The Problem with Current AI Agents
A problem set that governance frameworks and data access alone cannot solve.
At Moltin, our AI agents are rapidly moving from team-driven experimental prototypes to fully supported and operationalized production systems handling real business workflows. We are deploying agents that approve expense reports, manage the support ticket life cycle, route and respond to sales leads, and make countless other operational decisions. Yet despite the excitement and investment, these agents are hitting an invisible wall. A wall that governance frameworks and data access alone cannot solve.
The Data Paradox
At first glance, modern AI agents appear to have everything they need. They connect to CRMs, ERPs, support platforms, and data warehouses. They can read customer records, check shipment ETAs, review past transactions, and pull up policy documents. The data is there. The integrations work. So why do agents still struggle to make reliable decisions in complex scenarios?
The answer lies in what's missing: decision traces.
While organizations meticulously capture what happened (every transaction, status change, field update etc.), they systematically fail to capture why it happened. The context that justified a decision, the exceptions that were made, the competing priorities that were balanced—all of this evaporates the moment a decision is executed.
When Decisions Become Black Boxes
Consider a common scenario. A customer support agent (human or AI) needs to decide whether to escalate a refund request that exceeds standard policy limits. The decision gets made. The refund gets processed. The ticket gets closed. But what happened to the reasoning?
Perhaps the customer was a high-value enterprise client in the middle of a renewal negotiation. Maybe there was a known product bug affecting their account. The support manager might have considered recent churn rates in that customer segment, or referenced a similar exception made last quarter for a comparable situation. All of this context informed the decision, but none of it gets captured in the systems of record.
Six months later, when a similar case appears, that institutional knowledge is gone. The AI agent can see that an exception was made, but it cannot understand why the exception was justified. It cannot replay the state of the world at the moment of decision. It cannot use past precedent to guide current judgment.
The Audit Problem
This absence of decision traces creates a cascade of problems. The most immediate is auditability. When an AI agent makes a controversial decision, stakeholders inevitably ask: "Why did the system do that?" Without decision traces, there's no good answer. You can examine the current state of the data, but you cannot reconstruct what the agent knew or considered at decision time.
Was the customer tier different then? Had the SLA been recently updated? Was there an active incident affecting service levels? These contextual factors might have made the decision entirely reasonable, but if they're not preserved, the decision appears arbitrary or wrong in hindsight.
The Learning Problem
The lack of decision traces also prevents organizational learning. Every decision an organization makes (especially edge cases and exceptions) represents valuable information about how policies should adapt to reality. But without capturing why decisions were made, this learning cannot happen systematically.
Organizations end up in a loop where the same exceptional cases require the same manual intervention repeatedly. The knowledge exists in someone's head or in a Slack channel, but it never makes it into the systems that guide future decisions. AI agents, no matter how sophisticated, cannot learn from precedent they cannot access.
The Cross-System Context Challenge
The problem becomes even more acute when decisions depend on information scattered across multiple systems. A purchasing approval might hinge on budget data from the ERP, project timeline information from the project management tool, vendor performance metrics from the procurement system, and strategic priorities discussed in executive meeting notes.
Each system captures its slice of the truth, but no single system sees the complete picture that justified the decision. The AI agent can query these systems individually, but it cannot reconstruct the synthesis of information that led to the original judgment. The connective tissue between data points (the reasoning that tied them together) never existed in machine-readable form.
Why Governance Isn't Enough
Many organizations respond to these challenges by adding more governance: approval workflows, guardrails, human-in-the-loop checkpoints. These controls are necessary, but they don't solve the underlying problem. Governance can prevent bad decisions from being executed, but it cannot help agents make better decisions in the first place. It’s like putting a fence around a problem rather than solving it.
More fundamentally, governance after the fact cannot capture decision context. By the time a decision reaches an approval step, the full reasoning behind it is rarely documented. The approver makes a judgment call based on their own knowledge and experience, and that context disappears just as surely as the original agent's reasoning.
The Path Forward… A System of Precedents
The solution requires a hard look at how we think about business systems. We need infrastructure that captures not just the results of decisions, but the complete context that justified them. This means preserving the state of relevant data at decision time, recording the factors that were considered, documenting the precedents that were referenced, and maintaining the links between related decisions over time.
This isn't about logging every AI model inference or storing verbose chain-of-thought outputs. It's about building a system that treats decisions as first-class entities. A system with its own lifecycle, its own relationships, and its own place in the organizational memory.
Only when we solve this problem can AI agents move beyond executing simple, predefined workflows to making the kind of nuanced, context-aware judgments that complex business operations require. The data is necessary but not sufficient. What we need is a record of how that data was interpreted, synthesized, and applied. A decision trace that makes institutional knowledge not just preserved, but actionable for the agents we're building today and tomorrow.