What Can AI Agents Do Today (And What Can't They Do)?

The Expectation Problem 

Marketing materials describe impressive capabilities. Vendor demos show seamless operation. The gap between claims and operational reality creates implementation failures when expectations exceed actual capability. 

Mid-market companies need accurate understanding of what AI agents can and cannot do today to prevent misapplication and ensure appropriate oversight. 

What AI Agents Can Do Today 

Handle Standard Exception Processes 

AI agents execute exception processes following defined decision rules, coordinate communication with customers and vendors, document outcomes systematically, and escalate situations requiring human judgment. 

Agents review exception data, apply prioritization logic, initiate contact through phone or email, conduct structured conversations, document responses, and update ERP systems. 

Current Success Rates: 60-80% of standard exceptions handled successfully. 20-40% require escalation for complex situations or relationship sensitivity. 

Make Phone Calls That Sound Natural 

Voice AI conducts business conversations that customers and vendors accept as professional communication. Voice recognition handles standard English accents with 90%+ accuracy. Conversation quality is sufficient for routine business communication. 

Limitations: Heavy accents and multilingual conversations require specialized implementation. Voice AI does not handle emotional nuance well. Frustrated customers escalate to humans immediately. 

Apply Complex Business Rules 

Agents handle multi-factor decision logic.  

Example: Contact if (balance over $2,000 AND days overdue greater than 45 AND not flagged for special handling AND previous commitment broken) OR (balance over $5,000 AND days overdue greater than 30 AND no payment received in 60 days). 

Agents handle 10-15 decision factors simultaneously. Rule complexity limited by clarity of articulation, not technical capability. 

Document Everything Systematically 

Agents record complete details of every interaction immediately. Documentation consistency reaches 100% compared to 40-60% with manual handling under time pressure. 

Updates write directly to ERP systems through API. Information appears in customer notes, activity logs, or collection management fields. 

Identify Patterns Humans Miss 

Systematic documentation enables pattern analysis. Agents identify recurring customer payment issues, seasonal exception volume patterns, systematic vendor quality problems, and predictable back order triggers. 

What AI Agents Cannot Do Today 

Make Strategic Decisions 

Agents cannot set business policies, establish credit terms, approve exceptions to standard rules, or make judgment calls affecting company financial exposure. 

Example: Agent can request payment following defined scripts. Agent cannot decide to accept 50% payment settlement or waive late fees based on customer circumstances. 

Handle Complex Negotiations 

Agents cannot negotiate payment arrangements requiring multiple counteroffers, resolve disputes requiring investigation across systems, or create custom solutions for unique situations. 

Complex negotiations require investigation, authority to make concessions, and relationship context. Agents operate within defined parameters. 

Build Customer Relationships 

Agents cannot build rapport, understand unspoken context, recognize relationship history importance, or handle VIP accounts requiring personal attention. 

Relationship management requires historical knowledge, emotional intelligence, and discretion about when rules should flex. 

Respond with Emotional Intelligence 

Agents detect frustration or anger through tone analysis but cannot respond with empathy, de-escalate emotional situations, or handle customers requiring emotional support. 

Investigate Across Multiple Systems 

Agents operate within defined data sources. They cannot investigate problems requiring information from multiple systems, external research, or coordination across departments. 

Handle Situations Requiring Creativity 

Agents follow decision trees and defined logic. They cannot create novel solutions, adapt approaches to unique circumstances, or think outside defined parameters. 

The 60-80% Rule 

Most implementations achieve 60-80% success rates: 

60-80% of exceptions handled completely by the agent 20-40% of exceptions require human intervention 

This distribution is consistent across implementations following best practices. The success rate improves gradually as rules refine. 

Companies expecting 90-100% automation are disappointed. Companies expecting 60-80% automation with human oversight for complex situations find results meet expectations. 

Decision Framework: Capability vs. Need 

Your exception process has: 

• High volume (30+ monthly) 

• Definable decision rules 

• Standard conversation patterns 

• Systematic documentation needed 

• Human escalation acceptable for 20-40% of cases 

Then AI agent capability matches your need. 

Your exception process has: 

• Low volume (under 20 monthly) 

• Every situation requires unique expertise 

• Relationship context critical 

• Emotional intelligence required 

• 90-100% autonomous handling expected 

Then AI agent capability does not match your need. 

About the Author 

This content is published by ERP AI Agent, a consulting practice specializing in AI agents for mid-market ERP exception processes. 

Published: January 2025 Last Updated: January 2025 Reading Time: 6 minutes