Procurement Risk Intelligence – AI Case Study

1. Business Scenario

A large global industrial manufacturing company operates multiple production facilities across Europe, Asia, and North America. Every year, the organization processes more than 4,000 purchase orders (POs) from a diverse supplier base. The supply chain is complex, with long lead times, fluctuating material availability, and highly variable supplier performance.

In recent years, the company has experienced increasing supply chain disruptions, including:

• delayed deliveries,
• quality issues,
• non-responsive suppliers,
• unstable pricing and material shortages.

These disruptions have led to production delays, increased operational costs, and reduced delivery performance. As part of a broader digital transformation initiative, the CFO and CTO are sponsoring the development of an AI-driven early-warning system that identifies potentially risky purchase orders as early as possible.

The goal is to enable procurement and production teams to:

• escalate critical orders earlier,
• activate alternative suppliers,
• adjust safety stocks proactively,
• reduce the risk of production downtime.

2. Your Role as a Forward Deployed AI Engineer

As a Forward Deployed AI Engineer, you collaborate directly with stakeholders across Procurement, Supply Chain, and Production. You are expected to:

• understand business processes and operational bottlenecks,
• analyze the data model and interfacing systems,
• design and prototype AI-driven solutions,
• communicate insights to both technical and executive audiences,
• deliver scalable, production-ready components as part of the Data Intelligence Factory.

The customer provides access to a Supabase database with simplified, SAP-inspired procurement tables. Your task is to build a working prototype that predicts whether a purchase order is “at risk”.

3. Objective of the Case Study

Build a machine learning solution that predicts whether a PO is likely to become risky based on:

• PO header information,
• item-level information,
• supplier master data,
• event logs (e.g., delays, escalations, supplier issues),
• free-text descriptions in English and German.

The final model should classify each PO as:

• 0 = not at risk
• 1 = at risk

You will later present your results to a panel consisting of the CFO and CTO.

---

What was done in recent projects (Build-up of PO_EVENTS)

Procurement specialists communicate regularly with suppliers. For every email interaction, a PO_EVENT is generated, and a conversation summary is stored in the comment field. An internal ML model classifies the comment into one of six event types, while the severity of each issue is manually assigned by a human reviewer.

---

4. Tasks

4.1 Data Extraction and Understanding

Using the provided database connection:

1. Load the tables LFA1, EKKO, EKPO, and PO_EVENTS.
2. Explore and document each dataset’s content and structure.
3. Join the tables to construct a training dataset.
4. Conduct exploratory data analysis (EDA), including:
   • PO volume and distribution,
   • supplier distribution,
   • risk ratio,
   • text field analysis (header_text, item_text, comment),
   • event frequencies and severity patterns,
   • numerical feature distributions.

4.2 Feature Engineering

Design a meaningful feature set combining:

• Text features
  Embedded using a SentenceTransformer model.
• Numerical features
  e.g., quantities, values, lead-time information.
• Categorical features
  e.g., country, document type.

4.3 Machine Learning Model

Train an appropriate supervised classification model to predict PO risk.

4.4 Evaluation

Evaluate the model using suitable metrics.

---

5. Business Deliverables (Important)

In addition to the technical implementation, candidates should address business questions, such as:

• What is the estimated financial impact of deploying this solution?

• Are there specific suppliers, materials, or regions that require particular attention once the system goes live?

• What risks or limitations should the organization be aware of before operationalizing this solution?

These aspects will be relevant for the management-level discussion. When answering these questions, you might want to base your analysis on historical data rather than hypothetical events.

---

6. Final Deliverables

Candidates are expected to submit:

1. A fully working Google Colab notebook including:

   • data loading,
   • feature engineering,
   • preprocessing pipeline,
   • model training,
   • evaluation.

2. A short slide deck (4–6 slides) for a presentation to the CFO and CTO explaining:

   • the business problem,
   • your proposed AI solution,
   • key design decisions and assumptions,
   • expected business value and impact.
     Please use the slide master provided in our invitation.

Note: We understand that such an exercise could theoretically be expanded indefinitely. As guidance, we ask candidates to invest no more than approximately two days into the development of their solution and the preparation of the presentation. The evaluation will be based on the results and insights expected from this level of effort.

---

EKKO

Field	Type	Description
ebeln	text	Purchase order number (unique identifier of PO header)
lifnr	text	Supplier number (vendor ID assigned in SAP)
bukrs	text	Company code – identifies the legal entity
bedat	date	Creation date of the purchase order
bsart	text	Document type (e.g., NB – standard PO)
release_indicator	bool	Whether the PO is released/approved
header_text	text	Free-text PO header notes

EKPO

Field	Type	Description
ebeln	text	Purchase order number (links to ekko)
ebelp	int4	PO item number (line number within the PO)
matnr	text	Material number (identifier of the ordered material)
menge	numeric	Ordered quantity
eindt	date	Delivery date requested
price	numeric	Net price of the item
item_text	text	Item-level text (e.g., description or notes)

LFA1

Field	Type	Description
lifnr	text	Supplier number (primary key, links to ekko.lifnr)
name1	text	Supplier name (company name)
country	bpchar	Country code (ISO-style, e.g., DE, US)
risk_score	int4	Custom risk attribute (e.g., 1–10 or risk categories)

PO_EVENTS

Field	Type	Description
id	int8	Unique event identifier
ebeln	text	Related purchase order number (links to ekko)
event_ts	timestamp	Timestamp of the communication event
event_type	text	ML-classified event category (e.g., delay, confirmation, cancellation etc.)
severity	int4	Manually assigned severity level of the issue (ranking)
comment	text	Conversation summary extracted from email text