What “agentic” actually means for document processing

Agentic AI means software that plans and carries out multi-step tasks on its own: it decides what to do next, does it, checks the result and moves on, with a person supervising rather than driving. Applied to document processing, an agent might receive an email, recognise the attachment as an invoice, extract the data, check it against a purchase order and post it into your ERP without anyone touching a keyboard.

That is the promise on nearly every vendor slide this year. The part the slides skip: every one of those steps depends on the first one, reading the document correctly. An agent acting on a wrongly read total does not remove the mistake; it posts the mistake into your ERP faster than a person ever could.

So the useful question is not “is this agentic?” It is “how well does this read my documents, and what happens when it is not sure?” This article covers where the word came from, where agent-style automation already works in document processing, and the four things worth evaluating instead of the buzzword.

Why every vendor suddenly says it

The technology underneath the word is real. Large language models can now plan a sequence of steps, use tools and adjust when something unexpected happens, and that opened a real new class of software. The marketing moved faster than the software did. Between 2025 and 2026, “agentic” became the word analysts reward and procurement documents request, so products were renamed to match. In many cases the pipeline behind the new word is the same one that was called intelligent automation the year before, and workflow the year before that.

None of that makes the word meaningless. It makes it unreliable as a buying signal, which is a different problem.

What an agent is, in plain terms

Think of three generations of office automation. A rule does exactly one thing: if the invoice is from this supplier, put it in that folder. Robotic process automation replays a recording of what a person clicks, and breaks the day the screen changes. An agent is given a goal rather than a script: get this invoice into the ERP correctly, and it chooses the steps itself.

That last one is a decision-maker. In a business process, decisions have two requirements the demos rarely mention: they need to be right, and someone needs to be able to see afterwards why they were made. A goal-seeking system without an audit trail is not automation; it is an unsupervised employee.

The part that actually matters: the reading

Document processing has a fixed order of operations: before anything can be decided, the document has to be read. Not a clean PDF export. The real thing: the handwritten delivery note, the stamped and re-scanned contract, the statement photographed at an angle, the invoice format nobody has seen before. That reading step is where document automation succeeds or fails, and it is exactly the step the agentic conversation glosses over.

We process documents for a living, more than 500 million pages of them since 2014, and the pattern holds: when the reading is accurate, almost any downstream automation works, agentic or otherwise. When the reading is wrong, the sophistication of what follows only spreads the error further. If the term IDP is new to you, our explainer covers it; reading is the whole first half of that discipline.

Where agency already works in document processing, quietly

Here is the irony: unglamorous, rule-driven agency has been running in mailrooms for years. Software that watches inboxes, recognises what each attachment is, splits one email across the ERP, the CRM and invoice processing, files documents to the archive by rule, and escalates anything it cannot place. Multi-step, autonomous, goal-directed. Nobody called it agentic. It just worked.

At a South African scanning bureau, invoices, credit notes and debit notes once needed 26 people to download, classify and capture within a 24-hour SLA. Leap’s mailroom automation now reads the inboxes and extracts the data itself, up to 1,500 invoices a day; today 2 people validate the exceptions. Every action in that flow is a rule someone chose, with a log someone can read. That is what agency looks like when it has to survive an audit. (What a digital mailroom costs and saves covers the numbers side.)

What to evaluate instead of the buzzword

Four checks tell you more than any label on a slide:

1. Reading quality on your documents. Not a demo set. Send the vendor your fifty hardest: crumpled, handwritten, stapled, scanned at an angle, and look at what comes back.
2. Behaviour at low confidence. When the system is not sure, does the document go to a person, or does a wrong answer go to your ERP with a confident face? The honest answer to uncertainty is a queue, not a guess.
3. Life after extraction. Data leaves the page, and then what? Where is the document kept, how is it found next year, and who is tracking the paper original? Extraction without lifecycle is half a system.
4. The audit trail. For every automated decision: what was decided, by what rule or model, when, and who confirmed it. If that question gets a vague answer, so will your auditor.

And one more, about fit rather than function: ask how long until it works on your documents, in your systems. Full implementations of our platform go live in days to a few weeks; the confirmed range is two days to four weeks, configuration and client-specific fit included.

Humans in the loop, on purpose

The strongest document operations we see are not the ones with the fewest people; they are the ones where people stopped doing the machine’s work. Software does the reading, at volumes no team could match. People confirm the judgement calls: the smudged total, the unfamiliar layout, the exception that needs a phone call. Each confirmation teaches the system, so the queue gets shorter every month.

That is a design choice, not a limitation, and it is the subject of the third article in this series. Working document AI in days, with a human validating every decision: less exciting on a slide, considerably more useful in a mailroom.

Frequently asked questions

Is agentic AI the same as RPA?

No. RPA replays a scripted sequence of clicks and breaks when the screen changes. An agent is given a goal and chooses its own steps. Both, when applied to documents, depend entirely on how accurately the documents are read first.

Do I need agentic AI to automate document processing?

No. Most of the value in document automation comes from three things: accurate extraction on your real document types, rule-based routing you control, and human validation of exceptions. If a system delivers those, the label on it does not matter.

Where do people fit in an automated document process?

At the exceptions. The system handles everything it reads with confidence and queues what it is unsure of; people confirm or correct those cases, and the system learns from each decision. Judgement stays human; keystrokes do not.

What should I ask a vendor who says “agentic”?

Four things: show it on my documents, not yours; show me what happens when it is not sure; show me where the documents live afterwards; and show me the audit trail for one automated decision, end to end.