Google Cloud exec on software’s great reset and the end of certainty: we’re moving from predictability to probability

We’re presently seeing the largest clash in software’s history.

You’ve probably come across the words “deterministic” and “probabilistic” in AI conversations, but what do they actually signify? And how does this impact your business?

On one hand, there’s the deterministic model. This is the framework we’ve used to build software and businesses for half a century. Every software tool you’ve purchased – from your CRM platform to a simple spreadsheet – has been exact, governed by rules, and zero-tolerant of mistakes. Input A plus Input B always results in Output C. If that’s not the case, you have a bug that requires fixing.

On the other hand, Generative AI defies this rule. It’s probabilistic, creative, and context-aware. The same inputs can yield different outputs. It’s a reasoning tool, not a calculator.

This lets you pose questions that deterministic systems can’t address. What impact will tariffs have on my revenue this year? How would a conflict in the Taiwan Strait affect my commodity prices? These kinds of questions don’t have definite answers, but foundational AI models can sift through enormous datasets and simulate various outcomes to guide your choices.

The tension you’re experiencing in your operating model right now – whether it’s around compliance or quality control – stems from the fact that our business systems were designed to root out and remove uncertainty. But you can’t squeeze a probabilistic engine into a deterministic operating model. To unlock Generative AI’s full potential, leaders need to stop viewing AI as a speedier spreadsheet.

The companies that thrive in this new era will be those that stop trying to squelch uncertainty and start integrating it into their operations. Below are three changes needed to reconfigure your business and fully leverage the AI-driven future.

Measure Autonomy, Not Just Efficiency

In the deterministic world, software value is gauged by access (number of user seats) and efficiency (how quickly a person can work). We saw software as a tool to enhance human workers.

Generative AI turns this model on its head. We’re shifting from software-as-a-service to “service-as-software,” where value comes from the result, not the tool itself. If an AI agent writes a legal brief or resolves a customer issue, the metric isn’t how much time a human saved using the software – it’s whether the human needed to be involved at all.

This calls for different metrics. We need to stop tracking effort and start tracking autonomy. Was the AI agent consistently accurate? Did it cut down decision-making time? What’s the task completion rate? And the metric that’s most critical for boosting margins: Did the AI agent fix the issue without any human help? The aim isn’t a faster workforce – it’s a workforce that can scale endlessly because the bottleneck (humans) has been taken out of the process.

Manage Uncertainty, Don’t Eliminate It

Most companies try to force probabilistic AI into deterministic, rule-based operating models. It doesn’t work. When traditional leaders see an AI model produce hallucinations (inaccurate outputs), they panic. They want to shut it down until it’s “100% accurate.”

But 100% accuracy is a deterministic myth. The correct approach is to surround the probabilistic engine with guardrails that manage uncertainty. At , we refer to “grounding” and confidence scores. Leadership teams need to stop asking “Is this answer correct?” and start asking “How confident am I in this output?” At Google, we teach our employees that AI agents aren’t built to provide answers – they’re built to generate reasoning.

To prevent this, we need to create systems where AI works independently when confidence levels are high, and seamlessly hands off to a human expert for review when confidence dips.

Just as Google’s AlphaFold provides confidence ratings for its protein structure predictions, your business’s AI should give leaders a score they can act on. These interventions become the feedback loop that trains the model and fuels ongoing improvement.

Turn Data Into Feedback, Not Just Facts

This technology doesn’t replace people – it changes their main role from execution to expertise.

In deterministic systems, data was a record of truth used for reporting and historical analysis. With Generative AI, data becomes immediate feedback and action. Your past data trains your future workforce – your array of autonomous AI agents. Disorganized data leads to an ineffective digital workforce.

This requires a transformation of the human role. In the deterministic world, we hired large teams of junior employees to handle repetitive tasks. In the probabilistic world, AI does the tedious work. It creates the first draft, writes the initial code, and conducts the baseline analysis – all in an instant.

We’re now witnessing an evolution. First, humans do the work while AI supports them. Then, AI does the work while humans oversee it, stepping in when necessary. Eventually, AI operates on its own while humans audit periodically. If a human has to approve every AI agent’s decision, you’ve just built an expensive spell-checker.

This leads to a major shift in talent needs. We don’t need people who can just perform tasks – we need people who can audit. AI can produce average work in an instant. You need humans who are expert enough to tell the difference between “great” and “good” in seconds. We need editors-in-chief. We need talent with the expertise to look at an AI output and immediately spot the difference between “plausible” and “brilliant.” The era of learning through tedious work is over – we need to build an era of learning through judgment.

The Sailboat and the Train

The biggest competitive edge will go to leaders who can handle ambiguity in return for exponential speed.

Here’s a way to think about it: For decades, we’ve been building faster trains. Trains run on tracks (rules). They’re efficient, predictable, and go exactly where you plan. Today, we’re building sailboats. They depend on the wind (probabilistic data) and can reach places tracks can’t. But without a rudder (guardrails) and a compass (ground truth), you’ll capsize.

Leaders who insist on 100% certainty will stay stuck in the past, refining the efficiency of a fading model. The future belongs to those who learn to embrace and navigate probability.