Ampwise AI
    Back to blog

    Outlook PDF Invoice Processing for AP: Fix Limits, Map to ERP

    For a one-off PDF invoice, just save the attachment to your invoice folder and enter it manually. For ongoing intake, build a Power Automate flow with AI Builder that pulls the invoice ID, date, total, and line items straight from the email, scores its own confidence, and routes the result to your ERP for review or posting.

    TL;DR:

    • AI Builder accepts PDF, JPEG, and PNG files up to 20 MB and processes only the first 2,000 PDF pages; larger files need cloud links.
    • Route low confidence results, vendor mismatches, duplicate invoice numbers, and purchase order exceptions to human review instead of allowing automatic posting.
    • Test 50 to 100 representative invoices, then track touchless rate, processing time, error rate, and time to ERP before expanding beyond the pilot.
    • Skewed or low contrast scans and PDFs containing multiple invoices reduce extraction reliability, so request separate files and check scan quality before ingestion.
    Ampwise
    ampwise.ai
    Reduce Invoice Entry Work
    Ampwise processes invoices from Outlook or Gmail and sends data into business systems, helping reduce manual entry and related errors.
    Explore Ampwise

    Table of Contents

    How to extract a PDF invoice from Outlook on desktop, web, and mobile

    Before any automation can touch an invoice, you need a clean PDF sitting somewhere your workflow can reach it. The steps differ slightly depending on where you’re working:

    1. Desktop Outlook: right-click the attachment in the reading pane and choose “Save As,” or select the file and use “Save All Attachments” when an email carries more than one invoice.
    2. Outlook on the web: hover over the attachment thumbnail, click the download icon, and choose a destination folder, ideally one already synced to OneDrive or SharePoint.
    3. Outlook mobile: tap the attachment, select the share icon, and save to OneDrive rather than local device storage, since mobile files are harder to route into a flow later.
    4. Large or multiple attachments: replace the file with a OneDrive or SharePoint link in the email body before it’s forwarded, which sidesteps mailbox size limits entirely.

    If you’re scanning paper invoices instead of receiving PDFs directly, keep one invoice per file, scan at a resolution that keeps text legible, and avoid cramming multiple invoices onto a single page. A model extracting data from one scanned invoice per file performs far more reliably than one trying to split invoices out of a batch scan.

    Setting up Power Automate and AI Builder to extract invoice data

    Once PDFs are landing somewhere consistent, the actual automation is a short flow you can build in an afternoon.

    1. Choose your trigger. Use “When a new email arrives” for a personal mailbox, or point the trigger at a shared mailbox folder so the flow watches a centralized intake point instead of one person’s inbox.
    2. Add the AI Builder action. Insert “Extract information from invoices” and pass the email attachment as the file input. This is the prebuilt invoice processing model built specifically for this flow pattern: trigger, extract, map outputs.
    3. Map the returned fields. The action returns invoice number, invoice date, total amount, line items, and detected text, each with its own confidence score you can use to decide whether a record needs human review.
    4. Use the Pages parameter for long PDFs. When a vendor sends a multi-page statement with the actual invoice on page one, set the Pages parameter so the model only processes the pages that matter, which keeps costs down and accuracy up.
    5. Send the output somewhere useful. Write the extracted fields to Excel, SharePoint, or Dataverse, or call your ERP’s connector directly to create a pending invoice awaiting approval.

    Pro Tip: Run the flow against 15 to 20 real invoices from your most common vendors before trusting it with live volume. You’ll spot formatting quirks (odd date formats, non-English line items) that are easy to handle once you see them and painful to debug after the fact.

    Confidence scores matter more than people expect. A flow that treats every extraction as equally reliable will eventually post a wrong total straight into your ERP. Building a simple branch, where anything below your chosen confidence threshold routes to a review queue instead of auto-posting, catches most of the errors before they become accounting problems.

    File limits and mailbox setup that determine whether invoices get processed

    A flow that works perfectly in testing can quietly skip invoices in production if nobody accounts for format and size limits ahead of time.

    • The AI Builder prebuilt invoice model accepts PDF, JPEG, and PNG files up to 20 MB, and processes only the first 2,000 pages of a PDF.
    • Files larger than Outlook’s default send limits should travel as OneDrive or SharePoint links; Microsoft Graph supports attaching files using upload sessions for large files above common email size thresholds.
    • A shared, monitored inbox owned by accounts payable, rather than an individual’s personal mailbox, centralizes incoming vendor invoices and avoids the problem of one employee accidentally archiving or deleting a message a flow depends on.
    • Keep attachment counts per email reasonable. Business Central’s Payables Agent processes up to ten PDF attachments per message, creating one inbound document per PDF.
    • Anything that isn’t a recognized invoice format, a password-protected PDF, a corrupted scan, or an unexpected file type, should be flagged automatically and routed to manual review rather than silently dropped.

    Two end-to-end recipes: Outlook PDF to extracted fields to ERP

    Two patterns cover most of what AP teams need, and both start the same way: a monitored mailbox feeding a Power Automate flow.

    1. Shared mailbox to pending invoice. The flow watches a shared AP inbox, triggers on new mail with PDF attachments, calls AI Builder to extract invoice number, date, vendor, total, and line items, then writes the result to Dataverse or SharePoint before calling your ERP connector to create a pending invoice. Vendor identification against your supplier master list happens at this stage, so a new or misspelled vendor name gets flagged rather than silently matched to the wrong account.
    2. PO-matching flow. This variant extracts a purchase order reference from the invoice text alongside the standard fields, then checks it against open POs for a two-way or three-way match. Invoices that match cleanly post automatically; invoices with a quantity or price mismatch, or a missing PO reference, surface as exceptions for a human to resolve.

    Exception handling is where most of the practical value sits. Low-confidence extractions route to a review queue instead of posting directly. Duplicate detection compares invoice number and vendor combinations before anything new gets created, which catches the common case of a vendor accidentally sending the same PDF twice. And because the flow processes one attachment as one invoice record, a single email with three PDFs correctly generates three separate entries instead of one garbled one.

    Once the flow is live, keep an eye on daily processing counts and any invoices that got skipped rather than processed, since both numbers tell you whether volume is approaching a limit you haven’t planned for yet.

    Comparing AI models and tools for processing invoices from Outlook

    The AI Builder prebuilt model is the fastest starting point because it lives inside Power Automate already and requires no training data, but it works best as a general-purpose layer rather than a specialist for one vendor’s unusual invoice layout. Teams with a handful of high-volume suppliers often train a custom AI Builder model on those specific layouts and fall back to the prebuilt model for everything else, which keeps accuracy high without building a model for every vendor.

    Business Central users have a more integrated option in Payables Agent, which monitors a mailbox directly, runs OCR through Azure Document Intelligence, and creates draft purchase invoices without a separate Power Automate flow to maintain. That tighter integration comes at the cost of being tied to the Business Central ecosystem specifically.

    Outside Microsoft’s own stack, dedicated invoice automation platforms trade some setup flexibility for broader document support, including free-text emails that never had a PDF attached at all and scanned invoices with inconsistent layouts. For teams whose invoices arrive in a genuinely wide range of formats (PDFs, scans, and free-text requests mixed together) without standardized templates, a platform built to read varied document types without retraining can get you further faster than piecing together a one-off flow. The right choice depends mostly on how uniform your invoice formats already are and how much ongoing flow maintenance your team can absorb.

    Common problems in automated invoice processing and how to fix them

    Most automation failures trace back to a handful of repeat issues. Scanned invoices with skewed pages or low contrast reduce OCR accuracy more than any other single factor, so a quick scan-quality check before ingestion saves far more review time than tuning the model afterward. Multi-invoice PDFs, where a vendor bundles several invoices into one file, confuse field extraction because the model has no reliable way to know where one invoice ends and the next begins; splitting these at the source, or asking the vendor to send them separately, is usually faster than building logic to split them downstream.

    Inconsistent vendor formatting is another recurring headache. A supplier that changes its invoice template without notice can suddenly tank extraction accuracy for that vendor specifically, which is exactly the scenario where confidence-score routing earns its keep: the flow flags the drop instead of posting bad data silently. Free-text invoice requests embedded in an email body rather than attached as a PDF need a different extraction approach entirely, since there’s no document for AI Builder’s invoice model to read, and handling them usually means a separate text-parsing step or routing them to manual entry.

    Finally, watch for silent volume limits. A flow that works fine at ten invoices a day can hit page, size, or run-count limits at five hundred, so checking actual throughput against the documented file and page limits before scaling up avoids a nasty surprise during a busy month.

    Keeping invoice data secure and compliant in Outlook workflows

    Invoices carry banker details, tax identifiers, and vendor contact information, which makes the mailbox they land in a target worth protecting deliberately rather than by default. Restrict the shared inbox used for invoice intake to the accounts payable staff and the service accounts the flow actually needs, and never share that mailbox address externally beyond the vendors who need to send to it.

    Attachments should land in SharePoint or Dataverse with the same access controls your organization already applies to financial records, rather than sitting indefinitely in an inbox where retention policies are easy to overlook. Keep an audit trail of what was extracted, when, and by which flow run, since that trail is what lets you reconstruct a decision if a posted invoice later turns out to be wrong.

    Data residency and retention requirements vary by country and industry, so confirm where your organization’s Power Platform environment stores data and how long it retains processed documents before relying on default settings for anything regulated. A quick review with whoever owns compliance at your organization is worth doing once, early, rather than after an auditor asks.

    Mapping extracted invoice data into your ERP system

    Pulling clean fields out of a PDF is only half the job: the harder part is getting those fields into an ERP in a shape the system actually trusts. Start by mapping extracted vendor names against your supplier master list rather than creating a new vendor record automatically, since a slight name variation (“Acme Corp” versus “Acme Corporation”) can otherwise fragment one vendor into two in your books.

    Line items deserve their own mapping logic separate from header fields like total and date, particularly when your ERP expects items tied to specific GL codes or cost centers. A flow that maps the invoice total correctly but dumps line items into a single generic field creates more manual cleanup than it saves.

    Build validation rules before data ever reaches the ERP: does the extracted total match the sum of line items, does the invoice date fall within a plausible range, does the vendor have an open PO if your process requires one. Catching these mismatches in Power Automate, before the ERP connector fires, keeps your financial records cleaner than catching them after the fact. For the field-level detail behind this kind of mapping, our guide on invoice data extraction walks through accuracy benchmarks worth targeting before scaling a flow into production.

    Pilot first, measure honestly, then scale

    Before trusting any invoice automation with your full volume, run it against 50 to 100 representative PDFs from your actual vendor mix. That sample size is small enough to review by hand and large enough to expose the formatting quirks that matter.

    Track four numbers: touchless rate, average processing time per invoice, error rate, and time-to-ERP. Most teams see a clear signal within a couple of months, not a full fiscal year. If your invoices arrive in wildly inconsistent formats or your volume is already high, a managed solution built for varied document types may get you to a working pilot faster than a flow you maintain yourself.

    — Evert

    Ampwise AI: a managed option for Outlook-to-ERP automation

    If building and maintaining your own Power Automate flow isn’t where you want to spend engineering time, a managed IT services with a cybersecurity-first approach platform can handle the same problem without the setup. Such platforms read PDFs, scanned documents, and free-text invoice emails directly inside Outlook or Gmail, with no templates and no changes to how vendors send invoices, and connect straight to your existing ERP.

    Ampwise

    • The platform extracts invoice fields and routes them to your ERP without requiring a standardized vendor format.
    • It handles scanned documents and free-text requests alongside standard PDF attachments.
    • It integrates with your existing ERP system rather than requiring changes to existing workflows.
    Pilot element What to expect
    Suggested pilot length 2 to 4 weeks
    Suggested sample size 50 to 100 invoices
    What to measure Touchless rate and time-to-posting

    For a structured starting point, our pilot guide for AP teams walks through the same 2 to 4 week timeline we recommend to new customers. If you’d like to see how Ampwise AI handles your own invoice mix, request a demo through our product page and we’ll help you scope a pilot around your actual vendor volume.

    FAQ

    How do I create a PDF invoice template?

    Most accounting and invoicing software, including Excel, lets you build a reusable layout with your company details, line-item table, and totals, which you then export as a PDF for each new invoice. If you’re issuing invoices regularly, a dedicated invoicing tool or your ERP’s built-in template is usually more reliable than rebuilding one from scratch in a document editor.

    How do I extract a PDF file from an Outlook email?

    In desktop Outlook, right-click the attachment and choose “Save As”; on the web, click the download icon on the attachment thumbnail; on mobile, tap the attachment and share it to OneDrive. For files too large to attach directly, ask the sender to share a OneDrive or SharePoint link instead.

    How do I make an Excel invoice?

    Build a simple table in Excel with columns for item description, quantity, unit price, and line total, add a formula to sum the totals, and include your business details and the vendor’s or customer’s information at the top. Save a blank copy as a template so you’re not rebuilding the structure for every new invoice.

    How do I email a PDF file in Outlook?

    Compose a new message, select the attach icon, choose “Browse This PC” or “Browse Web Locations” if the file lives in OneDrive, and select your PDF. For files near or over standard size limits, attach a OneDrive link instead, since Outlook on the web’s default send size is 35 MB and some clients enforce lower limits.

    What file formats and sizes does automated invoice extraction support?

    The AI Builder prebuilt invoice model accepts PDF, JPEG, and PNG files up to 20 MB, processing up to the first 2,000 pages of a PDF. Files larger than that should be shared as a OneDrive or SharePoint link rather than a direct attachment.

    📅 Book an Ampwise AI demoBack to blog