Deterministic and Probabilistic AI

Confusing confident answers with guaranteed outcomes

By Tony Burlinson

Understanding the distinction between deterministic and probabilistic systems is critical to a AI First strategy. Some firms are trying to introduce AI into situations where deterministic outcomes are expected, simply because it sounds impressive to say they are using AI.

Traditional software systems are largely deterministic. Given the same inputs, rules and conditions, they produce the same output. A calculation, a payment instruction or a territory assignment should produce the same predictable result. There is little or sometimes zero room for interpretation.

AI is a fundamentally different species.

Generative AI is probabilistic. It predicts likely responses based on patterns learned during training and the context provided. The same question can produce different answers, particularly over time. Moreover, those answers can sound equally convincing, even when one is incorrect.

The distinction between deterministic and probabilistic systems matters enormously. Especially in financial services and other highly regulated industries.

A firm doesn’t want an AI model deciding probabilistically whether the payment amount, account number or beneficiary is correct. The answers need to be either correct or incorrect with no room for interpretation. The output is deterministic.

Yet firms are increasingly looking at AI to automate these types of activities.

Almost every firm has let loose the AI kraken and encouraged – if not shoved – their employees into using AI to automate everything and anything.

AI promises fewer people, faster processing and lower operating costs. But the assumption that AI can simply replace existing deterministic systems is hugely dangerous.

The problem isn’t allowing employees to use and embrace AI. The problem is failing to understand where AI belongs in a firm’s technology architecture.

AI can profile a client relationship, extract information from an unstructured document, identify data anomalies. But systems that support workflows that require guaranteed outcome every time must remain deterministic.

There is an important architectural distinction between using AI to interpret information and allowing AI to make decisions that necessitate guaranteed outcomes.

Firms need to clearly define where the line is between their deterministic systems and their probabilistic systems and how they work together to achieve a coordinated business outcome.

A probabilistic AI system that gets something right 99% of the time sounds impressive in a demo. It likely also earns the AI agent creators extra points for building something in AI in short order, which looks flashy and resonates with senior leaders. It is, however, inherently and extremely dangerous if the output needs to be deterministic.

AI should augment deterministic systems, not replace them.

Firms that fail to understand where the deterministic and probabilistic line lies in their architecture are building tomorrow’s operational disasters on today’s AI enthusiasm.

Tags: AI First
Share: LinkedIn