Red Pill Realities: The Truth About Data Extraction and Verification with the Payables Agent

Datahaven4Dynamics
Sep 30
7 min read

Updated: Oct 20

About the Blog Series:

Microsoft’s blue pill version of the Payables Agent looks convincing. But if you’ve worked with Business Central (and NAV before that) long enough, you know Microsoft works hard to establish a narrative when releasing new software. Sometimes the narrative holds up. Sometimes not so much.

This blog series steps outside the Microsoft narrative for the Payables Agent. Each post uncovers one of the red pill realities that you will discover if you were to use the Payables Agent in the real world.

You can’t rely on consistent recognition of vendors by the Payables Agent

The Payables Agent relies on exact text matching to identify vendors, which often causes recognition failures. It compares the vendor name and address extracted from the invoice against Business Central records and rejects anything that does not match precisely. Even minor formatting differences such as “Inc.” versus “Incorporated,” punctuation changes, or missing address elements can prevent a match. Image-based headers or logos without searchable text also fail to identify the vendor.

💡 The practical way to think about the issue

Expect to manage vendor identification manually for a significant portion of your invoices. The Payables Agent performs well when invoice data exactly matches your vendor records but struggles when vendors use alternate naming conventions or inconsistent formats. Maintaining clean, standardized vendor master data will reduce errors but will not eliminate them.

⚙️ The technical reason(s) underlying the issue

Vendor recognition depends on deterministic string matching between extracted text and Business Central’s vendor table. The AI model does not apply fuzzy logic, partial matching, or context-based recognition. If either the vendor name or address deviates from the stored record, the system cannot confirm a match. For scanned or image-based invoices, missing OCR text compounds the problem.

📉 The impact on AP operations

Inconsistent vendor recognition adds manual work and delays to invoice creation. AP users must select the correct vendor manually or create duplicates when the system fails to match. Over time, this undermines the accuracy of the vendor master file, increases the risk of duplicate vendors, and adds friction to approval and posting workflows.

Interacting with the expanded “chat” feature of the Payables Agent

When the Payables Agent fails to identify a vendor or encounters another issue, users can interact with it through a “chat” feature in the Copilot Task pane. In theory, this allows AP staff to guide the AI using natural language. In practice, it introduces another layer of complexity, as users must learn a specific syntax to communicate effectively.

💡 The practical way to think about the issue

The chat feature is better viewed as a technical tool for power users, not an intuitive solution for typical AP clerks. It can be helpful in specific cases, such as telling the system to create a vendor from OCR data, but it requires practice and precision. Most AP teams will find it faster to correct issues manually rather than spending time crafting the right phrasing to “instruct” The Payables Agent.

⚙️ The technical reason(s) underlying the issue

The chat interface relies on Copilot’s task framework, which interprets typed instructions as structured actions within The Payables Agent’s workflow. The AI can only act on instructions relevant to the current step and context. Commands that fall outside the scope of that step, such as altering unrelated records or skipping tasks, are ignored. This makes the feature sensitive to wording and context, limiting its usefulness in routine AP work.

📉 The impact on AP operations

Rather than simplifying problem resolution, the chat interface often slows it down. AP clerks must remember phrasing rules and test instructions through trial and error. The learning curve reduces adoption and increases reliance on manual corrections. What should be a quick fix becomes a training and usability issue that diminishes the practical value of the Payables Agent in high-volume environments.

The data extraction results of the Payables Agent can be highly variable

The Payables Agent’s ability to extract invoice data is inconsistent and heavily dependent on the quality and structure of each document. It performs reasonably well on clean, digital PDFs but struggles with multi-page invoices, poor-quality scans, and documents containing complex tables or dense line items. When data is missed or misread, users must manually correct it in Business Central, leaving little of the process truly automated.

💡 The practical way to think about the issue

You will have to think of the Payables Agent as being limited to creating initial drafts of invoices. It can populate fields, but it cannot ensure accuracy. AP staff must manually confirm every key field and re-enter any data the AI misreads or skips. Because the system doesn’t highlight where extracted values came from or allow corrections directly on the invoice image, users must toggle between the PDF viewer and the Business Central form, effectively redoing the work the AI was meant to eliminate.

⚙️ The technical reason(s) underlying the issue

Data extraction depends entirely on Azure Document Intelligence, which uses OCR and layout analysis. Its model is pre-trained and not tenant-specific, meaning users cannot retrain or extend it. Extracted fields are limited to Microsoft’s schema and cannot include custom compliance or reporting fields. The built-in PDF viewer is read-only so it does not support annotations, click-to-correct, or incorporate a human-involved feedback loop. Microsoft intentionally disabled learning at the data extraction level, restricting AI refinement to coding suggestions rather than improving how invoices are read.

📉 The impact on AP operations

The combination of variable accuracy and limited correction tools keeps human effort high. Each extraction error adds manual re-entry, increasing fatigue, error risk, and verification time. Over time, these limitations flatten efficiency gains and discourage adoption. What should be a touchless process becomes a manual quality-control loop. In other words, AP teams must repeat the data correction step for every invoice.

You can’t ensure the accuracy of header and line data without significant human involvement

Despite its AI foundation, the Payables Agent’s data verification process depends on users to verify nearly every result. It uses Microsoft’s Azure Document Intelligence service to read invoices, but it does not provide an interactive review or correction step before creating a draft purchase invoice. This design shifts the responsibility for data accuracy to the user without offering efficient tools for validation.

💡 The practical way to think about the issue

You will have to think of the Payables Agent as a draft invoice generator, not an autonomous data extraction tool. It can speed up data entry but cannot ensure accuracy without a heavy dose of human participation. Users must manually confirm every extracted field against the original invoice image. Without side-by-side viewing or field highlighting, verification becomes a slow, error-prone process that reduces the efficiency gains the AI is supposed to deliver.

⚙️ The technical reason(s) underlying the issue

The Payables Agent sends PDF invoices to Azure Document Intelligence for text recognition and data extraction. Once the service returns the results, the system immediately generates a draft purchase invoice. There is no intermediary step for interactive review or data correction. The built-in PDF viewer displays only the document image, without visual cues showing where the extracted values originated. The viewer cannot be detached from the page, making true side-by-side comparison impossible.

📉 The impact on AP operations

AP teams could spend more time verifying and fixing extraction errors than they save from automated entry. For invoices with many lines or complex layouts, verification becomes a bottleneck. Missed or misread values must be corrected manually, and because the AI does not learn from these edits, the same issues recur. The result is a process that looks automated but still functions like manual data entry at scale.

The Payables Agent can’t “learn” from user corrections to extracted data

When the Payables Agent misreads data on an invoice, your corrections don’t teach it anything. Whether the error occurs in a header field or at the line level, each fix applies only to that one draft. The next invoice from the same vendor can repeat the same mistakes, requiring the same manual corrections all over again.

💡 The practical way to think about the issue

You will have to think of every correction as verification, not training. The Payables Agent still depends on humans to review, validate, and fix extraction errors. Until Microsoft introduces a true feedback loop that lets the system learn from user edits, accuracy issues will persist. The AI can improve its account-coding suggestions based on historical postings, but it will not get better at reading documents itself.

⚙️ The technical reason(s) underlying the issue

Corrections to extracted data—like adjusting totals, fixing tax amounts, or correcting vendor names—are not captured as training input. Extraction accuracy is governed entirely by Microsoft’s global Azure Document Intelligence model and by the quality of each source document. Because the model is shared across all tenants, it does not adapt to your company’s specific layouts or recurring corrections.

📉 The impact on AP operations

Without learning from corrections, your touch rate never declines. The same errors resurface repeatedly, adding friction and rework. AP staff must keep verifying the same fields, driving up both manual effort and Copilot credit consumption. Instead of compounding efficiency, the process stalls at a steady level of oversight and exception handling.

The Result:

Each red pill reality in this blog series exposes the limitations of the Payables Agent. We offer Datahaven 365 to help you unify documents, data, and workflows throughout Business Central—not just AP. The result is complete visibility, faster approvals, and automation that scales with the complexity of your business.

Check out the other Blogs in this Series:

About the Author

About Brad Bimson As the CEO of Datahaven 365, Brad is passionate about helping companies streamline their processes leading to savings in time, costs, and improved customer service. Brad wants to connect w/ Microsoft Dynamics users, partners & ISVs who want to leverage Dynamics.