What Is Intelligent Document Automation? — Lapu AI

Intelligent document automation is AI that reads invoices, contracts and forms — then acts on the data. Unlike template-based capture or cloud-only intelligent document processing suites, a desktop AI agent runs it where the file already sits, keeps the extracted data on your machine and asks permission before writing to any system of record.

Most guides collapse three separate categories into one. Robotic process automation (RPA) replays clicks. Intelligent document processing (IDP) extracts fields into a record. Intelligent document automation reasons about the whole document and runs the workflow that follows. The distinction matters because the failure modes are different — and the buyer question "should we build this on our RPA platform, our IDP platform or a desktop AI agent?" turns on which of the three layers your work actually needs.

What is intelligent document automation?

Intelligent document automation is AI software that reads a document, extracts the data that matters and executes the next step in the workflow — an approval, a system update, a routed exception — without a human queuing each action. It combines document understanding (what does this document say?) with decision execution (what should happen next?).

Three moves separate it from the older stack:

• No templates. A vision-language model looks at the page and decides which region is the invoice total, the contract effective date or the shipping address. There is no invoice-vendor-42.template to build.
• Multi-step reasoning. After the extract step the model decides which tool to call — post to the accounting system, route to a reviewer, request a missing field from the sender — instead of handing a flat record to a downstream job.
• Exception handling in the same loop. Low-confidence fields become questions the agent asks (or escalations it queues), not silent rows in a "review later" bucket.

This is the shape MIT Technology Review names when it describes 2026 as the year "AI agents actually go to work" — the shift from a browser or a code snippet to processes an agent runs end-to-end. Document work — the largest single category of back-office labor — is where that shift lands first.

Intelligent document automation vs IDP vs RPA

The three categories are stacked, not competing. The visual above traces the difference; here is the buyer version.

RPA (UiPath, Power Automate Desktop, Blue Prism) drives a stable UI on a fixed script. It is fast, cheap on a per-task basis and deterministic. It also breaks the day a screen refactor moves a button, and it cannot read a document it has not been trained on.

Intelligent document processing (ABBYY, Rossum, Hyperscience, Kofax) uses OCR plus machine learning to extract fields from a document into a structured record. It handles new suppliers and layout drift better than RPA. It stops at the record — the "process automation" bit is still someone else's job.

Intelligent document automation is the top layer. The model reads the document and decides the next action and calls the tool for that action — approving a purchase, updating a customer record, filing a claim, escalating a mismatched invoice. Anthropic's computer use announcement puts the enabling piece plainly: developers can now direct a model to "use data from my computer and online to fill out this form" — the model perceives the screen, moves the cursor, types the values and clicks through the flow.

Practical rule: RPA is the right layer when the UI and the document are both stable. IDP is the right layer when the document is unstable but the downstream workflow is a straight pipe. Intelligent document automation is the right layer when the workflow itself has judgment in it — routing, exception handling, cross-app coordination.

How AI reads a document without templates

Template-based extraction says "the invoice total is at (x=482, y=1220)" for each new supplier. Model-based intelligent document automation reads the page like a person and asks "which text on this page is the invoice total?" — then verifies against the sum of line items, checks the currency and returns a confidence score.

The three-step read is the spine of every modern document agent:

1. Perceive the page — the model receives the page as an image plus (optionally) an OCR text layer, and looks at both. Vision-language models handle scanned receipts, stamped forms, redacted contracts and mixed-language documents that pure OCR pipelines choke on.
2. Extract the entities — the model returns the fields you asked for (vendor_name, invoice_number, line_items, net_terms) as structured JSON, with a per-field confidence score and the bounding box it read the value from. No template, no per-vendor training.
3. Verify and decide — the agent checks the extraction against known constraints (line-item sum, tax math, purchase-order match), asks a question when confidence is low and calls the next tool — write to the ERP, route to Slack, file the exception.

The PDF data extraction guide on the desktop walks the same pattern on the specific PDF-to-Excel case. The shape holds for every document type: perceive, extract, decide.

Where intelligent document automation runs: desktop vs cloud

Most intelligent document automation suites are cloud services. You upload the document, the vendor processes it, the result comes back. That shape fits a public-web workload — expense receipts, unclassified PO scans, ID uploads — where a cloud round-trip is the right trade-off.

It is the wrong shape for three real cases:

• Compliance-restricted documents. Signed contracts under legal privilege, health-plan enrollment forms with PHI, mortgage packets with SSNs, patient records — the buyer cannot ship these to a third-party server on principle, and the vendor's "we don't retain data" bullet is not a policy the auditor accepts.
• Documents that live in legacy Windows apps with no API. The invoice starts in an on-prem accounting system that only opens through a Win32 UI. There is no upload button; the file only exists as a rendered screen. See the Windows automation hub for the wider pattern.
• Workflows that cross more than one desktop app. Read the invoice PDF, look up the vendor in an on-prem ERP, check the PO in a spreadsheet, write the approval back to email. A cloud IDP handles step 1; the other three steps are on the operator's machine.

Anthropic's Claude Cowork describes the desktop shape directly — the agent "moves between" local files, folders and applications, "synthesizing information across multiple sources," and returns a finished deliverable. That is intelligent document automation without a cloud round-trip on the file body. Lapu AI takes the same shape: the model reads the document where it sits, the extracted data stays on the machine and every write to a downstream system goes through a permission gate that lands in a local audit trail. For the mechanics of that permission-gated loop, see how desktop AI agents work.

Common intelligent document automation use cases

The cluster of jobs intelligent document automation handles today, in rough order of how often they show up in real buying conversations:

• Invoice processing. Read the invoice, match it to the purchase order and receiving record, post the approval, escalate the mismatch. This is the workhorse and the reason "invoice automation" is the largest single spend line in the category — see the invoice processing agent for the buyer view.
• Data entry from mixed documents. Read a form, a scanned application, a supplier onboarding packet or an ID card and write the values into an internal system with no exposed API. The data entry automation agent is the anchor page for this cluster.
• Contract intake and classification. Read new contracts, extract counterparty, effective date, renewal terms and jurisdiction, route to the right business owner. Reduces the "unmanaged contracts folder" problem without a full CLM rollout.
• Insurance claims triage. Read the FNOL packet, classify the loss type, extract the fields required by the adjuster system, flag missing documents back to the submitter.
• Government filings and back-office paperwork. Read the packet, populate the destination system, produce the acknowledgment. Sits under the broader back office automation umbrella.
• Onboarding and KYC. Read the passport, extract the fields, verify against the sanctions list and post the account request.

Each of these is more than an OCR-plus-mapping job. Each has a routing decision, an exception path and a downstream write — the reason it is intelligent document automation, not intelligent document processing.

What to look for in an intelligent document automation tool

Five questions separate the tools that survive contact with real documents from the ones that demo well:

• Does the extractor work without a per-document template? If the vendor asks you to draw bounding boxes for every new supplier, this is IDP-with-a-model, not intelligent document automation.
• Where does the document body actually live during processing? If the answer is "our cloud," ask the security team before the pilot. If the answer is "the operator's machine," you can pilot next week.
• What does the agent do when confidence is low? A tool that silently writes low-confidence rows into your accounting system is worse than one that stops and asks. Look for an explicit exception queue with the reason, the bounding box and the suggested value.
• Can it drive a downstream app that has no API? Most systems of record in real back offices are twenty-year-old Win32 apps. A tool that only integrates through REST calls covers the easy 20% of the workflow.
• Is there a real audit trail for every write? Anything the agent puts into a system of record needs a timestamped, per-decision record that the internal audit team can review — not a summary log after the fact.

If the answer to more than one of those is no, the tool is the wrong shape for the workflow — pick a different one before the pilot.

To try intelligent document automation on your own documents with a desktop AI agent — no upload, no template, permission-gated on every write — download Lapu AI and point it at the folder your invoices, contracts or forms live in today. The wider category context sits on the document automation agent page and the Windows automation hub. For the underlying model + tool loop that makes this work, how desktop AI agents work walks the mechanics.

FAQ

What is intelligent document automation?

Intelligent document automation is AI software that reads a document, extracts the data that matters and executes the next step in the workflow — an approval, a system update, a routed exception — without a human queuing each action. It handles unstructured layouts (invoices, contracts, forms) without templates, because the model reasons about a document it has never seen instead of matching pixels to a fixed map.

Is intelligent document automation the same as intelligent document processing?

No, but they are close. Intelligent document processing (IDP) captures, classifies and extracts data from documents and hands the result to another system. Intelligent document automation goes further: after the extract step it decides the next action, calls the right tool or API and closes the loop. IDP fills the record; intelligent document automation runs the process the record was for.

How is intelligent document automation different from RPA?

RPA replays a fixed script of clicks and keystrokes against a stable UI. Intelligent document automation reasons about a document it has never seen, adapts when a field moves or a new supplier layout arrives and calls the right tool for the next step. RPA breaks when the layout changes; intelligent document automation re-reads the layout on every run.

Does intelligent document automation require templates?

No. Modern intelligent document automation is model-based, not template-based. A vision-language model looks at the page, decides which region is the invoice total or the contract effective date and returns the value. There is no template file to build or maintain for every new supplier or contract counterparty.

Can intelligent document automation run on the desktop instead of the cloud?

Yes. A desktop AI agent reads the document from your local file system, runs the extraction in-process and writes the result back to your machine — with a permission gate on any external system it touches. That shape fits documents that cannot leave the network for compliance reasons, and workflows that cross legacy Windows apps with no exposed API.

What documents does intelligent document automation handle?

Any semi-structured document with recognizable fields — invoices, purchase orders, receipts, remittance advices, insurance forms, mortgage packets, government filings, contracts, HR letters, ID documents. It also handles fully unstructured text (long-form contracts, memos, emails) by extracting entities and clauses instead of fixed fields.

How accurate is intelligent document automation compared to manual entry?

Category-defining IDP vendors report accuracy above 90% on common document types, with human-in-the-loop review reserved for the low-confidence tail. Accuracy on any given document depends on print quality, layout complexity and how much training data the model has seen for that document family — never trust a vendor number in isolation, always test on your own worst 100 documents.

Sources

1. Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku — Anthropic (2024-10-22) · accessed 2026-07-07
2. 10 Things That Matter in AI Right Now — MIT Technology Review (2026-04-21) · accessed 2026-07-07
3. Claude Cowork — Anthropic's agentic AI for knowledge work — Anthropic (2026-03-24) · accessed 2026-07-07