One habit separates engineering from trial and error: prediction.
Before a bridge is built, engineers predict how much weight it can support. Before an aircraft flies, engineers predict how it will behave. Before a rocket launches, engineers predict its trajectory.
Yet in software engineering, production changes are often deployed without anyone writing down what they expect to happen.
That should strike us as strange.
Every production deployment makes a claim about reality.
This will improve performance. This will reduce cost. This will make recovery easier. This will reduce operational complexity.
Those are hypotheses. Unfortunately, many engineering organizations never state them precisely enough for reality to evaluate them.
A vague claim like "performance will improve" is easy to approve and hard to evaluate.
A better prediction is specific:
> P95 API latency will decrease by at least 25% under normal weekday traffic without increasing infrastructure cost.
That sentence immediately improves the review. It raises questions about measurement, baseline, scope, side effects, and time window.
The prediction does not merely define success. It improves the engineering discussion before implementation begins.
Predictions should evaluate models, not punish people.
If a prediction fails, the organization learned something. The system behaved differently from the team's mental model. That is valuable information.
Organizations that punish wrong predictions eventually stop getting honest predictions. Organizations that reward learning become better at understanding complex systems.
AI will generate plausible changes at extraordinary speed. The bottleneck will not be implementation. It will be deciding which recommendations deserve confidence.
Predictions create a reviewable standard. An AI proposal that includes measurable expectations, confidence, evidence, and failure conditions is much more useful than one that merely produces code.
Every significant change should answer:
Those questions transform a deployment into an experiment.
The code changes production. The prediction changes the engineers who build the next system.