The conversation around AI often treats autonomy as a measure of progress. If an AI agent can complete more work without involving a person, the assumption is that the system has become more capable.
For collections, that conclusion is incomplete.
McKinsey reported in 2025 that 88% of surveyed organizations were using AI in at least one business function, yet only 7% had fully scaled the technology across their organizations. The gap highlights an important distinction between having access to AI and building an operating model that can support it at scale.
For collection organizations, one of the most important decisions in that operating model is how much freedom an AI agent should have.
Too little autonomy can undermine the economic value of automation by sending excessive volumes of interactions back to employees. Too much autonomy can allow conversations to move into circumstances the system was not designed to handle.
The goal should always be appropriate autonomy.
The Consumer Makes AI Autonomy Complicated
An organization can determine how an AI agent opens an interaction, what information it can access, which actions it can perform, and what policies should govern its responses.
It cannot determine what the consumer will say next.
Consumers introduce unusual circumstances, incomplete information, emotional responses, unexpected questions, and subject changes. They may move a routine interaction into a sensitive situation within seconds.
Nir Laznik of Sedric.ai described this challenge during a recent appearance on the Receivables Podcast. He said organizations can control one side of an AI interaction, but the consumer remains an uncontrolled variable.
That distinction is fundamental to designing AI for collections.
Testing can evaluate thousands of expected scenarios, but production environments inevitably introduce edge cases. The governance question therefore cannot be limited to whether an AI agent has been trained for every possible conversation.
No practical system can anticipate every possible interaction. The better question is whether the system knows what to do when it reaches the edge of its capabilities.
The Degrees-of-Freedom Trade-Off
AI agent design involves a basic trade-off.
Restricting an agent's degrees of freedom can reduce the range of unexpected outputs. The system operates within narrower conversational and operational boundaries, potentially lowering hallucination risk.
The cost is flexibility. More conversations may reach circumstances the AI cannot resolve within those boundaries. Those interactions then need to be transferred to employees.
Expanding the system's degrees of freedom produces the opposite effect.
The AI may resolve a broader range of interactions without human assistance, improving automation rates and potentially reducing operational workload. At the same time, greater flexibility creates more opportunities for unexpected behavior.
This means hallucination and human escalation are connected design variables. The right balance depends on the workflow.
An AI Autonomy Ladder for Collections
A practical way to evaluate this balance is through an AI Autonomy Ladder.
Rather than classifying a system as either automated or human-operated, organizations can think about autonomy as a progression.
At the lowest level, AI supports an employee without communicating independently with the consumer. It may retrieve information, summarize interactions, or suggest responses, while the employee retains decision-making authority.
At the next level, AI can perform defined communication tasks within narrow boundaries. The system handles predictable interactions but escalates when the conversation moves outside predefined conditions.
A higher level allows the AI to manage broader conversational variation and complete more complex workflows independently, with monitoring and escalation controls operating in the background.
At the highest level, AI would have broad authority to navigate interactions and make decisions with minimal human intervention.
Not every workflow should move to the top of the ladder. That is the critical point.
The greater the potential consequence, the stronger the case for tighter boundaries, more intensive oversight, or earlier human escalation.
This creates a risk-based approach rather than a technology-based approach.
The question shifts from "How autonomous is our AI?" to "How autonomous should AI be for this particular interaction?" That is a far more useful operating question.
The Best AI Strategy May Be Selective Autonomy
The next phase of AI in collections should not become a race toward completely autonomous consumer interactions.
A stronger strategy is selective autonomy.
Give AI greater freedom where the workflow, evidence, and control environment support it. Maintain narrower boundaries where consumer impact or regulatory sensitivity is higher. Escalate deliberately when human judgment provides greater value.
That approach recognizes that AI and people do not need to compete for ownership of every interaction. The operating model can assign each the work it is best positioned to perform.
The objective is not to reach the highest rung. It is to operate at the right one.
For more analysis on artificial intelligence, compliance, technology, and operational strategy across the receivables industry, explore the industry resources available at ReceivablesInfo.com.
About Adam Parks
Adam Parks has become a voice for the accounts receivable industry. With almost 20 years working in debt portfolio purchasing, debt sales, consulting, and technology systems, Adam now produces industry news, hosts hundreds of episodes of the Receivables Podcast, and manages branding, websites, and marketing for over 100 companies within the industry.
Sign in to leave a comment.