Can an AI agent read your documents and pull out the data?

Can an AI agent do document processing?

Yes, with a person on the uncertain cases. An AI agent for document processing can take files from an inbox or folder, read every page, work out what each document is, split bundled PDFs, extract the fields you define and check them against your rules. Organizations run this in production today. AWS reports that Built Technologies, a real estate finance software provider, processes commercial construction loan draw packages this way, cutting classification and extraction work that took 3–9 days to minutes per package. Microsoft reports that Direct Pojišťovna, a Czech insurer, cut windshield claim handling from 15 minutes to about 2 and fully automated 60% of those claims. Both figures come from the vendors.

Both keep people involved, in different ways. Built gives each field a confidence score and routes results below its required threshold (over 95 percent confidence) to reviewers. Microsoft's story says Direct's system was built "with a human always in the loop", and Direct's head of IT calls human-in-the-loop essential; the same story says 60% of windshield claims were fully automated and that uncomplicated claims are paid out inside the AI-supported process. Our recommendation is to treat this as a supervised hire: the agent does the reading and keying, and a person checks what the agent is unsure about and owns any decision made on the data.

This page covers documents in general: forms, claims, contracts, certificates, applications. For supplier invoices see accounts payable, and for typing data that is not on a document, see data entry.

What the agent does and what stays with you

Agent

3of 9

• Collect incoming documents from an inbox, upload folder or portal
• Read the text, tables and handwriting on each page (OCR)
• Hand the structured data to the downstream system

Agent with review

4of 9

• Classify each document and split bundled PDFs into separate documents
• Extract the fields you defined into structured data
• Check extracted data against business rules or other records
• Draft the extraction schema for a new document type

Human

2of 9

• Correct low-confidence fields and keep the test set up to date
• Act on the data (approve, pay, file, reply to a customer)

The split follows one rule: the agent does work a reviewer can check against the page, and a person does work that changes what the company does next. Collecting, reading and handing over data can be checked by counts and samples, so they go to the agent. Classification, extraction and rule checks carry a confidence score, and anything below your threshold gets a human look. Deciding what to do with the data stays human.

The confidence score is the control that makes this work. Microsoft's documentation describes a field confidence of 0.95 as a prediction likely to be correct 19 times out of 20, and says confidence can be used to decide whether to accept a value automatically or flag it for review. Microsoft also notes that not every field returns a confidence score, so ask any vendor which fields have one.

A worked example: one certificate of insurance

Assume, for this example, that a property manager receives certificates of insurance from contractors and must confirm each one names the company as additional insured with at least $1,000,000 in general liability cover.

A contractor emails a 4-page PDF that holds a cover letter, the certificate and an endorsement. The agent picks it up from the intake mailbox, reads the pages and splits the file into three documents. It extracts the insured name, insurer, policy number, expiry date, liability limit and additional insured wording from the certificate. The limit reads $1,000,000 with high confidence. The expiry date is partly covered by a stamp, so its confidence falls below the threshold. The rule check passes on the limit but cannot confirm the expiry date.

The certificate lands on the review queue with the date field highlighted on the page image. A reviewer reads the date from the endorsement, corrects it and marks the document complete. The record goes to the vendor-compliance system. The agent sent nothing to the contractor and approved nothing; if the cover had been too low, the property manager would decide what to ask for.

What you need to provide

The agent touches whatever is in the documents, which often includes personal data (names, addresses, ID numbers) and financial data (amounts, account numbers). Give it the least access that gets the job done:

• Read access to one intake mailbox, folder or portal, limited to the document types in scope.
• Real examples of each document type, including bad scans. For custom models, Microsoft advises including every variation of a document in the training set, for example digital versus scanned PDFs, and adding at least five samples of each PDF type if the model must analyze both.
• A written field list for each type: what to extract, in what format, and where it usually appears. Built's team found that a field description saying where a value appears and what it is called guides the model better than a field name alone.
• Your business rules in plain language.
• A labeled test set: 50–100 documents with the correct values filled in by the person who owns the process. This is the yardstick for the trial.
• A user in the target system that can create records but cannot approve, pay or send.

Ask the provider where page images, extracted text and logs are stored, and for how long. Your own retention rules apply to every copy.

Job brief template

Use this as the brief for a vendor, an operator or an internal team setting up the agent. Replace the bracketed values with your own.

Keep the labeled test set even after go-live. It is how you check every schema change, model change or new document type before it touches live data, which is how Built runs its own changes.

What it costs

The human figure is the 2025 national median annual wage for data entry keyers (SOC 43-9021), $41,340 or $19.88 an hour, as published on O*NET OnLine from BLS data. It is wage only. The agent figures are list prices for two cloud extraction services, billed per page. Amazon Textract charges $0.015 per page for Analyze Document with Queries for the first 1 million pages a month in US West (Oregon). Google Document AI, priced pay as you go, charges $30 per 1,000 pages for its Custom extractor up to 1 million pages, or $0.03 per page, plus $0.05 an hour to host a custom processor version. Google also lists prebuilt parsers priced per document, such as $0.10 for an invoice of up to 10 pages.

These prices cover reading and extraction only. A full agent also pays for model calls in its classification, validation and reasoning steps, for running the pipeline and for building the integration. Packaged document-processing products may bundle those costs; their prices were not checked for this page. Reviewer time on flagged documents is a further cost outside both numbers, so measure it in your trial.

Risks and human checkpoints

The risk that matters most is the confident mistake. A flagged field gets a human look; a wrong value with a high score goes straight into your system. Microsoft notes that a document set can look similar to people but dissimilar to a model, and that reported accuracy can be inconsistent when analyzed documents differ from the training documents. A weekly sample of unflagged documents is the only way to see this.

The quieter risk is lost volume. Pipelines throttle, time out and skip files. Built's team reports having to resolve Amazon Textract throttling limits when processing large documents at high concurrency. A daily count of documents in against documents out and documents queued catches gaps before a customer or supplier does.

How to run a trial

Run the agent next to your current process for 3–4 weeks before you rely on it.

1. Pick 2–3 document types with steady volume and a mix of clean and messy copies.
2. Build the labeled test set first, and score the agent on it before it sees live documents.
3. Set the agent's permissions to write only to a staging area or review queue. Your team keeps processing as usual.
4. Each week, compare 50 agent records with what your team keyed. Count wrong types, wrong fields, missed flags and false flags against the brief's acceptance criteria.
5. Time your reviewers. Minutes per flagged document, times the share flagged, is the number that decides whether the agent saves money.
6. Feed in a few deliberately awkward files: a rotated scan, a bundled PDF, a new layout. Confirm each lands on the review queue rather than in the system.

1. 1

   Pick 2–3 document types

   Steady volume and a mix of clean and messy copies

2. 2

   Build the labeled test set first

   Score the agent on it before it sees live documents

3. 3

   Limit writes to a staging area or review queue

   Your team keeps processing as usual

4. 4

   Each week, compare 50 agent records with your team's

   Wrong types, wrong fields, missed flags and false flags against the acceptance criteria

5. 5

   Time your reviewers

   Minutes per flagged document times the share flagged decides whether it saves money

6. 6

   Feed in a few awkward files

   A rotated scan, a bundled PDF, a new layout; each should land on the review queue

If the results meet the criteria for two weeks in a row, let that slice write to the live system and keep the weekly sample. For more on structuring the test, see how to run a paid trial with an AI agent, and for writing the brief, how to write a job description for an AI agent. Related jobs: expense review and audit, the administrative assistant role and the operations overview.

Sources

1. 1
   Built Technologies builds an AI-powered document intelligence solution on AWS to power agents across real estate finance

   Amazon Web ServicesVendorAccessed Nov 14, 2026

2. 2
   Direct Pojišťovna automates claims, cutting processing time by 87% with Azure AI

   MicrosoftVendorAccessed Nov 14, 2026

3. 3
   Interpret and improve model accuracy and confidence scores

   Microsoft LearnPrimaryAccessed Nov 14, 2026

4. 4
   43-9021.00 - Data Entry Keyers

   O*NET OnLine (U.S. Department of Labor)PrimaryAccessed Nov 14, 2026

5. 5
   Amazon Textract pricing

   Amazon Web ServicesVendorAccessed Nov 14, 2026

6. 6
   Document AI pricing

   Google CloudVendorAccessed Nov 14, 2026