01
Start with a decision, not a model
The useful question is rarely “where can we add AI?” Start with a recurring decision, delay, exception, or information bottleneck. Describe who owns it, what evidence they use, what happens when they are wrong, and what a better outcome would change.
This framing prevents a capable model from being attached to work that has no stable owner, no accessible data, or no measurable definition of improvement.
02
Score the opportunity across five constraints
Evaluate volume, value, data readiness, acceptable error, and integration feasibility. High volume alone is not enough. A task can be frequent but cheap, or valuable but too safety-critical for unsupervised output.
The strongest early candidates usually combine repeatable inputs, visible review, a meaningful time or quality cost, and a system where recommendations can be tested before they become actions.
03
Separate assistance from authority
Drafting, summarizing, retrieval, classification, and recommendation can create value without giving a model final authority. Design the human checkpoint deliberately: who reviews, what evidence is shown, and how disagreement improves the system.
For low-risk, reversible steps, more automation may be justified after performance is measured. For high-impact decisions, a persuasive response is not a substitute for qualified judgment.
04
Prototype the risky behavior
Do not begin by building every integration. Create a representative evaluation set, include normal and adversarial cases, compare providers or approaches, and measure quality, latency, and cost. The prototype should answer a decision about feasibility, not merely produce a demonstration.
05
Design the operating system around the AI
Production readiness includes source freshness, access control, evaluation, monitoring, fallback, incident handling, and a way to explain or reproduce important outputs. Those product and operational layers are usually where business value becomes dependable.