Uncategorized

A Reminder: AI Is Not Deterministic

0 Comments

While “AI”, and more specifically Large Language Models (LLMs), continue to improve rapidly in accuracy, speed, and reasoning capability, there is one absolute rule that must be understood before deploying them in production systems:

Large Language Models are not deterministic and will produce unexpected, inconsistent results at some non-zero error rate.

A deterministic system is one in which outcomes are never random; they are reliably derived from the inputs into the system. The same input produces the same output, every time. This property is foundational to traditional software, automation, and most enterprise systems executives have relied on for decades.

Consider the prompt:

“Write a short thank-you note to a colleague for helping with a project.”

Deterministic system (guaranteed output):

“Thank you for your help on the project. I appreciate your support and collaboration.”

Every time this prompt is executed, the output is identical. There is no interpretation or variation. The advantage here is absolute reliability; the downside is that it is rigid, uncreative, and often less useful in situations where tone or nuance matters.

Probabilistic system (variable outputs):

“Thanks so much for jumping in and helping with the project. I really appreciated your insight and teamwork.”

“I wanted to say thank you for your help on the project. Your support made a real difference, and I’m grateful for the collaboration.”

Here, the system is not retrieving a fixed answer; it is generating one. The intent is the same, the quality is acceptable, but the result is different each time. That variability is what makes probabilistic systems useful for creativity but unsuitable for tasks that require exact repetition.

Why Won’t You Survey My Customers?

Over the last year Salesforce provided a real-world example that illustrates why this distinction matters. At the end of 2024, the company announced that Vivint was adopting its Agentforce platform to provide 24/7 autonomous customer support. The promise was compelling: scalable, always-on AI agents handling customer interactions without human intervention. Roughly a year later, Salesforce executives publicly acknowledged trust issues with autonomous AI deployments and began scaling back certain uses.

One of the underlying challenges Vivint encountered was simple: sending customer surveys. The requirement was straightforward: after every customer interaction, a survey must be sent. Agentforce was instructed to do exactly that. Yet, despite those clear instructions, some surveys were missed. There was no obvious error; the LLM simply didn’t behave consistently.

The reported fix was Vivint introducing what they described as “deterministic triggers”. In other words they used traditional automation rules that guaranteed surveys were sent every time an interaction concluded.

This pattern highlights a broader architectural lesson that applies to AI use across the board and is candidly the only way AI can be incorporated into larger and more impactful workflows: some tasks are inherently deterministic and should remain so.

LLMs are extraordinarily powerful. They can summarize, classify, reason, converse, draft, and analyze in ways that dramatically reduce human effort and unlock new capabilities. However, they are best used behind deterministic guardrails. When a business requirement demands 100% reliability (billing, compliance notifications, patient identification, contractual communications, etc.) traditional automation should own the final decision or action.

From a systems architecture perspective, there are a few best practices that ensure tasks are being completed with the right level of determinism:

  • Separation of concerns: Deterministic systems manage state, triggers, and guarantees; LLMs provide intelligence, interpretation, and flexibility.
  • AI as an advisor, not the executor: The model proposes actions, drafts content, or classifies intent, while deterministic code decides whether and when to act.
  • Validation and control loops: In higher-risk workflows, a second model or rule-based system validates outputs before execution.
  • Graceful tolerance for error: LLMs are deployed freely in areas where a small error rate is acceptable or easily reversible.
  • Human-in-the-loop escalation: For critical paths (legal, financial, or safety-related) humans remain the final authority.

Even techniques such as “LLM-on-LLM” validation can only reduce error rates, not eliminate them. Probabilistic systems will always retain some degree of uncertainty.

The ultimate takeaway here is not “don’t use AI.” LLMs can deliver extraordinary value to businesses and society more broadly, but only when leaders understand where they belong in the system. The most successful organizations will be those that design architectures combining probabilistic intelligence with deterministic reliability.

Sources & References

This article draws on a combination of public reporting, industry announcements, and generally accepted concepts in AI systems design.

Leave a Reply

Your email address will not be published. Required fields are marked *