Artificial Intelligence / Use Cases / Document Processing
Invoices, contracts, and forms processed without re-entering data
Extract the fields, verify them against your systems, and send them toERP or to Dynamics 365 — and keep a human where the decision matters.Every day, documents arrive—via email, mail, or a portal—and someone enters their contents into a software program. Intelligent document processing reads the document, extracts the data, verifies it against what your systems already know, and only requires human intervention for discrepancies. It operates in Microsoft 365 and Azure ; when documents are subject to regulatory requirements, the scanning can be performed on-site.
Three scenarios common in accounting, purchasing, and back-office operations
The document has arrived. Anything that happens between its arrival and when it is entered into the system is a waste of time and opens the door to potential errors.
The invoice arrives as a PDF, and someone types it out by hand
Three hundred invoices a month, each one opened, read, entered into theERP, and reconciled with the purchase order by hand. One error in fifty, discovered during bank reconciliation. And the person who knows how to do this is also the one who closes out the month.
Data entry is a job no one chose to do.
The contract has been signed, but the dates aren't listed anywhere
The renewal date, the notice period, indexation, the termination clause: it’s all in the PDF; none of it is in a system that sends notifications. The contract renews automatically, and you don’t find out until you receive the next bill.
A contract you don't read over again renews itself automatically.
The form is filled out by hand and then re-entered
A membership application, an insurance claim, a delivery slip signed on the loading dock, a pickup form: filled out by one person, photographed or scanned, then retyped by another. That’s twice the work, and the handwriting is never fully legible.
Paper isn't going away; retyping is.
What the solution does, in plain language
Six steps, in the order in which they occur. The fourth step is what distinguishes a reading tool from a sustainable process.
The system processes the document as soon as it arrives
An email, a file SharePoint, a scanner, a portal, a photo taken on the platform. Each channel feeds into the same processing system.
The system reads the document, regardless of its format
Invoices, contracts, forms, delivery slips: the document-reading services recognize the type of document and extract the relevant fields—amounts, dates, parties, lines, checked boxes, and signatures.
The system converts the image into data
A supplier, a number, an amount, a due date, and line items: named fields, each with its own confidence level. This confidence level determines the next steps.
The solution checks against what your systems already know
Does the supplier exist? Does the purchase order match? Does the total of the line items equal the total on the invoice? Is the date plausible? Any discrepancies are flagged, and the system then alerts a human.
The system only alerts a person about discrepancies
The document that meets the requirements is approved. The document with discrepancies is displayed to one person, with the discrepancy highlighted and the original document shown alongside it. The person corrects it or rejects it in a matter of seconds, rather than having to retype it.
The solution is integrated into your systems and keeps track of everything
The accounting entry, the contract record with its due dates, the customer file, and the reconciled delivery slip—all with the original document attached and a log of what was read, checked, and approved, and by whom.
A concrete, step-by-step example
A supplier invoice arrives at the ` factures@ ` address of an industrial SME that manages its accounting on a web ERP and its purchasing in Dynamics 365.
What Remains Human
The goal is not to eliminate approval: it is to require it only when it is necessary. Five decisions remain in the hands of appointed officials.
Decisions That Cannot Be Delegated
They are documented in the pre-commissioning procedures and reviewed during each maintenance check. A process that writes data to a “ ERP ” without verification is more dangerous than the manual data entry it replaces.
- Any discrepancy between the document and what the systems have on record—such as an unknown supplier, a total that doesn't match, or a missing order—is resolved by a person.
- Any document exceeding a certain threshold amount or classified as sensitive—such as a contract or an amendment—must be approved before being posted, regardless of what the audit says.
- The confidence thresholds below which a read field is presented to a human are determined by the business, and are tightened or relaxed depending on the errors observed.
- Rejecting a document—and sending a message to the supplier or customer—are still human actions.
- A monthly review of past errors and common discrepancies: that’s what improves the process—not the model alone.
The data used, and what they require
Here, two types of data intersect: incoming documents and the reference systems used to verify them. The quality of the latter determines the usefulness of the processing.
Sources
- Incoming documents: native PDFs, scans, photos, emails, and scanned paper forms.
- Channels: Shared Inboxes Exchange, folders SharePoint, scanners, portals, and mobile field applications.
- Reference documents: suppliers and customers in theERP or Dynamics 365, purchase orders, contracts, payment terms.
- Business rules: thresholds, approval workflows, delegations—often just in people’s heads, waiting to be written down.
What they must have
- Consistent reference sets: a supplier with three different spellings makes verification impossible. The reconciliation process begins with the reference set.
- Legible documents: A blurry photo, a scan at 100 dots per inch, or tightly packed handwriting are difficult to read. The quality threshold is determined based on the prototype, using your actual documents.
- A representative sample for the prototype: the last 100 documents received, not the 10 cleanest ones.
- A process owner who sets the thresholds and reviews errors each month.
- Retention: how long the original document is kept, and where—this is a legal issue rather than a technical one.
Where the solution works, and which one we would examine first
A use case that supports three architectures. The sensitivity and volume of the documents determine the choice, and the same processing chain can be moved without needing to be rewritten.
At Microsoft
RecommendedServices Azure document reading services, Power Automate for orchestration, Dataverse , orERP for writing. Billed per page, hardware-free, and up and running in a few weeks.
At your facility
PossibleThe same reading services run in containers on Azure locally, on your datacenter, when documents are subject to regulations or when the volume of data makes page-by-page reading too expensive.
Partly at Microsoft, partly at your office
PossibleOrchestration and validation in Power Platform ; viewing sensitive documents on your premises. The distinction depends on the type of document.
Next to the machine
Possible, at the dockScan delivery slips and labels on the warehouse loading dock, on-site, when the connection is too slow to wait for the call center.
Why online, and what makes people choose to visit your location
Three criteria are decisive for this use case. The third one is the most surprising: it has to do with volume, not confidentiality.
| Criterion | What tips the scales | Toward what |
|---|---|---|
| Where are the systems that receive the data? | ERP, Dynamics 365, SharePoint, the shared inbox: if everything lives in Microsoft 365 and Azure, orchestration is built right in, and page-by-page playback is the easiest to set up. | Online |
| The Sensitivity of Documents | Patient records, banking data, contract documents: when a policy prohibits these from being accessed by an online service, the data is processed in containers on your premises—while the orchestration can remain with Microsoft. | Hybrid or on-premises |
| Monthly page volume | For a few thousand pages per month, per-page pricing is the least expensive option. For volumes exceeding hundreds of thousands—such as an insurance company, a scanning service, or a government agency—on-premises computing becomes the more cost-effective option. The threshold is calculated over a three-year period. | On-premises above the threshold |
Microsoft lists its document reading, language, and vision services among those that run on Azure Local, including in offline operations. We verify this on a per-batch basis before each proposal: the list changes from quarter to quarter.
What the solution connects to
A document that has been read but isn't written down anywhere is useless. The value lies in the connection —and that has been our business for a long time, long before artificial intelligence.
Infrastructure, When It Is Needed
For most companies, none: at Microsoft, you pay per page read. When volume, latency, or privacy make local execution preferable, we also scale the necessary computing infrastructure.
Hundreds of thousands of pages per month
An insurer, a scanning service, a government agency, a company bound by a confidentiality agreement: the scanning services run in containers on Azure locally, on validated Dell nodes, using graphics processing units when the volume warrants it. Dell storage receives the original documents and handles their retention. This is a variation of this use case, not a separate one: the same workflow, a different execution location.
How We Carry Out the Project
Five steps. The prototype is based on your last 100 documents, not on the 10 cleanest ones—that’s the only way to determine the true pass rate without manual intervention.
Framing: The Process as It Is
Who receives what, through which channel, who re-enters it where, who approves it, and what discrepancies arise. And the status of the reference databases: a supplier with three different spellings is listed here.
Architecture and Rules
Online, hybrid, or on-site, depending on sensitivity and volume. Confidence thresholds, controls, and approval processes are developed in collaboration with the business.
The prototype, based on your last 100 documents
The complete workflow, based on actual documents from the past month. We measure the rate of fields entered correctly, the rate of documents processed without intervention, and the validation time for the remaining documents.
A realistic yield rate rarely exceeds what the prototype shows: that's the figure to keep in mind, not the one in a brochure.
Connection and Commissioning
ERP, Dynamics 365, SharePoint, the shared inbox, and the validation screen in Teams. First as a document type alongside manual entry, then on its own.
Maintenance
Errors reviewed each month, thresholds adjusted, new document types added, and usage tracked.
Estimated timeline: eight to twelve weeks from the first assessment to go-live for the first type of document—usually supplier invoices—and one channel. This is a rough estimate, not a commitment: measuring progress is part of the work. The second type of document takes a fraction of the time required for the first: the infrastructure and connections already exist.
Scope and Limits
What the processing does not do, and what we do not promise. No document is read with 100 percent accuracy, and the process is designed with that in mind.
A field is sometimes misread, with a high degree of confidence
A 3 mistook for an 8, a date in an unexpected format, a missing currency. The checks on your systems catch most of these errors—that’s their job—but not all of them. Thresholds and checks are tightened for the fields that matter most, with amounts and due dates at the top of the list.
The pass-through rate without intervention is not the same as the one listed in the brochure
It depends on the quality of your documents, your reference systems, and the variety of your suppliers. We determine this based on a prototype using your actual documents, and that figure is included in the proposal.
Inaccurate reference frameworks make monitoring ineffective
If a supplier appears three times, the reconciliation fails three out of four times, and everything is sent for manual verification. Cleaning up the databases takes place before the system goes live—and it remains valuable for everything else.
Handwriting and damaged documents have their limitations
These services read handwriting, but with a higher error rate than printed text. A poorly photographed form or a low-quality scan is difficult to read. The acceptable quality threshold is determined during the prototyping phase, using your actual documents.
The cost is charged per page
Online, each page viewed costs something, and a ten-page invoice costs ten times as much as a one-page invoice. We base our pricing on your actual volume, and we calculate the threshold beyond which the service becomes less expensive.
Automation does not replace decision-making
It reads, verifies, and records what matches. Paying a bill, approving an insurance claim, and renewing a contract remain human decisions, with the rights and delegated authority you already have.
Related cases and their corresponding pages
When another use case is the right one, and the technologies behind it.
Automated Email Processing
When the document itself is not the focus, but rather the written request that accompanies it: sort through it, prepare a response, and open the file.
View the case A similar casePrivate RAG
When the goal is not to extract data from fields but to query a confidential corpus using everyday language.
View the case TechnologyPower Apps : the confirmation screen and the surrounding application
When a process deserves an application, not just a workflow: the spreadsheet that becomes a full-fledged application.
View the page TechnologyDataverse : the extracted data, along with their permissions
The foundation of business applications, where read fields are stored along with their status and history.
View the page DepartmentBusiness Solutions: What Needs to Fit Your Business
Dynamics 365, Power Platform and the development surrounding it—the department responsible for these processes.
View the department ArticleWhere Artificial Intelligence Actually Reduces Costs
Jobs where earnings can be measured—and those where they cannot.
Read the articleLet's talk about this use case
Please send us your last 100 documents
The type of document that takes the most time to process, the channel through which it arrives, the system where it ends up, and a monthly order of magnitude. We’ll come back to the architecture we’d be exploring, the cost threshold, and what a prototype using your actual documents would allow us to measure—first and foremost, the rate of documents processed without intervention.
What we offer is the opportunity to meet the engineers who will do the work. Assessment, prototyping, and installation are all part of the project. The initial consultation, however, is free of charge.
Renens, Sion, Châtel-Saint-Denis
Microsoft Solutions Partner and Dell Technologies Gold Partner. Meter reading, regulations, connection to theERP , and—when the volume requires it—infrastructure, all handled by the same point of contact.
Renens VD +41 21 806 37 15
Sion VS +41 27 552 00 22
Châtel-Saint-Denis FR +41 26 322 59 05

