Home · Solutions · Other solutions

Solution · Other solutions

Invoices to processing, statements to reconciliation, payment questions answered from SAP

The AP mailbox sorted every hour, payment questions answered

Every message in the accounts-payable mailbox is classified within the hour, routed to the process that owns it, and payment questions are answered from SAP in the supplier's language.

DepartmentalMicrosoft TeamsHuman in the loopAI where it earns its place
6,000messages a month land in one shared mailbox that serves as intake, help desk and archive at once for this illustrative facilities group.

Executive summary

Challenge

Your suppliers ask a question your SAP already answers, and a person spends ten minutes typing it out.

What changes

We turn the mailbox into a queue with intents.

Business value

Invoices reach processing the day they arrive rather than the day somebody opens them, which puts three days back into every cycle time.

Systems involved

SAP S/4HANA for status and history; the invoice pipeline, statement reconciliation and vendor verification; Power BI

Business problem

Accounts payable

The accounts-payable mailbox is an intake, a help desk and an archive at the same time, and it was designed for none of them. Invoices have to be found among everything else before processing can start. Payment questions need a lookup and a written answer in English or Polish, and each one takes a person ten minutes to produce a status that was in SAP the whole time.

Because sorting is manual, priority goes to whoever pushes hardest. Messages about one invoice scatter across weeks and threads, so nobody can say what a supplier was told without searching Outlook. Statements arrive and are filed unread, which is where missing credits go to hide. Bank-change requests arrive in the same stream and are sometimes answered as ordinary questions, which is the entry point every payment-diversion attempt is looking for.

The problem survives because it looks like a staffing problem. A fourth clerk brings the unread count down for a quarter, the next contract adds subcontractors, and the count returns. Meanwhile the real cost lands elsewhere: invoices found three days late lose three days of cycle time, and contractors who cannot get a payment date stop showing up on site.

How it works today

  1. PersonA clerk opens a message and decides what it is: invoice, question, statement, bank change or noise
  2. PersonInvoices are saved to the processing folder; questions are left for later
  3. WaitingThe supplier sends a reminder and a second thread now exists for the same invoice
  4. PersonThe clerk searches SAP by invoice number, then by amount, and drafts a reply in the supplier's language
  5. Risk of errorA bank-change request is answered as a routine question, and a statement is filed without being read
  6. PersonA site manager asks the same question in Teams, and another clerk searches SAP again
PersonWaitingRisk of error

Why the current process costs more than it appears

Time that disappears before anyone measures it.

  • Every message is touched at least twice, once to sort it and once to act on it, and only the second touch feels like work worth recording.
  • Delay compounds quietly: an invoice found on day three has already lost three days of cycle time, along with any early-payment discount attached to it.
  • Unanswered questions turn into calls, and calls turn into escalations, until suppliers learn that the fastest route to a payment date is the finance director's assistant.
  • Knowledge sits with whoever knows which supplier writes in Polish and which one threatens legal action after two reminders, which makes that person impossible to replace and impossible to promote.

Cost of inaction

Twelve months of sorting before anything is processed≈ €147,000
Three more years of the same mailbox≈ €441,000
When new contracts take the mailbox to 7,500 messages a month≈ €184,000

A fourth clerk fixes this for two quarters. Then the next contract adds subcontractors, the unread count returns, and the mailbox is worked in the evening again. Suppliers learn which route gets an answer and use it, so escalations arrive on the executive floor rather than in the queue where they started.

The quieter risk is in the messages nobody classified. A statement filed unread hides a credit note that stays unbooked. A bank-change request answered as an ordinary question is how a year's incident begins. And the invoices buried in the pile lose days that later appear as late payment somewhere else in finance, where nobody attributes them to a mailbox.

Illustrative scenario

A plausible organisation with realistic proportions. The figures are there to be recalculated on your data; they are not a client result.

Organisation

A facilities-management group providing cleaning, security and maintenance across the United Kingdom, Ireland and Poland, working with about 1,400 subcontractors and suppliers, with accounts payable run from a shared-service team in Poland on SAP S/4HANA.

Volume

About 6,000 messages a month in English and Polish: a little over half are invoices, a fifth are payment questions, and the remainder are statements, bank changes, internal questions from site managers and noise. Three clerks share the mailbox alongside processing work.

Current process

Each message is opened and classified by hand, invoices are moved to a folder, questions are researched in SAP one at a time, and the oldest unanswered question is usually two weeks old.

Bottleneck

Three minutes of sorting per message and twelve minutes per payment question, both of them repeated whenever a reminder arrives or a site manager asks the same thing in Teams.

Solution

UiPath Communications Mining classifies every message and attachment by intent, routes invoices, statements and bank changes to the processes that own them, and a robot answers payment questions from SAP with a reply drafted in the language of the request and approved by a clerk.

Potential outcome

In the modelled case the mailbox holds no undecided messages for longer than an hour, payment questions are answered the same day, and roughly two people's worth of monthly time returns to processing and supplier work. These are modelled figures rather than a client measurement.

Proposed solution

We turn the mailbox into a queue with intents. UiPath Communications Mining reads each message and its attachments and places it in a taxonomy your team defines: invoice, credit note, statement, payment question, bank-change request, internal question, noise. It also pulls out what the message refers to, the supplier, the invoice numbers, the amounts. Invoices move to the invoice pipeline, statements to reconciliation and bank-change requests to the vendor-verification control, each one tagged, filed and acknowledged to the sender.

Payment questions are answered rather than routed. A robot looks the invoices up in SAP and returns the state they are in: received, in approval, blocked with a reason, scheduled for a date, or paid with a payment reference. UiPath GenAI Activities compose a reply from that data and an approved template, in the language of the request, and the draft goes to a clerk as a task in UiPath Action Center. Nothing is released automatically on day one. Once accuracy has been measured for long enough to argue about, routine status replies can go out within written limits, with a daily sample reviewed.

Internal questions stop being email at all. A UiPath conversational agent published to Microsoft Teams gives site managers the same SAP status, including a supplier's previous invoices and totals, without anyone in accounts payable being involved. Every message about one invoice is linked into a case record in UiPath Data Fabric, and a Power BI dashboard reports the mailbox by intent, response time and backlog age.

Native capabilities used

UiPath Communications Mining intents, extracted fields and sentiment; UiPath GenAI Activities for drafting and translation; UiPath Action Center tasks completed in Microsoft Teams and Outlook; UiPath Data Fabric entities for the case record; a UiPath conversational agent published as a Microsoft Teams app; UiPath Orchestrator queues, triggers and credential store

What we build

The intent taxonomy and routing rules; the SAP status lookup and its handling of partial matches; the reply templates and written release limits; the case model linking messages to invoices; the Teams agent's tools and guardrails; the Power BI dashboard

Custom integration

The shared mailbox through the Microsoft Outlook 365 connector for triggers, folders and sending; SAP S/4HANA invoice and payment status through the SAP BAPI and OData connectors; notifications through the Microsoft Teams connector

How the automated process works

  1. AutomationA new message in the shared mailbox triggers intake; Communications Mining classifies it and extracts supplier, invoice numbers and amounts
  2. AutomationInvoices, statements and bank-change requests are routed to their own processes, filed, and acknowledged to the sender
  3. SystemFor a payment question the robot retrieves the status, history and totals for the invoices named, from SAP
  4. AutomationGenAI Activities draft the reply in the language of the request, from the retrieved data and an approved template
  5. PersonThe clerk reviews the draft in Action Center and sends it; disputes, legal threats and anything unclassified stay in a short human queue
  6. AutomationMessages are linked into a case record, and site managers get the same status from the Teams agent without writing to the mailbox
AutomationSystemPerson

Human-in-the-loop model

Automation handles

  • Classifying every message and attachment by intent and extracting the references it contains
  • Routing invoices, statements and bank changes to the processes that own them, with an acknowledgement
  • Retrieving payment status, history and totals from SAP for a named supplier or invoice
  • Drafting the reply and linking every message about one invoice into a single case

People decide

  • Whether each drafted reply goes out, until measured accuracy supports releasing routine ones within written limits
  • Disputes, legal threats and anything the classifier could not place with confidence
  • Which supplier needs a call instead of a reply, and what is said on it
  • The taxonomy and the templates, reviewed monthly as suppliers and phrasings change

Before and after

BeforeAfter
Sorting the mailbox3 min per message, by handclassified on arrival, 15% still opened by a person
Answering a payment question12 min of searching and writing5 min of reviewing a draft, falling as accuracy is proven
Oldest unanswered questionabout two weekswithin the working day
Where a bank-change request goesinto the same queue as the questionsinto the verification control, tagged and logged

Systems and integrations

The stack is deliberately short: one engine, one execution layer, one place where a person decides.

Inputs

  • the accounts-payable shared mailbox in Outlook
  • attachments of every format
  • questions asked by site managers in Microsoft Teams

Automation layer

  • UiPath Orchestrator
  • UiPath Robots
  • UiPath Communications Mining
  • UiPath GenAI Activities
  • UiPath Action Center
  • UiPath Data Fabric

Target systems

  • SAP S/4HANA for status and history
  • the invoice pipeline, statement reconciliation and vendor verification
  • Power BI

Human touchpoints: draft approval in Action Center; the short queue for disputes and unclassified messages; the Teams agent for internal questions

the accounts-payable shared mailbox in OutlookUiPath OrchestratorUiPath RobotsSAP S/4HANA for statusdraft approval in Action Center

Technologies used

UiPath Communications Mining

classifies every message and attachment by intent and extracts supplier, invoice and amount references

A
UiPath GenAI Activities

drafts the reply from SAP data and an approved template, in the language of the request

A
UiPath Robots + Orchestrator

retrieve status from SAP, hold the queue, keep secrets in the credential store and log every step

A
UiPath Action Center in Microsoft Teams

draft approval, exception tasks and the short human queue

A
UiPath conversational agent in Microsoft Teams

answers "was this paid?" for site managers and buyers, from the same SAP data

A
SAP S/4HANA (BAPI and OData connectors)

invoice status, payment dates and references, supplier history and totals

A
Power BI

mailbox dashboard: volume by intent, response time per intent, backlog age, drafts corrected

A
Averified product capability (vendor documentation)

Illustrative economic model

Numbers you can check against your own data.

Illustrative model
6,000 messages a month × 3.95 minutes released each= 395 h / month
395 h × €31 fully loaded clerk cost= €12,245 / month
× 12 months≈ €146,940 / year
Annual capacity released (illustrative)≈ €146,940

Every message in this model carries 3.95 released minutes, which is a blend rather than a stopwatch reading: sorting 6,000 messages at three minutes is 300 hours a month and answering 1,200 payment questions at twelve minutes is 240 more, while the target state keeps 45 hours of sorting for the 900 messages a person still opens and 100 hours of reviewing drafted replies. The 395 hours between them, spread across all 6,000 messages, give the figure below. €31 an hour is an assumed fully loaded cost for an accounts-payable clerk in a Central European shared-service team, and none of these numbers was measured at a client.

Run the numbers on your data

hours released per month
of annual capacity released

An illustrative estimate from your own inputs. It models released capacity; it is not a promise of savings.

Business benefits

  • Invoices reach processing the day they arrive rather than the day somebody opens them, which puts three days back into every cycle time
  • Payment questions are answered within the hour from the system of record, so reminders stop multiplying and calls to the finance director's assistant stop arriving
  • Bank-change requests are never answered as routine questions, because the classifier hands them to the verification control with a log entry
  • Site managers get their answer in Teams without accounts payable in the loop, which removes an entire category of internal email
  • Every conversation about an invoice is one case with a history, so what a supplier was told is a record rather than a search through three mailboxes

The management view

  • Workload becomes measurable by intent rather than by unread count, so the team scales with contracts won instead of with backlog
  • Reply quality is governed by templates and approval statistics rather than by each clerk's writing style and mood
  • Which suppliers ask most becomes visible, and that list is usually a list of invoices stuck for a reason worth fixing

Board-level KPIs

backlog age in the AP mailboxresponse time per intentshare of payment questions answered from the system of recorddrafts corrected before sendinginvoice cycle time from arrival to processing

Security and governance

Control is not an add-on.

  • The mailbox is reached by a service account through Microsoft Graph application permissions scoped to that one mailbox, not to the tenant, and the robot reads SAP with an account that cannot change payment data
  • Replies leave from the accounts-payable mailbox under templates the AP manager owns; automatic release applies only to intents and limits written down in advance, with a daily sample reviewed by a person
  • Bank-change requests are routed and logged, never answered by the automation, because that decision belongs to the verification control and to a call
  • Supplier messages carry personal data and bank details, so the generative step runs under the UiPath AI Trust Layer with PII masking and an allow-listed model; the classifier is trained on your own messages inside the EU region of UiPath Automation Cloud, your Microsoft 365 data stays in the EU Data Boundary, and case records are retained under Microsoft Purview labels

Why now

01

E‑invoicing mandates are moving documents out of email into structured channels, so the mailbox that remains will hold questions rather than invoices, and questions are exactly what a folder rule has never handled

02

Classification has moved past keyword rules to reading intent across attachments and mixed-language threads; the answer was always in SAP, and what changed is that a robot can fetch it and compose the sentence while a person still signs it off

03

The waiting cost is visible in the model: about €12,245 a month of clerk time, of which the larger part is not answering suppliers but deciding what each message is

Relevant executive roles

CFO

Supplier escalations stop reaching the executive floor, and the mailbox stops being the door through which a bank-change fraud walks in

Head of Shared Services

Workload becomes measurable by intent, so the team scales with contracts won rather than with the unread count

AP Manager

Clerks process instead of sort, and every supplier conversation has a case history somebody else can read

Common questions and objections

Outlook rules can sort this mailbox.

Rules sort by sender and subject. They cannot tell a payment question from a bank-change request written in the same polite tone, nor read the attachment that decides which it is.

We cannot let a machine write to our suppliers.

It drafts and a clerk sends. Routine status replies go out on their own only after the correction rate has been measured, only within limits you write down, and only with a daily sample on somebody's desk.

Our suppliers write in several languages.

The classifier is trained on your own messages in each language rather than a generic corpus, and the draft is composed in the language the supplier used, with a person reviewing it.

When this is not the right solution

  • A mailbox with a few hundred messages a month, where folder rules, a tracker and a clear rota are cheaper than a model
  • A supplier portal that already carries status questions and that most suppliers actually use
  • No willingness to standardise reply templates; without them there is nothing to control the drafting against

A question for the next management meeting

Our suppliers already know the answer sits in SAP, so what does it cost this company that a person has to find it and type it out, twice, before they hear it?

Implementation approach

A scope without ambiguity, before anything is signed.

We deliver

  • An analysis of one month of your mailbox: intents, languages, volumes and response times, labelled with your team
  • The intent taxonomy, routing rules and reply templates, agreed with the AP manager who will own them
  • A trained and validated Communications Mining model, plus the SAP status lookup and its handling of partial matches
  • Reply drafting with the approval flow in Action Center, the case model and the conversational agent in Teams
  • The Power BI mailbox dashboard, clerk training, a controlled start and run-phase support

We need from you

  • One month of the mailbox as an export, with permission to label a sample
  • A service account with access to that mailbox, and read access to invoice and payment data in SAP
  • A named owner in accounts payable for the taxonomy, the templates and the release limits
  • Your escalation rules, so the short human queue inherits them rather than inventing new ones

Stages

Discovery

One month of messages exported and labelled; taxonomy and templates agreed

Design

Routing rules, SAP lookups, draft templates, release limits, the case model

Build

Model training, robots, approval flow, Teams agent and dashboard in your environment

Pilot and scale

Live classification with every reply approved and accuracy measured weekly, then statements, the case record and limited automatic release

Departmental. Effort follows the number of languages, the spread of the intent mix, how clean the SAP status data is for partial matches, and how much template standardisation the team is willing to accept.

1,146 unread, and SAP already knows the answer to most of them.

Export one week of your accounts-payable mailbox and let us label it with your team. We come back with the intent mix, the share that could be routed without a person, and the share of questions answerable from your own data today.

Label one week of your AP mailbox

The neighbouring process usually has the same problem

Industries we deliver this in most oftenManufacturing & industryRetail & e‑commerceServices & ITShared services

Browse all 173 solutions