Intelligent Process Automation: What AI Actually Adds

"Intelligent process automation" used to be a marketing label for ordinary robotic process automation with an OCR step bolted on. That is no longer quite true. A model that reads a free-text description and decides which of several paths to take is a genuinely different capability from a bot that fires on a fixed trigger, and most tools sold under this name now stack both layers together. Knowing which layer is doing which job in a specific process is the part that actually decides whether buying one helps.

The two layers, and what changed

The old layer is deterministic: a trigger fires, a fixed action runs, every time, the same way. It still does most of the actual work inside IPA and it should — find the repetitive part is still the right first question for any step you are tempted to hand to it. The new layer sits on top: a model reading something that does not arrive in a fixed shape — an invoice, a contract clause, a support ticket — and producing a structured decision the deterministic layer can act on. Automation software covers this split in general; what is specific to IPA is that vendors are now shipping agent-style orchestration on top of both, so one request can route across several systems rather than firing a single action.

You can see that third layer being sold directly: Claude Cowork and Microsoft's Copilot Cowork both package an agent that plans a multi-step task across connected tools rather than executing one fixed action, which is the same shift IPA vendors describe when they say "agentic." The pitch is real. Whether it is reliable enough for a given process without a person checking it is a separate question, and it is the one worth spending your evaluation time on.

A worked example: three-way matching

Accounts payable has run a version of intelligent process automation for years under the name "three-way match": a purchase order, a goods-received note and a supplier invoice have to agree on item, quantity and price before the invoice gets paid. Historically a clerk opened all three documents and compared them by eye.

Weak setup: point a tool at an inbox of invoice PDFs and ask it to "process accounts payable." That hides the extraction step, the matching rule and the exception-handling decision inside one instruction, with no way to tell which one failed when a payment goes out wrong.

Better setup: separate the three jobs explicitly before configuring anything.

• —Extraction — read the invoice, purchase order and goods-received note, and pull item, quantity and unit price from each into fixed fields. A narrow reading task the model genuinely handles well.
• —Matching rule — a plain comparison: do the three documents agree within a stated tolerance (commonly a few percent on price, zero on quantity). Not a model opinion on whether a mismatch "looks fine."
• —Exception queue — anything outside tolerance goes to a person with all three documents attached and the specific field that disagreed highlighted, rather than a generic "review this invoice" flag.
• —What stays manual — approving a mismatched invoice, chasing a missing goods-received note, and anything involving a new supplier with no purchase history to match against.

That shape — a Gartner survey of 183 CFOs and senior finance leaders found accounts-payable automation was the second most common AI use case in finance, at 37%, behind only knowledge management — is close to the most mature application of IPA that exists, and it is still built on separating a reading task from a judgement call rather than automating the whole thing.

Where the gain is real, and where it is overstated

A frequently misquoted detail is who benefits most. A large field study of 5,179 customer support agents using an AI assistant found issue resolution per hour rose 14% on average — but 34% for novice and lower-skilled workers, and close to nothing for people who were already experienced. The same pattern shows up in a separate study of 758 BCG consultants working with GPT-4: inside tasks the model handles well, completion rose over 12% and quality over 40% — but the paper's own point is that the frontier is "jagged," meaning the same tool fails badly on tasks that look similarly hard to a person but are not inside that frontier.

For IPA specifically, that means the honest pitch is narrower than most vendor pages: biggest gains on the newest person on a team handling a well-defined, repetitive step, not a blanket productivity multiplier across a whole department. Intelligent automation use cases walks through the same reading-then-routing shape applied to a single invoice field by field, which is a useful next read if three-way matching is not quite your process.

Where it breaks

The extraction step fails quietly rather than loudly. A model reading a scanned or oddly formatted invoice produces a confident-looking quantity and price even when it misread a digit, and a wrong number close to the real one is harder to catch than an obviously broken one, because the matching rule still runs and the exception queue stays empty. The matching rule fails differently: a tolerance copied from an old policy, a currency mismatch on an international supplier, or a rule comparing pre-tax to post-tax totals — none of those are AI failures, just configuration drift the extraction layer makes easier to miss because the output still looks complete.

Checks before it runs unattended

1. 01Sample a slice of auto-matched invoices every week and compare the extracted fields against the original documents by eye.
2. 02Confirm the tolerance in the matching rule against the current written policy, not a figure someone remembers from when the flow was set up.
3. 03Check currency and tax handling explicitly for any supplier outside your home country — a silent mismatch here is the most common real failure.
4. 04Decide in advance what a silent failure would look like: an exception rate that drops to zero, or a jump in payments later disputed by a supplier, are both worth checking rather than celebrating.

That last habit — deciding up front what a quiet failure would look like, rather than waiting to notice one — is the practical core of the NIST AI Risk Management Framework, applied here to one finance process instead of a whole organisation.

Why this is worth learning rather than just buying

Microsoft and LinkedIn's 2024 Work Trend Index, surveying 31,000 people across 31 countries, found 75% of knowledge workers already using generative AI at work, and 78% of those users bringing their own tools rather than waiting for one provided to them — but only 39% had received any training from their employer. That gap between use and training is exactly where a well-run IPA deployment either earns its budget or quietly produces more exception-queue noise than it resolves, and it is the gap business process automation strategy covers in more depth for scoring which process to automate first.

What to do Monday

1. 01Pick one document-matching or routing step you currently do by hand and separate it into extraction, a plain comparison rule, and an exception path.
2. 02Configure the automation for the mechanical reading step only, with a fixed set of fields, not an open-ended instruction.
3. 03Write the comparison rule from your current policy document, not from memory of how it used to work.
4. 04Run it against a batch of historical cases you already know the right answer for before pointing it at anything live, and sample-check the output weekly after that.

No-code process automation covers the same discipline applied to a multi-step approval chain with handoffs between departments, and data automation tools covers it for a recurring data-cleanup task rather than a document match. Examples of automation at work and IT process automation are worth reading next if the process on your list sits outside finance entirely. Coursium teaches this kind of practical, tool-specific judgement — separating what a model can check from what a person still has to decide. Stay ahead of AI by learning the tools on your phone.