Procurement · Production

Five real-world AI use cases from Celonis AI Lab in Columbus

The problem

Across five companies at Celonis AI Lab, shared challenges included procurement spend not linked to master catalogs, manual helpdesk operations requiring multiple FTEs, category managers lacking visibility into contract leakage, chargeback teams manually resolving large volumes of daily rejections, repeated manual vendor contact for PO confirmations, and insurance adjusters manually checking compliance documents across multiple states.

First attempt

Manual random-sampling for insurance compliance revealed blind spots in regulatory adherence, while manual helpdesk data retrieval caused processing delays and strained vendor relationships, and repeated vendor contact for PO confirmations resulted in a high rate of missed confirmations.

Workflow diagram · grounded in source
1
Vendor inquiry triggers agent
trigger
“Upon receiving a vendor inquiry, an agent is triggered automatically that categorizes the request, assigns both a severity level, and identifies the appropriate resolution team”
2
AI categorizes and routes inquiry
ai_action
“categorizes the request, assigns both a severity level, and identifies the appropriate resolution team”
3
AI drafts response options
ai_action
“The agent then drafts comprehensive response options for the resolution teams, automating time-consuming manual searches”
4
Human review and finalization
human_review
“enabling FTEs to review and finalize the agent's pre-written responses”
5
LLM extracts PDF confirmation data
ai_action
“the solution leverages a Large Language Model (LLM) to extract relevant confirmation data”
6
Confirmation data written to system
integration
“ensuring it is accurately written back into Celonis”
Reported outcome

Prototype solutions aimed to deliver significant cost savings via catalog conformance, speed up vendor inquiry resolution with improved transparency, automate C-suite Spend Under Management reporting, stop chargeback rejections before they occur, eliminate repeated manual vendor contact, and remove the burden of manually checking insurance compliance documents.

Reported metrics
FTEs required for helpdesk operationssix full-time employees (FTEs)
Daily partner communication lines processedhundreds of thousands of partner communication lines every day
Daily chargeback rejection rateranging between one and two percent
Procurement cost savingssignificant cost savings
Show all 7 reported metrics
FTEs required for helpdesk operationssix full-time employees (FTEs)
daily partner communication lines processedhundreds of thousands of partner communication lines every day
daily chargeback rejection rateranging between one and two percent
procurement cost savingssignificant cost savings
vendor inquiry resolution timespeed up resolution times
vendor transparencysignificantly improve transparency
planning reliabilityenhances real-time planning reliability
Reported stack
Process CopilotsAnnotation BuilderPrediction BuilderOrchestration EngineLarge Language Model (LLM)
Source
https://www.celonis.com/blog/ai-on-display-in-columbus-five-real-world-use-cases-from-celonis-ai-lab
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Prototype solutions aimed to deliver significant cost savings via catalog conformance, speed up vendor inquiry resolution with improved transparency, automate C-suite Spend Under Management reporting, stop chargeback…

What tools did this team use?

Process Copilots, Annotation Builder, Prediction Builder, Orchestration Engine, Large Language Model (LLM).

What results were reported?

FTEs required for helpdesk operations: six full-time employees (FTEs); Daily partner communication lines processed: hundreds of thousands of partner communication lines every day; Daily chargeback rejection rate: ranging between one and two percent; Procurement cost savings: significant cost savings (source-reported, not independently verified).

What failed first in this deployment?

Manual random-sampling for insurance compliance revealed blind spots in regulatory adherence, while manual helpdesk data retrieval caused processing delays and strained vendor relationships, and repeated vendor contac…

How is this procurement AI workflow structured?

Vendor inquiry triggers agent → AI categorizes and routes inquiry → AI drafts response options → Human review and finalization → LLM extracts PDF confirmation data → Confirmation data written to system.

Related procurement cases