Home · Solutions · Other solutions
Solution · Other solutionsReading, coding and routing happen without a person; the decisions still need one
Invoices read, coded and routed; people see only exceptions
Every invoice is read in three languages, checked against SAP, coded from supplier history and routed to the budget owner in Teams; AP works the exceptions.
Executive summary
Somebody still types the PDF into the ERP, guesses the cost centre and chases the approver.
The pipeline starts where invoices already arrive, not where anyone wishes they arrived.
Cost per invoice falls because keying disappears for the matched majority and clerks work only on the exception classes the rules present.
SAP S/4HANA (MIRO posting via BAPI); SharePoint archive with retention labels
Business problem
Accounts payable
Invoice handling breaks in three places, and each one belongs to a different person. Capture breaks because the supplier chooses the format: a PDF attachment, paper through the post room, a portal download, occasionally a photograph taken on a phone. Coding breaks because an invoice without a purchase order needs somebody to decide the account, the cost centre and the tax code, and that somebody is almost never in accounts payable. Routing breaks because the approval rule lives in people's heads: marketing goes to Anna, unless it is above ten thousand, and then it goes to the CMO.
The result is a queue nobody can see in full. Suppliers telephone, controllers accrue for invoices they cannot list, and auditors ask why one general ledger account carried five different services last year. Everything about the process is visible to somebody and to nobody in particular.
It persists because every fix so far has been a person. A temporary clerk at the peak who then stays. A mailbox rule. An Excel tracker maintained by the most conscientious member of the team, who goes on holiday in August and takes the routing rules with her.
How it works today
The pattern below is what we find in shared-service centres before automation, whatever the ERP.
- PersonThe clerk saves the attachment from the shared mailbox or collects the scan from the post room
- SystemShe looks up the vendor and any purchase order number in SAP, then keys header and line data into MIRO
- PersonFor an invoice without a purchase order she guesses the account and cost centre, or asks somebody who knows
- WaitingThe question sits in a Teams chat or an inbox until the controller has time, typically two to four days
- PersonShe emails the presumed approver, who forwards it, and follows up after the supplier's reminder arrives
- Risk of errorThe approval is copied from a chat into the tracker; the invoice is posted or parked depending on the day
- WaitingAt month end the accrual is reconstructed from the tracker and the mailbox, and the numbers are late
Why the current process costs more than it appears
Behind every exception is an hour nobody logged.
- Cost per invoice, as finance reports it, counts the keying and nothing else. The investigation around it, who ordered this, was it delivered, which cost centre pays, is unrecorded work spread across three departments.
- Controllers accrue for invoices they cannot enumerate, so the month-end figure is an estimate defended rather than a number produced, and the correction lands in the following period.
- Approvals given from a phone without the order in front of the approver are a formality, not a control, which is precisely the point auditors return to every year.
- Early-payment discounts expire in the mailbox, and no report shows what they were worth, because nothing measures the days between arrival and OK-to-pay.
- A new plant adds a language and a set of formats, and the only lever anyone has is another clerk, so unit cost never falls with volume.
Cost of inaction
Formats arrive faster than people can be hired. Every new supplier brings a layout, every new entity brings a language, and the queue grows with revenue while the centre grows with budget rounds. Within a year the peak is covered by a temporary clerk who is still there the following December, cycle time stays near two weeks, and suppliers price the delay into their next quotation.
The dependency is the part that does not appear in any budget. One person knows which supplier applies the wrong tax rate, which plant controller answers quickly and which portal needs a manual download on the first working day. Her holiday is a risk, her resignation is an incident, and none of what she knows is written down anywhere the company can read.
A plausible organisation with realistic proportions. The figures are there to be recalculated on your data; they are not a client result.
An industrial manufacturer with four plants in Poland, Czechia and Germany. Accounts payable runs from a seven-person shared-service centre on SAP S/4HANA, with Microsoft 365 as the workplace.
About 3,200 supplier invoices a month in German, Polish and English. Roughly 60% quote a purchase order; the rest are services, utilities, rents and marketing. They arrive as mailbox attachments, as paper scanned by the post room and as portal downloads.
Attachments are saved and keyed into MIRO; coding for invoices without an order is settled by asking a controller; approvals happen in email and Teams chats, and a controller keeps a spreadsheet for the accrual. Receipt to OK-to-pay averages twelve working days.
Around fifteen minutes of effort per invoice, most of it not keying: the coding query, the search for the right approver and the reminders that follow.
A robot collects every document from the mailbox, the scan library and the portals; UiPath Document Understanding reads it in all three languages; rules check the vendor and the tax treatment in SAP and propose the coding from supplier history and the contract register; the budget owner approves in Microsoft Teams and the robot posts through SAP interfaces.
In the modelled case roughly six invoices in ten reach posting without a human touch, the remainder take about six minutes each instead of fifteen, and twelve working days to OK-to-pay becomes a matter of days. These are modelled figures, not a client result.
Proposed solution
The pipeline starts where invoices already arrive, not where anyone wishes they arrived. A robot collects attachments from the shared mailbox, files from the scan library on SharePoint and downloads from the supplier portals, splits multi-document files and puts each invoice into an Orchestrator queue. UiPath Document Understanding, using the pre-trained Invoices model, reads header and line data in German, Polish and English. Anything below the confidence threshold becomes a Validation Station task in which a clerk corrects only the fields in doubt, and each correction feeds back into the model.
The rules layer is where most of the value sits, because it replaces the questions rather than the typing. It checks the vendor against SAP master data, finds the purchase order from supplier and delivery references when the invoice does not quote one, verifies the tax treatment for the plant's country, and screens for a document already received. For invoices with no order it proposes the general ledger account and cost centre from that supplier's posting history and the contract register, with the confidence and the reason shown, never applied silently.
Approval reaches the budget owner as an Action Center task inside Microsoft Teams, carrying the image, the proposed coding and the delegation limit that applies. Approved invoices post to SAP through BAPI calls under a service account; the image, the extraction, the rules applied and the decision are archived on SharePoint under a retention label. Anything the rules cannot settle becomes an exception with a named class, so the AP lead manages a queue instead of a mailbox, and UiPath Insights reports volume, touchless rate and ageing by supplier and invoice type.
UiPath Document Understanding pre-trained Invoices model and Validation Station; UiPath Orchestrator queues, triggers, credential store and audit; UiPath Action Center tasks as actionable notifications in Microsoft Teams; UiPath Insights; SharePoint versioning with Microsoft Purview retention labels
The collection and dispatching workflows across mailbox, scan library and portals; the rule layer for vendor, order discovery, tax treatment and repeat documents; the coding proposal logic and its confidence handling; the approval matrix with delegation and reminders; the exception classes and the reporting
SAP S/4HANA lookups and posting through UiPath SAP automation (BAPI and OData); supplier-portal downloads where the portal exposes no interface
How the automated process works
- AutomationA new message in the shared mailbox, a new file in the scan library or a scheduled portal run triggers intake; documents are split and queued
- AutomationDocument Understanding classifies each document and extracts header and line data in the three languages, with a confidence score per field
- PersonFields below the threshold go to a Validation Station task, where a clerk corrects only what is uncertain
- SystemRules check vendor, order, tax treatment and prior receipt in SAP, and propose the account and cost centre for invoices without an order
- PersonThe budget owner approves or returns the invoice in Microsoft Teams, seeing the image, the proposed coding and the limit that applies
- AutomationApproved invoices post to SAP and are archived with the extraction, the rule trail and the decision; reminders escalate after the agreed service level
- AutomationReturned invoices go back to the supplier with a reason code, and Insights reports touchless rate, ageing and exceptions by class
Human-in-the-loop model
Automation handles
- Collection from mailbox, scan library and portals, including splitting and queueing
- Extraction of header and line data in German, Polish and English
- Vendor, order, tax and prior-receipt checks in SAP, and the coding proposal
- Routing, reminders, posting, archiving and the daily status view
People decide
- Uncertain fields in validation, corrected once and learned by the model
- Coding the rules could not settle, or a proposal the controller wants to change
- The approval itself, within a delegation limit finance owns
- Disputes, credit notes and anything the rules flag as unusual
Before and after
Systems and integrations
Everything below runs on licences and systems you already hold, or would need anyway.
Inputs
- Outlook shared mailbox
- scan library on SharePoint
- supplier portals
- the contract register
Automation layer
- UiPath Orchestrator
- UiPath Robots
- UiPath Document Understanding
- UiPath Action Center
- UiPath Insights
Target systems
- SAP S/4HANA (MIRO posting via BAPI)
- SharePoint archive with retention labels
Human touchpoints: Validation Station tasks in Teams; budget-owner approvals in Teams; the exception queue for the AP lead
Technologies used
reads header and line data from PDFs, scans and photographs in three languages; Validation Station for uncertain fields
Aqueues every invoice, runs the rules, posts to SAP, retries, logs and audits
Awatches the shared mailbox and the scan library and writes the archive
Avalidation, coding and approval tasks completed without leaving Teams
Avendor, order and tax lookups; invoice posting under a service account
Athe archive of image, extraction, rules applied and decision
Atouchless rate, ageing and exception classes for the AP lead
Aaccount and cost centre derived from supplier history and the contract register
CIllustrative economic model
A model, not a promise.
Netting comes first here. 3,200 invoices at fifteen minutes is 800 hours a month today, and the future state keeps 160 of them: six minutes on each of the 1,280 exception invoices and one minute of spot-check on each of the 1,920 touchless ones. That leaves 640 hours released, which is why the calculator carries 2,560 rather than 3,200 as its volume. €36 an hour is an illustrative fully loaded cost for a shared-service clerk in Central Europe, and the touchless share of 0.6 is an assumption the pilot replaces with a measurement.
Run the numbers on your data
An illustrative estimate from your own inputs. It models released capacity; it is not a promise of savings.
Business benefits
- Cost per invoice falls because keying disappears for the matched majority and clerks work only on the exception classes the rules present
- Cycle time shortens because the approval arrives with everything needed on the card and the reminder runs without anyone composing it
- Coding quality improves because the proposal comes from that supplier's posting history rather than a guess under time pressure
- The accrual becomes a report from the queue rather than a reconstruction from a tracker and a mailbox
- A new plant, language or entity is absorbed by the model and the rules instead of by recruitment
- Approval evidence exists per invoice, with user, timestamp and the document version that was seen
The management view
- Workload is planned rather than discovered: the queue shows how many invoices are in each state, how long they have waited and who holds them
- Approval service levels become measurable per approver, which ends the argument about whether finance or the business is slow
- Touchless rate by supplier and invoice type tells the AP lead which suppliers and which rules produce the exceptions
- The process stops depending on the person who remembers which supplier always applies the wrong VAT rate
Board-level KPIs
Security and governance
An auditor should be able to reconstruct every decision.
- The robot posts under a named SAP service account limited to the transactions the process needs, and its secret lives in the Orchestrator credential store rather than in a workflow
- The person who enters an invoice can never be the person who approves it: approval stays with the human budget owner and is recorded with user, timestamp and document version
- The shared mailbox is reached with application permissions scoped to that mailbox alone, not with anybody's personal account
- Documents and extracted data stay in the EU: your Microsoft 365 tenant inside the EU Data Boundary, robots in the EU region of UiPath Automation Cloud
- Extraction runs under the UiPath AI Trust Layer with a model allow-list and logging of every call, and the SharePoint archive carries retention labels matching the statutory period
Why now
European e‑invoicing mandates are pushing structured invoices into the ERP, and a pipeline built now takes XML through the same validation, coding and approval as a PDF, which is the part the mandates do not solve
Pre-trained invoice models read multi-language, multi-format documents well enough for a human-validated pipeline without months of training data, so the extraction step is no longer the project risk it once was
Approvals can live in Teams with the document on the card, which removes the portal nobody opened and the email nobody answered, and that is where most of the twelve days actually go
Relevant executive roles
Cost per invoice, days to OK-to-pay and approval evidence become three numbers on one page rather than three separate arguments
The accrual is read from the queue instead of being reconstructed, so the close stops waiting for a folder of unposted documents
Growth and new entities are absorbed by the same seven people, whose work shifts from keying to exceptions
Blocked invoices stop turning into blocked deliveries and irritated suppliers on the plants' doorstep
Common questions and objections
Pre-trained models handle the common layouts on day one, uncertain fields go to a person, and the pilot measures the real rate on your documents before anyone commits to a number.
OCR gives you characters. The value is everything after it: validation against SAP, order discovery, a coding proposal from history, the approval and the audit trail.
Structured invoices remove the extraction step, not the validation, the coding or the approval. The same pipeline takes XML as an input and applies the same rules.
When this is not the right solution
- Fewer than a few hundred invoices a month, where a clerk with a good mailbox routine costs less than a pipeline
- A mature invoice-management module already in use in the ERP, where the real gap is discipline rather than tooling
- Vendor master data so poor that vendor and order matching would fail on most documents, in which case master data comes first
A question for the next management meeting
Twelve working days pass between receipt and OK-to-pay: on how many of them is anyone actually working on the invoice, and what would the centre do with the rest?
Implementation approach
The first week looks the same at every client: we look at the data.
We deliver
- One month of invoices classified by channel, language, order status, approver and exception class, with the volumes behind each
- Configuration and tuning of the Document Understanding Invoices model on your suppliers' documents in all three languages
- The rule layer: vendor validation, order discovery, tax treatment, prior receipt and the coding proposal with its confidence handling
- The approval matrix with delegation limits, reminders and escalation, delivered as Action Center tasks in Teams
- SAP posting and error handling, the SharePoint archive with retention labels, and Insights reporting for the AP lead
- A parallel run for one plant, then rollout entity by entity with hypercare and a runbook for the team
We need from you
- Three months of invoice history with the matching SAP postings
- The approval matrix, delegation limits and the current coding conventions
- Service accounts for SAP test and production and access to the shared mailbox and the scan library
- A process owner in accounts payable and a controller who owns the coding rules
Stages
Discovery
Channels, languages, order share, exception classes and volumes mapped with the process owners
Design
Target flow, rule set, coding logic, approval thresholds and the security model
Build
Document model, robots, SAP integration, Teams and Action Center touchpoints in your environment
Validation
Parallel run on real invoices for one plant until the touchless rate is measured rather than estimated
Go-live
Controlled start per entity with supervision, then hypercare
Optimisation
Rule and model tuning, new suppliers, further entities and languages
Departmental. Effort follows the number of legal entities and document formats, the share of invoices without an order, and the state of the vendor master.
Three languages, three channels, and one clerk who knows every supplier's quirks.
Send us one month of invoices exactly as they arrived, PDFs, scans and portal exports. We come back with the touchless rate a pilot would start from and the exception classes behind it.
Measure the touchless rate on one monthThe neighbouring process usually has the same problem
Stop paying your finance team to move numbers from PDFs into the ERP.
View solution Finance & accountingInvoices from KSeF booked to the right budget lineA structured invoice states the amount. It does not state which unit ordered the work.
View solution Other solutionsPrice and quantity mismatches resolved the day they appearA variance released under payment-run pressure becomes next quarter's price, and nobody negotiated it.
View solution Other solutionsThe same invoice by e‑mail, portal and post, paid onceA hyphen, a rescan or a reissued number is enough for the same invoice to be paid twice.
View solution Case studyInvoice disputes without escalationA disputed invoice is not one task — it is an investigation: order, delivery, contract, correspondence.
View case study Case studyAutomated P2P query handlingInvoice copies, statuses and statements delivered instantly — around the clock.
View case studyIndustries we deliver this in most oftenManufacturing & industryRetail & e‑commerceServices & ITShared services