Data Requirements: What AI Agents Need to Learn Your Process

The Data Question 

"What data do AI agents need?" determines implementation preparation. Understanding data requirements - historical examples, decision criteria, outcome definitions - enables realistic timeline planning. 

Reality: Data gathering takes 2-3 weeks, not months. Implementation partner guides the process. 

Data Category 1: Process Examples 

Historical Exception Scenarios 

What's needed: 20-30 examples of typical exceptions with outcomes 

For AR collections: 

• Customer account details 

• Invoice amount and age 

• What staff did (called, emailed, sent letter) 

• Customer response 

• Outcome (payment commitment, dispute, no response) 

Example: 

• Customer: ABC Company 

• Invoice: $12,500, 25 days overdue 

• Action: Called customer, spoke to AP manager 

• Response: "We'll pay next week, expecting large payment from our customer" 

• Outcome: Payment commitment for 7 days, followed up, payment received 

Purpose: AI learns typical scenarios, appropriate responses, expected outcomes 

Edge Cases 

What's needed: 5-10 examples of unusual or complex situations 

Examples: 

• Customer bankruptcy filing 

• Disputed invoice requiring investigation 

• Payment misdirected to wrong account 

• Merger/acquisition changing payment entity 

• Natural disaster affecting customer 

Purpose: AI learns to recognize situations requiring human escalation 

Data Category 2: Decision Rules 

Explicit Decision Criteria 

What's needed: Your business rules documented 

For collections: 

• When to call vs. email (e.g., >$10K balance always call first) 

• Payment plan limits (e.g., up to 3 months for balances under $25K) 

• Escalation thresholds (e.g., disputes over $5K to AR manager) 

• VIP account definitions (e.g., >$100K annual revenue, strategic partners) 

• Communication frequency (e.g., contact every 7 days until resolved) 

Format: If-then rules or decision trees 

Example: 

IF balance > $10,000 AND overdue > 30 days 

  THEN call before email 

IF customer requests payment plan 

  AND balance < $25,000 

  AND plan duration <= 3 months 

  THEN approve 

  ELSE escalate to controller 

Implicit Knowledge 

What's needed: Unwritten rules staff follows 

Discovery process: 

• Interview staff about decision-making 

• Ask "how do you decide..." questions 

• Document reasoning patterns 

• Identify judgment criteria 

Examples discovered: 

• "I'm more flexible with long-time customers" 

• "If they've always paid eventually, I give more time" 

• "Construction companies I expect seasonal payment patterns" 

• "If they're combative, I escalate immediately" 

Challenge: Converting implicit knowledge to explicit rules 

Approach: Implementation partner facilitates workshops, asks probing questions, documents patterns 

Data Category 3: Customer Segmentation 

Account Classification 

What's needed: How you categorize customers 

Common segments: 

• VIP/Strategic (require special handling) 

• Standard commercial (typical process) 

• Small business (may need more flexibility) 

• Government (specific payment processes) 

• International (payment method differences) 

For each segment: 

• Defining criteria 

• Process variations 

• Special considerations 

• Escalation rules 

Contact Preferences 

What's needed: How customers prefer communication 

Data to provide: 

• Phone vs. email preference (if known) 

• Best time to contact (if known) 

• Decision-maker identification 

• Language preferences 

• Any special instructions 

Often this data doesn't exist systematically: Implementation identifies what's available, works with what exists, flags gaps for future enhancement 

Data Category 4: Outcome Definitions 

Success Criteria 

What's needed: What constitutes successful resolution 

For collections: 

• Payment commitment with specific date 

• Payment received 

• Dispute resolved with credit/adjustment 

• Payment plan agreement 

For vendor bills: 

• Bill approved for payment 

• Discrepancy resolved 

• PO amended to match bill 

• Bill rejected with documentation 

Purpose: AI knows when exception is resolved vs. requires follow-up 

Follow-Up Requirements 

What's needed: When and how to follow up 

Rules to document: 

• Payment commitment follow-up (e.g., day after commitment date) 

• No response follow-up (e.g., 7 days after initial contact) 

• Dispute follow-up (e.g., 3 days after escalation) 

• Payment plan monitoring (e.g., verify payment received each period) 

Data Category 5: Communication Templates 

Existing Templates 

What's needed: Current email templates, letter templates, scripts 

Purpose: 

• Understand tone and messaging 

• Identify key information communicated 

• Maintain brand voice consistency 

What to provide: 

• Collection reminder emails 

• Payment plan offer letters 

• Dispute resolution communications 

• Follow-up messages 

AI will adapt, not copy: Templates provide guidance, AI generates contextual messages 

Tone and Voice 

What's needed: Communication style preferences 

Considerations: 

• Professional vs. friendly 

• Firm vs. understanding 

• Formal vs. conversational 

• Industry-specific norms 

How documented: 

• Examples of good communications 

• Examples of poor communications 

• Staff feedback on tone 

• Customer feedback if available 

Data Gathering Process 

Week 1: Discovery Workshop 

Participants: 

• Controller or AR/AP manager 

• Staff who handle exceptions 

• Implementation partner 

Activities: 

• Current process walkthrough 

• Decision rule documentation 

• Edge case discussion 

• Customer segment identification 

Deliverables: 

• Process flow documentation 

• Initial decision rule matrix 

• Exception scenario list 

Time required: 6-8 hours spread over 3-4 sessions 

Week 2: Data Compilation 

Staff activities: 

• Export 20-30 recent exception examples from ERP 

• Gather existing email templates 

• Document implicit rules identified in workshop 

• Identify VIP account list 

Implementation partner: 

• Synthesize workshop notes 

• Create decision tree drafts 

• Identify data gaps 

• Prepare for review session 

Time required: 4-6 hours staff time 

Week 3: Review and Refinement 

Activities: 

• Review decision rules with staff 

• Test rules against historical scenarios 

• Refine edge case handling 

• Finalize escalation criteria 

Deliverables: 

• Complete decision rule matrix 

• Escalation criteria 

• Customer segmentation 

• Communication templates 

Time required: 2-3 hours review session 

What Data You DON'T Need 

Complete Historical Database 

Not needed: Every exception from past 5 years 

What's sufficient: 20-30 representative recent examples (last 6-12 months) 

Why: AI learns patterns, not memorizes history 

Perfect Data Quality 

Not needed: 100% clean, complete data 

What's acceptable: 

• Some missing contact information 

• Incomplete notes on past interactions 

• Gaps in documentation 

Approach: Work with available data, improve during pilot 

Fully Documented Processes 

Not needed: 200-page process manual 

What's sufficient: 

• High-level process flow 

• Key decision criteria 

• Escalation rules 

• Edge case examples 

Why: Over-documentation creates analysis paralysis. Basics sufficient to start, refinement during pilot. 

Data Sources 

From ERP 

Extractable: 

• Customer master data 

• Transaction history 

• Payment history 

• Account notes (if documented) 

How to extract: 

• Standard reports 

• Data export 

• API queries (if testing integration) 

From Staff Knowledge 

How to capture: 

• Discovery workshops 

• Interview sessions 

• Observation (shadow staff handling exceptions) 

• Documentation review 

Time required: 6-10 hours over 2-3 weeks 

From Existing Documentation 

What to gather: 

• Email templates 

• Process documentation (if exists) 

• Training materials 

• Any decision matrices or approval workflows 

Often limited: Many companies have minimal documentation. That's normal and acceptable. 

Data Privacy Considerations 

What Implementation Partner Sees 

During data gathering: 

• Customer names and contact info 

• Invoice amounts and ages 

• Communication examples 

• Payment history 

Protections: 

• Non-disclosure agreement (NDA) 

• Data encryption in transit 

• Deletion after implementation 

• No sharing with third parties 

Sensitive Data Handling 

Exclude from examples: 

• Social security numbers 

• Bank account details 

• Credit card information 

• Any regulated PII not needed 

Include only: 

• Data necessary for exception handling 

• Customer business relationship info 

• Transaction details 

• Communication history 

When Data Is Insufficient 

Scenario: Limited Historical Examples 

Problem: Only have 5-10 exception examples 

Solution: 

• Use those 5-10 as starting point 

• Staff provides hypothetical scenarios 

• Test with live data during pilot 

• Refine based on actual results 

Impact: 2-3 weeks additional refinement during pilot 

Scenario: Undocumented Decision Rules 

Problem: "We just know what to do" 

Solution: 

• Implementation partner facilitates discovery 

• Ask "walk me through how you handled last exception" 

• Document patterns from examples 

• Create initial rules, test and refine 

Impact: 1-2 additional workshop sessions 

Scenario: Poor Data Quality 

Problem: Missing contact information, incomplete notes 

Solution: 

• Work with available data 

• Flag accounts missing info for manual handling initially 

• Data cleanup as secondary project 

• Expand AI coverage as data improves 

Impact: Lower initial automation rate (50% vs. 70%), improves over time 

The Reality 

AI agents need: Process examples (20-30 typical scenarios, 5-10 edge cases), decision rules (explicit if-then rules, implicit knowledge documented), customer segmentation (VIP identification, process variations), outcome definitions (success criteria, follow-up requirements), communication templates (tone, voice, existing templates). 

Data gathering process: Week 1 discovery workshop 6-8 hours, Week 2 data compilation 4-6 hours staff time, Week 3 review and refinement 2-3 hours. Total 12-17 hours over 3 weeks. 

Don't need: Complete historical database, perfect data quality, fully documented processes. Basics sufficient, refinement during pilot. 

Data sources: ERP exports, staff knowledge (workshops, interviews), existing documentation (templates, process docs). 

Insufficient data manageable: Limited examples use 5-10 plus hypotheticals. Undocumented rules facilitated discovery. Poor data quality work with available, expand coverage as improves. 

Implementation partner guides entire process. Not a solo staff effort. 

About the Author: This content is published by ERP AI Agent. 

Published: January 2025 | Reading Time: 7 minutes