Reading Faxed Purchase Orders With AI Builder — A Realistic Design for Cutting Manual Data Entry, and Where It Stops

“Forty percent of our orders still come in by fax. We look at the purchase order coming out of the multifunction printer and key it into the sales management system one at a time.” We hear this constantly in consultations with manufacturers and wholesalers. What usually follows is: “We aren’t in a position to ask our trading partners to switch to web ordering,” and “the bigger the partner, the more their ordering system assumes a fax is being sent, and the less we can move them.”

There are broadly two routes to improving fax-based order intake. One is to migrate to web ordering or CSV import and “stop using fax.” The other is to keep receiving faxes but automate the reading and the data entry — “living with the fax you cannot get rid of.” The former was covered in a separate article, “Moving FAX Orders to the Web — Designing the Dual-Operation Period and a Phased Migration.” This article is about the latter: receiving faxes as PDF data, reading them with AI Builder document processing, inserting a human check, and connecting the result to an order ledger and the core system.

Let me say up front that this mechanism will not be “fully automatic,” and it should not be. Even so, if you can compress one to two hours of daily data entry into a few dozen minutes of “just checking the extraction results,” it will be worth the investment for many companies. Below I set out how far you can automate and where the limits begin, based on specifications you can verify on Microsoft Learn.

1. The Bottom Line First

• The precondition for automating fax reading is receiving faxes as data (PDF), not as paper. It starts with turning received faxes into files using an MFP’s fax forwarding feature or a cloud fax service, and consolidating them into SharePoint.
• AI Builder’s document processing custom models can extract fields and tables from forms with bespoke layouts such as purchase orders. Training is done per collection — a group of forms sharing the same layout — and each collection needs at least 5 samples (up to 20).¹²
• Both training types — structured documents and unstructured documents — support Japanese, and the FAQ explicitly states that handwritten text extraction is supported too. How much of a real fax-quality form can actually be read, however, must always be verified against your own samples.³²
• The crux of the flow is triage by confidence score. Each extracted field carries a score between 0 and 1, so build a branch that sends high-confidence results to the ledger automatically and low-confidence ones to a human. Do not abandon the assumption that a person reviews everything — design for “reviewing gets easier.”⁴
• Do not try to read every trading partner’s forms. Because the mechanism separates collections by layout, you get results from just the standard purchase orders of your highest-volume partners. For low-quality forms, the official guidance is to increase the sample count to 15–20.²⁵
• AI Builder requires metered credits on top of everything else. Document processing consumes per page, and custom models carry a higher rate than prebuilt models. Following the phase-out of AI Builder credits announced in October 2025, the model is migrating to Copilot credits.⁶⁷
• Reading is not a cure-all. Depending on volume, partner count, and how standardized the forms are, migrating to web ordering or EDI may be the more fundamental answer (see the decision table in Section 8).

2. “Stop It” vs. “Read It” — Do Not Mix the Two Routes

When we take on a fax-order consultation, the first thing we check is: “Is this a partner you can stop faxing with?”

Migrating to web ordering is an initiative that asks your trading partners to change their behaviour. As set out in the phased-migration article, you move partners across in order of volume and system readiness, cutting manual entry where the impact is largest. Meanwhile, some partners simply cannot be moved. Their ordering process runs on the assumption of fax; their staff are older and cannot be asked to key things into a web form; or you are simply not in the position of power to ask. It is realistic to assume faxes from partners like these will remain for years.

That is where the “read it” route comes in. The key point is that these two routes are not mutually exclusive.

The biggest advantage of the “read it” route is that it demands no change whatsoever from your trading partners. It also functions as a way to reduce the processing cost of the fax channel during the dual-operation period of a phased migration. Its limits are equally clear: extraction accuracy will never reach 100%, so a review step remains permanently. And if you try to read every form from every partner with a different layout, you will burn yourself out maintaining models. That is exactly why you should think of it as a combination: move the partners you can move, and read the faxes from the ones who remain.

Processing purchase orders that arrive as PDF attachments by email is covered in “Automating Purchase-Order and Invoice PDFs That Arrive by Email With Power Automate — Designing Storage, Routing, Notification, and Extraction.” If you receive faxes via email forwarding, everything downstream of receipt is the point where this article and that one converge.

3. Prerequisite: Receive Faxes as “Data” — Nothing Starts While It Is Still Paper

What you can hand to AI Builder is a file. An operation where you re-scan paper printed by the MFP is not automation — it just swaps data entry for scanning. The first thing to do is create a path where received faxes are saved as files without a human touching them. There are two realistic options.

• The MFP’s fax forwarding feature. Most business MFPs can save a received fax to a designated folder without printing it (scan to folder), or forward it as an email attachment. Check the configuration with the model’s manual and your maintenance provider.
• A cloud fax service. Move the fax number itself to a cloud service and receive incoming faxes as PDFs by email or API. This digitizes receipt without replacing the MFP; whether number portability is possible and how the pricing works needs checking service by service.

Whichever path you take, we recommend landing everything in a SharePoint document library. The SharePoint connector has a “When a file is created (properties only)” trigger and a “Get file content” action, so the save itself becomes the starting point of the downstream extraction flow.⁸ If you receive by email forwarding, put a flow in front that receives into a shared mailbox and saves the PDF attachment to SharePoint (this design is exactly Sections 3 and 4 of the email attachment article).

One caveat about file formats. Depending on the MFP’s forwarding settings, faxes may be saved as TIFF. AI Builder document processing can only use PDF, JPG, and PNG for model training, although a trained model running in a cloud flow can process TIFF as well.³ Given the need to prepare training samples, though, it is simplest to select PDF output in the forwarding settings. While you are at it, note the other limits: files can be up to 20 MB, and images must be between 50×50 and 10,000×10,000 pixels.³

4. What AI Builder Document Processing Can Do

How Custom Model Training Works

AI Builder document processing is a custom AI model that you train, using sample forms, on “where on this form each piece of information sits.” When creating a model, you first choose a document type.¹

For a trading partner’s standard purchase order, the answer is basically structured documents. Next you define the information you want to extract. You can specify fields (order number, order date, partner name, delivery destination and so on), tables (line items with product, quantity, unit price), and checkboxes.¹

The unit of training is the collection. A collection is “a group of forms sharing the same layout,” so forms with different layouts — partner A’s purchase order versus partner B’s — go into separate collections. You upload at least 5 sample documents into each collection and tag the positions of the fields and tables to train it. A collection can hold up to 20 samples, and a single model can have up to 200 collections.¹³ The official FAQ notes that 5 is often sufficient for high-quality documents, while 15 to 20 is recommended for low-quality scans.²

Japanese and Handwriting Support

The points worth confirming for handling Japanese forms can be verified on Microsoft Learn as follows.

• Japanese is among the supported languages for both the structured document and unstructured document training types.³
• The official FAQ explicitly states that “document processing can extract printed text and handwritten text.”²

That said, let me be honest here. There is a gap between “it is supported” and “your own faxes can be read at production-grade accuracy.” Faxes are low resolution, and faded strokes, filled-in characters, and skew happen daily. Whether a handwritten quantity correction or a scribbled note can be read depends on how the form was filled in. Training and testing a document processing model is free,⁹ so before you decide to adopt it, always insert a step where you train and test on the PDFs of faxes you have actually received and look at the accuracy on your own forms. Training on clean original PDFs and then operating on faxes is a classic failure: the image quality differs so much between test and production that accuracy never materializes.

How It Differs From Prebuilt Models

AI Builder also has prebuilt models that need no training. The flagship is the invoice processing model, which extracts common fields such as invoice number, invoice date, and amount due with no training, and its supported languages include Japanese.¹⁰ For invoice processing, trying this first is the fast route.

There is, however, no prebuilt model for purchase orders. Forms like purchase orders, order confirmations, and delivery notes, whose layouts differ from company to company, are cited in the official FAQ as the classic example of a structured document custom model,² so the faxed purchase orders that are the subject of this article fall squarely in custom model territory.

5. Overall Flow Design — Insert a Human Check

The overall picture looks like this. The crux is the “human check” in the middle; a fully automatic configuration that omits it is not recommended, for reasons given below.

flowchart TD
    Fax[Fax received from a trading partner] --> Digitize[MFP fax forwarding / cloud fax service<br/>Delivered as PDF to a folder or mailbox]
    Digitize --> Save[Saved to a SharePoint document library<br/>Unique file name including the receipt timestamp]
    Save --> Trigger[Cloud flow starts<br/>When a file is created]
    Trigger --> AIB[AI Builder: Process documents<br/>Extract order number, partner, due date, line items]
    AIB --> Score{Confidence score check<br/>e.g. all key fields at 0.9 or above?}
    Score -- High confidence --> Ledger[Record in the order ledger list<br/>Status: auto-extracted, not yet confirmed]
    Score -- Low confidence --> Review[Send a review request to staff<br/>Teams notification + link to the original PDF]
    Review --> Fix[Staff compare with the original, correct and confirm<br/>Status: reviewed]
    Ledger --> Batch[Staff review and confirm the list in bulk<br/>Status: reviewed]
    Fix --> Entry
    Batch --> Entry[Enter into the core system — only rows with status reviewed<br/>CSV import / Power Automate for desktop]

The Extraction Step

In the flow, you pass the trained and published model plus the file content retrieved from SharePoint to the “Process documents” action (this action was renamed in May 2025 from “Extract information from documents”). The output includes the value of each field you defined ({field} value) and its confidence score ({field} confidence score), plus per-cell values and scores for tables. For multi-page documents, you can also specify the page range to process and hold down consumption.⁴

An implementation caveat: extracted values all come back as strings. To put quantities and amounts into numeric columns of your ledger, convert them with int/float expressions and strip currency symbols and whitespace with replace. Normalize dates with formatDateTime too.⁴ Normalizing full-width/half-width variation in fax numbers and order numbers here as well will make the downstream matching much easier.

The Review Step — This Is the Crux

There are two shapes that work well for reviewing extraction results.

1. Drive the review through a status column on the ledger list. Hold the order ledger as a SharePoint list, register extraction results with a status of “needs review,” and have staff compare them against the original PDF (attached or linked) on the list, correct them, and change the status to “reviewed.” The change to “reviewed” then triggers the downstream processing. Moving a ledger from Excel to a SharePoint list is covered in “Replacing an Excel Ledger With a SharePoint List.”
2. Drive the review through a Teams approval. Use the Approvals connector’s “Start and wait for an approval” action with custom responses, giving staff options like “register as is” or “needs correction.” How to build approvals and handle responses is covered in detail in “Building an Approval Workflow in Power Automate — Digitizing Paper and Email-Based Approval Requests.”¹¹

In either case, consolidate the notifications into a Teams channel and always attach a link to the original fax PDF.¹² The actual substance of reviewing is “comparing the extraction result against the original,” so if opening the original is fiddly, the review will degenerate into a rubber stamp.

From the Ledger to the Core System

How you get reviewed data into the core system depends on what entry points the core system offers. If it has a CSV import feature, formatting the reviewed data as CSV and importing it is the solid choice. If there is no import feature and only screen entry, UI automation with Power Automate for desktop becomes an option. That design is covered in “Automating Data Entry Into Legacy Core Systems With Power Automate for Desktop.” For error handling across the whole flow (notification on extraction failure, re-running), the thinking in “Power Automate Error Handling and Retry Design” applies directly.

6. Living With Accuracy — You Get Value Without Reading Everything

Sorting by Confidence Score

Once you accept that “extraction can be wrong,” the centre of the design becomes how you use confidence scores. Scores run from 0 to 1, with values closer to 1 indicating a higher probability that the extracted value is correct.⁴ In practice, you might split it like this.

• All key fields — order number, partner, due date, line items — at or above a threshold (say 0.9) → record in the ledger as “auto-extracted.” Staff check the list in bulk and only need to perform the confirmation action.
• Any one of them below the threshold → send to staff as “needs review.” Spell out the low-scoring fields in the notification and have them corrected against the original.

Note that even on the high-confidence side, you are not skipping the review itself. A score only indicates a high probability of correctness; it is not a guarantee of correctness. Misreads happen even at high confidence, so the gate that only rows a human has explicitly confirmed as “reviewed” are passed to the core system is maintained regardless of score. What the score is allowed to change is the weight of the review (confirm in bulk, or correct one by one), not whether a review happens.

Do not fix the threshold up front; adjust it after go-live while watching “how many items were treated as automatic but turned out to be wrong.” The important thing is that the cost of mistakes is asymmetric. Erring towards “needs review” only adds a little checking effort, whereas letting a misread through automatically leads to a wrong shipment. When in doubt, set the threshold high.

The Fate of Fax Image Quality

The document processing requirements say that “scans from paper should be high-quality images” and that “PDFs with embedded text (text PDFs) are preferable, since they avoid garbled characters and positional drift.”³⁵ Faxes are the polar opposite of that ideal. Even so, there are moves you can make.

• Use actual faxes as training samples. As noted above, training at the same image quality as production comes first. For low-quality images, the official guidance is to increase the number of samples to 10–15 or more.⁵
• Work on quality at the sending end. With your top partners, you can sometimes ask them to “please send in high-quality (fine) mode.” Even on a route that asks nothing of your trading partners, a single setting next to the send button is at the easy end of the scale to ask for.
• Do not force an unreadable form. Forms that are badly faded, or mostly handwritten, will end up permanently in the “needs review” pile. Sometimes the right call is to exclude that partner from extraction and keep keying them in as before.

The Problem of Layouts Differing by Trading Partner

There are as many purchase order layouts as there are trading partners. Document processing is designed to absorb this through collections: you split collections by layout and roll them into a single model (up to 200).³ But the more collections you have, the more maintenance you take on — collecting samples, tagging, verifying accuracy, and keeping up with layout changes. If a partner changes their order form, that collection has to be retrained.

This is exactly why the way to start is not “read everything” but “read only the standard purchase orders of your highest-volume partners.” If fax orders run at 600 a month and your top three partners account for 350, a model with 3 collections (5–20 samples each) covers just under 60% of the data entry. The rest can carry on being keyed in as before. This is the extraction-side version of the “move the highest-impact things first” principle from the phased-migration article, and the partner classification (A/B/C) used for web migration transfers over directly. Building the pattern in a pilot, measuring the benefit, and only then adding collections is what prevents model maintenance from grinding you down.

Some finer constraints are worth confirming too. Fields that span pages, and line items that wrap across a page boundary, are currently unsupported.³ For partners who often send multi-page purchase orders, the line-item design needs care.

7. Licensing and Cost — The Credit Consumption Model

AI Builder actions consume metered capacity on every run, separately from your Power Automate licensing. Start building without knowing this and you will run into NoCapacity-class errors in production.⁹

Here is how it works. Document processing consumes credits according to the number of pages processed (including pages with no data to extract).² Consumption rates differ per capability: the official rate table defines custom model document processing at 100 AI Builder credits per page, prebuilt capabilities such as invoice processing at 32 credits per page, and, under the Copilot credit model, content processing at 8 Copilot credits per page.⁶

Historically there were two routes to obtaining credits.⁹

• AI Builder capacity add-on: 1 million credits per add-on
• Seeded credits bundled with a Premium license: 5,000 credits per Power Automate Premium license

Credits are pooled at the tenant level, and administrators allocate them to environments. Doing the arithmetic, the 5,000 credits bundled with Premium cover 50 pages of custom model document processing. So a single-page purchase order at up to 50 a month fits within the bundled allowance, whereas 600 a month means you will need an add-on or Copilot credits — that is the order of magnitude.

This model is, however, in the middle of a transition. In October 2025 Microsoft announced the phase-out of AI Builder credits. From 1 November 2025, new customers cannot purchase the AI Builder capacity add-on, and on 1 November 2026 add-on renewals end and the seeded credits bundled with Premium licenses are discontinued. AI Builder’s functionality itself is not going away and remains available through Copilot credits. During the transition, the order of precedence is: AI Builder credits are consumed first, then Copilot credits once those run out, and execution is blocked if neither is available.⁷

In short: “extraction carries an additional cost proportional to how much you run it, and the currency it is charged in is changing.” This article does not quote specific figures; estimate consumption from your monthly fax volume × page count against the rate table, and check it alongside the latest pricing information. Power Automate’s own licensing (the boundary between standard and premium connectors) is set out in “Power Automate Licensing and the Boundary Between Standard and Premium Connectors.”

8. Which Route to Take — A Decision Table

AI Builder extraction is not the only answer to faxed purchase orders. Drawing the line by volume, number of trading partners, and how standardized the forms are gives roughly the following.

As a feel for the decision, the quickest measure is to count “how many sheets of the same layout arrive per month.” If there are a handful of partners with a large number here, extraction pays off. If the per-layout sheet count stays low while the number of partners is high, maintaining extraction models does not pay, and web migration, EDI, or simply continuing manual entry is the more rational choice.

One more thing not to forget is how you position the review step. The benefit of extraction automation is not that “data entry goes to zero” — it is that “data entry turns into reviewing.” When measuring the benefit, measure the pre-adoption entry time and the post-adoption review and correction time on the same yardstick. Tracking the review rate (the proportion falling below the threshold and going to a human) alongside the number of misreads that were treated as automatic gives you the material for adjusting thresholds and adding collections.

9. Summary

Automating faxed purchase orders can be designed as a route entirely separate from “stopping fax.” To recap:

First, turn received faxes into PDFs using the MFP’s forwarding feature or a cloud fax service, and consolidate them into SharePoint. Nothing starts while it is still paper. Next, train an AI Builder document processing custom model, narrowed to the standard purchase orders of your highest-volume partners. At least 5 samples per collection, more samples for fax quality, and train on faxes you actually received. Then sort automatic from needs-review by confidence score, and pass things to the ledger and the core system only after a human check. Not aiming for full automation is precisely what makes adoption faster and operation more stable.

On cost, estimating metered credit consumption and confirming the migration schedule from AI Builder credits to Copilot credits are both essential. And extraction remains, ultimately, a means of “living with the fax you cannot get rid of.” Move the partners who will cooperate to web ordering or CSV import, and lighten the remaining faxes with extraction. Combining the two routes is what makes the total volume of manual entry smallest.

If you would like to proceed while verifying how far you can get with your own forms — from digitizing fax receipt through designing the extraction flow, building out the review step and core-system integration, and combining it with a web migration — please get in touch through the consulting areas below.

Related Articles

• Moving FAX Orders to the Web — Designing the Dual-Operation Period and a Phased Migration
• Automating Purchase-Order and Invoice PDFs That Arrive by Email With Power Automate — Designing Storage, Routing, Notification, and Extraction
• What Is EDI? How It Streamlines Business-to-Business Ordering — From Fax, Email, and Manual Entry to Data Integration
• Automating Data Entry Into Legacy Core Systems With Power Automate for Desktop
• Can the Labour-Saving Investment Subsidy Pay for Moving FAX Orders to the Web? — How to Think About Order-System Investment Under the General Type

Related Consulting Areas

KomuraSoft LLC can advise on everything from designing extraction automation for forms arriving by fax and email, through building out the review flow, to integration with your existing sales management and core systems.

• Technical Consulting & Design Review
• Windows App Development
• Contact

References

Related Articles

Recent articles sharing the same tags. Deepen your understanding with closely related topics.

Related Topics

These topic pages place the article in a broader service and decision context.

Frequently Asked Questions

Common questions about the topic of this article.

Can AI Builder read Japanese fax purchase orders?

There is a fair chance it can. AI Builder's document processing custom models support Japanese under both training types — structured documents and unstructured documents — and the FAQ explicitly states that handwritten text extraction is supported as well. That said, faxes are low resolution and prone to faded strokes and skew, so how much can actually be read depends on your forms and your line quality. Before deciding to adopt it, we strongly recommend training a model on PDFs of faxes you have actually received, and checking the accuracy against your own forms.

How many sample documents does model training need?

You need at least 5 sample documents per "collection" — a grouping of forms that share the same layout. A single collection can hold at most 20, and while 5 is often enough for high-quality documents, 15 to 20 is recommended for low-quality scans such as faxes. If your trading partners each use a different purchase-order layout, create a separate collection per layout (up to 200 collections per model), and realistically start with just the standard order forms from the partners who send the most volume.

Can I feed the extracted results straight into our core system?

We do not recommend letting it flow through unattended. AI Builder attaches a confidence score between 0 and 1 to every field, so always insert a branch: high-scoring results go automatically into the order ledger, while low-scoring ones are sent to a member of staff who checks and corrects them against the original fax image. If you build it so that only reviewed data goes on to core-system entry — CSV import, or data entry via Power Automate for desktop — you prevent an extraction mistake from turning directly into a shipping mistake.

What does it cost to use AI Builder?

AI Builder actions consume metered credits every time they run. Document processing consumes credits according to the number of pages read, and custom models are set at a higher consumption rate than prebuilt models. Historically this was covered by AI Builder credits — either a capacity add-on or the 5,000 credits bundled with Power Automate Premium — but in October 2025 the phase-out of AI Builder credits was announced, and the migration to Copilot credits is under way. The seeded credits bundled with Premium end on 1 November 2026, so always confirm the current licensing model when adopting it.

Author Profile

Profile page for the article author.

Go Komura

Representative of KomuraSoft LLC

Focused on Windows software development, technical consulting, and investigations into failures that are difficult to reproduce.