Asurion Took AP From 20% to 70% Touchless Without New OCR

If your accounts payable team is shopping for a better OCR engine, check where your invoices actually stop first. Asurion did, found the problem was in the step after extraction, and raised touchless processing from below 20% to more than 70% across North and South America without replacing its capture tool or its ERP.

That is the result in Microsoft's Asurion customer story, published October 8, 2026. Asurion's finance team runs Dynamics 365 Finance with Microsoft's invoice capture front end. Its fix was a Copilot Studio agent that sits in the error state of invoice capture and repairs the item-number and purchase-order mismatches that were sending nearly every invoice to a person. The number is a vendor-published customer story, so read it as Microsoft's account of Asurion's results, but it names volumes, exception rates and a week-one figure, which most AP case studies leave out.

Why Asurion's Invoices Failed Before the Agent

Asurion's invoices failed at matching, after the OCR had already read them. According to the Microsoft story, the company handles about 30,000 direct purchase-order invoices a year, and roughly 90% of them carried at least one exception. Another 25,000 indirect invoices a year had header-level issues more than 75% of the time.

The root cause was item mapping. Vendors sent their own external item numbers, and those did not map cleanly to Asurion's internal item structure. One OEM part could correspond to different internal SKUs depending on which vendor sent it. Some invoices ran to hundreds of lines, and analysts fixed errors in the capture tool and then repeated some of the same updates inside the finance system.

For scale, compare the exception rate with the market. Corpay's summary of APQC benchmarking puts the average invoice exception rate at 22% and top performers at 9%. Asurion was at 90% on PO invoices. No extraction model fixes that, because the extracted value was correct. It just did not match anything in the ERP.

Microsoft's own documentation shows why this happens. In the invoice capture workspace docs, the item number is derived from three inputs: the derived legal entity, the derived vendor account, and the item description or external item number. "If derivation of a field fails, the process stops." The same page says the shipped derivation logic "includes the most common scenarios" and that "more specific validations can be implemented through customer extension." Asurion's agent is that customer extension, built with a low-code agent instead of X++ code.

---

What Asurion Built, Step by Step

The design is narrow: an agent called only when an invoice lands in an error state. The Microsoft story describes the flow. When invoice data enters an error state, Power Automate extracts the relevant details and passes them to an "Invoice Capture Derivation" agent built in Copilot Studio. The agent applies a defined set of instructions: it trims or adjusts item numbers, compares them against the specific purchase order to find the right item, and modifies PO numbers when a vendor has added extra characters that block the match. A second flow then reprocesses the invoice so it can move into Dynamics 365 Finance.

Three choices in that design are worth copying.

1. Invoices that derive cleanly never reach the agent. The model only sees exceptions, which keeps both the credit bill and the blast radius small.
2. The instructions are deterministic rules in plain language: trim this, compare against that PO, strip that suffix. The model applies business rules your AP leads already know and is never asked whether an invoice should be paid.
3. After the agent's fix, the invoice goes back through the capture tool's derivation and validation, so the agent's edit faces the same checks a clerk's edit would.

Asurion also said it evaluated third-party OCR alternatives first and concluded the mapping, derivation and exception problems would follow it to any product. So it kept its existing capture investment. Teams built working proof points in about a week, covering four of the biggest adoption blockers, per the story.

Don Baldwin, Senior Director of Software Engineering at Asurion, put the result plainly in the story: "Embedding the agent let us resolve invoice errors before they ever hit finance."

How Fast the Numbers Moved, and What They Cover

Touchless processing for North and South American vendors passed 55% in the first week in production and later exceeded 70%, according to Microsoft. Asurion's stated global goal is roughly 80% touchless, and the story describes a path toward 85% or more as its MCP server work advances from QA into production.

Read the scope carefully before you put this number in a business case:

• The baseline is loose. The story's summary says touchless processing started "below 20%", but the one body sentence that gives the starting point puts it at "less than 20% touchless processing for two vendors." The before and after figures may not cover the same set of vendors.
• The 70% covers the Americas only. Extending the pattern to APAC and Europe is listed as next work, along with VAT tax-to-charge-code conversion and unit conversion, which are the cases most likely to break item and charge mapping.
• The 85% is a projection that depends on work still in a QA environment as of the October 8 story.
• The story gives no definition of "touchless." It does not say whether it counts invoices with zero human touches end to end, or invoices that reached Dynamics 365 Finance without a capture-side review.
• No error rate for the agent is published. The story reports how many invoices went through untouched, not how many the agent fixed wrongly.

For a benchmark, the same Corpay summary cites The Hackett Group's 2025 research putting the touchless rate of top performers at 49.2%. Definitions differ across studies, so treat the comparison loosely, but a 70% rate on a book of invoices that used to fail 90% of the time is a large move.

---

The MCP Piece Still in QA

The remaining gap to 85% sits on MCP, and the product that work depends on is already generally available. Asurion is exploring MCP server capabilities in a QA environment while it validates security safeguards and large-file handling, according to the story. The Model Context Protocol is an open standard that lets an agent discover and call tools exposed by another system. In this case, that system is the ERP itself.

Microsoft's Dynamics 365 ERP MCP server documentation describes data tools for create, read, update and delete operations, form tools that let an agent open forms, set values and press buttons, and action tools that call custom classes. It requires version 10.0.47, 10.0.46 PQU-2 or 10.0.45 PQU-7, and it does not run on Cloud Hosted Environments. The older static MCP server, with 13 fixed tools, was scheduled for retirement on October 1, 2026, so anyone who piloted on it has already had to migrate.

Two details in that documentation explain why Asurion is moving slowly:

• The agent inherits a security role: "The security role of the authenticated user for the agent determines which objects are returned." An invoice agent with a broad finance role can edit far more than invoices. Asurion's story says agents need the right permissions, "read-only access where possible," and human review in the right places.
• Attachments are a listed limitation. The MCP server "doesn't support the document viewer, DocuUpload, and FileUpload controls," and Asurion's near-term roadmap includes large invoice PDF processing, while its QA work covers large-file handling.

On the agent side, Copilot Studio's MCP page says it supports MCP tools and resources and requires generative orchestration to be turned on. That is a change from the deterministic rule-following agent Asurion runs today, and it is the step where an AP leader should ask for evaluation results before go-live.

What It Costs, and Where Microsoft Is Catching Up

Microsoft does not publish what Asurion pays, so budget from the rate card. Copilot Studio sells capacity at $200 per pack per month for 25,000 Copilot Credits. The billing rates page charges 5 credits per agent action and 13 credits per 100 agent flow actions, and when the agent calls the Dynamics 365 ERP MCP server from Copilot Studio, the MCP documentation says each tool call bills as an agent action with no separate server charge. The Microsoft 365 Copilot inclusion on the billing page covers an authenticated, licensed person using the agent. An agent triggered by a flow on an exception queue has no such person, so plan to pay for it.

Here is an illustrative estimate; the per-invoice action count is an assumption, not an Asurion figure. At Asurion's reported volumes, about 27,000 PO invoices and roughly 19,000 indirect invoices a year hit an exception. If each one takes ten agent actions, that is around 2.3 million credits a year, or eight packs a month at list price: about $1,600 a month. Power Automate cloud flows sit on Power Automate licensing, not Copilot Credits, per the same billing page, and invoice capture itself draws on Electronic Invoicing credits, according to Microsoft's invoice capture FAQ. Set that against APQC's spread of $10.89 per invoice on average versus $2.78 for top performers, and the agent is unlikely to be the line item that decides the case.

The steel-man for waiting is that Microsoft is shipping some of this natively. A Message Center notice, MC1465507, announced AI-based invoice header derivation in public preview from September 15, 2026 in version 10.0.49, matching headers to purchase orders and procurement categories. Line-level derivation is listed as future work. Asurion's hardest problem was at the line level (external item numbers on invoices with hundreds of lines), so a native header feature would not have closed its gap. If your exceptions are mostly header problems, like Asurion's 25,000 indirect invoices, test the preview before you build.

The FAQ also says continuous learning derives legal entity, vendor and items from the last transferred invoice, which "can increase the touchless rate." Turn that on first, since it learns from the fixes your clerks already make.

---

What to Do Before You Buy Another Capture Tool

This Week:

1. Pull last quarter's exception log from your capture tool and sort it by reason code. If item or PO mismatches are more than half, a new OCR vendor will not move your touchless rate.
2. Ask your AP lead to write down, in plain sentences, the three fixes clerks make most often. Those sentences are the first draft of the agent's instructions.
3. Check whether continuous learning is on in invoice capture and whether your environment is on 10.0.47 or later.

This Month:

1. Build one exception agent for the single most common mismatch, triggered only from the error state, with the invoice going back through standard derivation and validation afterwards.
2. Give it a dedicated identity with the narrowest finance role that can do the job. Have your controller sign off on the role before the agent touches production.
3. Define "touchless" in writing before you measure it, and add a sampled audit of agent-fixed invoices so you track a wrong-fix rate next to the touchless rate.

Before Your Next Capture Renewal:

1. Run the MC1465507 header derivation preview against your indirect invoice book and compare it with what a custom agent would do.
2. Set a per-agent monthly credit cap in the Power Platform admin center, which the billing page says you can do, so a vendor sending a bad batch cannot drain your tenant's capacity. At 125% of prepaid capacity, custom agents are disabled.

The Bottom Line

AP automation has been sold on extraction accuracy since the first OCR rollouts, and Asurion's numbers show the touchless rate was set by what happened after extraction. Microsoft's own docs left that step to "customer extension," which used to mean a development project. Asurion wrote the rules its clerks already applied as instructions to an agent and had proof points in about a week. Its next targets, according to the story, are journal entry review, month-end reconciliation and cash application with a human in the loop. Each has its own exception queue, and the first question for each is which exceptions people fix the same way every time.

Pull the exception log before you take the next OCR demo.

Continue Reading

• Textract vs Azure vs Gemini: Split OCR From Extraction
• Human-in-the-Loop for AI Agents: Most Approval Gates Rubber-Stamp
• Microsoft Hit 20,000 Agent Makers on a Risk Rubric It Didn't Publish
• MCP Server Governance: Enforce at the Client, Not the Registry
• Microsoft Work IQ Learns Dynamics 365 From Views You Never Pruned
• Cisco's CFO Automated 80% of Reporting, Then Deployed AI to 90,000 Staff