How generative AI works as a “probability machine”
Artificial intelligence, and particularly generative AI (GenAI), has captured the imagination of leaders everywhere. As we look for ways to embrace GenAI, it’s important to understand the fundamentals; the underlying technology and software principles powering generative AI don’t operate like traditional software.
What do I mean by this? Traditional software tools are deterministic: given the same input, it will produce the same output every time. We expect this fundamental principle from the varied software applications we engage with on a daily basis and most end-users, who may not write a single line of code, expect this to be true.
Generative AI is not deterministic, and that difference has major implications for where AI adds value—and where it creates risks.
Think of generative AI’s core process as a probability machine. Given a specific token (or set of tokens), the tool finds the most likely next token that should be returned. In a single token scenario, this process could possibly be deterministic – given token A, token B is always the most likely next token. That would not be a technological advantage – existing software solutions already do that. Generative AI works because it takes a sequence of tokens and predicts the next most likely token based on previous tokens – this is called context, and gives rise to the probabilistic nature of the technology.
As AI models grow larger and process more tokens, each predicted word carries a small chance of introducing an error or hallucination into the context. Over long interactions or under heavy use, these small probabilities add up, making strange or unpredictable outputs increasingly likely. All it takes is one rogue, non-deterministic response in a system that should behave predictably to cause real business consequences.
How does this apply when making business decisions?
Traditionally, when making decisions around software for business use cases, we did not have to worry much about indeterminate outcomes. Many business processes are dependent on deterministic outcomes – think accounting and finance, order processing, manufacturing, billing, etc. As traditional software is deterministic, applying software solutions to these problem spaces was the obvious choice. Now with generative AI in the system, there are many business areas where decisions around software must be approached with careful consideration of integrating an entirely new kind of technology.
Let’s look at some business areas and examine the potential impact of applying probabilistic software instead of deterministic.
| Business Area | Sample Software | Traditional (Deterministic) | Probabilistic AI Solutions |
|---|---|---|---|
| Customer Service | Automated chat system | Customer is engaging with a “logic engine” that asks specific questions and can only return specific answers or do the actions it’s been programmed to do based on the customer’s input. | Customer is engaging with an AI engine, which will use the entire text of the conversation to inform its “decisions.” The AI can make more powerful and dynamic decisions to address the customer’s inquiry. |
| Risk Analysis | Application Review Tools | A rules engine is applied to financial applications to look for specific markers of fraud or high risk. Applications must match the programmed values to get flagged. | An AI tool is given instructions and examples of what fraud and high risk look like. Applications that would not otherwise have met the criteria could be flagged. |
| Accounting | Deposit Revenue Recognition | A programmed process matches a bank deposit amount with the invoices the deposit clears. All the invoice totals need to sum up to the deposit, and discrepancies need to be dealt with in specific ways. | An AI system can analyze contextual patterns to suggest potential matches or highlight anomalies. It may surface connections that deterministic rules would miss but also introduce risk of misclassification. |
| Reporting | Subscriber Churn Prediction | A report is generated based on a set of known rules to predict what subscribers are at risk of cancelling. | An AI system learns from historical behavior and signals to probabilistically predict churn. This can uncover risks outside of known rules but may vary in accuracy. |
| Supply & Logistics | Inventory Management | A rules-based system tracks stock levels and triggers replenishment when certain thresholds are met. | An AI tool uses demand forecasting, seasonality, and external signals (shipping delays, promotions, macro trends) to predict inventory needs. It can optimize stocking dynamically, but predictions may shift unexpectedly. |
What is the “So What” here?
In some of these examples, the ability of AI to generate answers outside the bounds of a normal deterministic response proves to be an advantage – the business value and utility of the software improves.
However, losing the determinism of the software could create significant risk to the business – introducing possible outcomes that were never on the radar of that business unit before. This highlights the complicated nature of the new decisions facing business leaders.
Let’s review the previous examples and see how the decision is impacted by the introduction of AI in each of these scenarios:
| Software System | Deterministic → Probabilistic Impact |
|---|---|
| Automated Chat | Do we want to introduce the possibility that customers could get resolutions out of our chat bot that we have not explicitly allowed? |
| Application Review Tools | Are we prepared for the system to flag or reject applications that don’t strictly meet predefined criteria but “look suspicious” according to AI patterns? This could reduce missed fraud but may also frustrate legitimate applicants. |
| Deposit Revenue Recognition | Are we comfortable with AI making probabilistic matches that could incorrectly align deposits and invoices? It may simplify or accelerate reconciliation but could also create accounting errors that have both fiscal and regulatory consequences. |
| Subscriber Churn Prediction | Can we rely on more variable predictions when allocating resources to customer retention? The AI may uncover hidden churn risks but also produce false positives that could waste time and investment. |
| Inventory Management | Do we want to shift from simple reorder points to forecasts that adapt based on seasonality and external data? This could optimize stock levels but may also produce sudden, unexpected swings in supply planning. |
As we look for opportunities to apply AI, we should carefully review with the business new tradeoffs included in moving from predictable deterministic outcomes to variable probabilistic ones. We should ensure the business is prepared for the outcomes that can come from this decision.
Axian has been helping clients implement innovative technologies for over 30 years. Leverage Axian’s experienced staff to assist you with insights and industry practices for implementing GenAI.