concrete first move. Here is how the ladder maps onto the three workhorse finance processes, and what to do first in each.
The close is a sequence of dependencies, which makes it seductive to automate end-to-end and dangerous to do so. Start narrow. Rung 1 automates the mechanical: pulling trial balances, running standard recurring journals, refreshing consolidation templates. Rung 3 adds the leverage: anomaly detection that flags unusual entries *before* they require restatement, and continuous reconciliation that shifts work out of the close window entirely (the "continuous close" concept lives here). Rung 4 drafts the flux commentary from the variance data for a controller to review and sign.
Monday-morning move: identify the three tasks in your close that (a) recur every period identically and (b) sit on the critical path. Automate those with RPA first. They compress the timeline and prove the concept.
AP is the most reliable place to demonstrate cross-rung value because it spans all four. IDP (Rung 2) captures invoice data regardless of format. RPA (Rung 1) posts approved invoices and executes payment runs. ML (Rung 3) predicts GL coding based on history and flags duplicate or fraudulent invoices. Agents (Rung 4) handle supplier inquiries and exception correspondence.
Monday-morning move: measure your straight-through-processing rate—the percentage of invoices that flow from receipt to payment with zero human touch. That single metric tells you where you are on the ladder and where the next investment goes. A world-class AP function runs 80%+ STP. If you're at 30%, the constraint is almost certainly invoice capture (Rung 2), not posting (Rung 1).
Reconciliations are the purest ML use case in finance. The work is high-volume, pattern-rich, and judgement-laden in exactly the way that punishes rigid RPA and rewards learning models. Rung 3 auto-matches transactions within a confidence threshold and routes only true exceptions to humans.
Monday-morning move: pull your account reconciliation inventory and segment by match complexity. The high-volume, low-complexity accounts (bank recs, credit card recs) are ML-ready today and often deliver the fastest step-change in analyst hours redeployed. Do not start with the gnarly, low-volume, high-judgement accounts—the ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → isn't there and the model won't have enough data to learn.
Across all three, one discipline is non-negotiable: every automation that touches the ledger needs a documented control that the automation itself cannot bypass. Auditors will ask who owns the bot, how exceptions are handled, and how you know the ML model is still accurate. Build that evidence trail from day one. Retrofitting controls onto a sprawling bot estate is far more expensive than designing them in—and it's the difference between an automation program that scales and one that gets frozen by risk and audit.
1. Automate the process, not the mess. Run eliminate–simplify–standardize *before* you automate. Never automate a process you wouldn't freeze for 18 months—instability is what turns bots dormant.
2. Sequence up the ladder, don't leap to the top. Bank fast RPA wins on stable, structured tasks to fund the program, add IDP where input variation is the bottleneck, then layer ML on high-judgement volume. Early wins buy the credibility and budget for the expensive agentic work later.
3. Model total cost of ownership, not build cost. Assume 15–30% of build cost in annual maintenance. This single adjustment reveals which "quick wins" are actually value-destroying and which deserve investment.
4. Treat automation as a capability, not a project. Stand up a prioritization committee, a center of excellence, and named bot custodians. Track bots *retired* and net FTE hours *actually redeployed*, not just bots built.
5. Instrument each process with the right metric. Straight-through-processing rate for AP, critical-path task count for the close, match-complexity segmentationsegmentationDividing a market into distinct groups of customers who share similar needs, characteristics or behaviours, so each group can be served with a tailored approach.View full definition → for reconciliations. The metric tells you which rung you're on and where the next dollar goes.