The Question We Ask Before “Should We Add AI”
Most AI feature requests we see start from the technology, not the problem: “we should have a chatbot,” or “can we add AI to this.” The ones that actually ship and get used start from a specific, recurring point of friction a real user hits — and AI turns out to be one possible way to remove it, not the goal itself.
A Simple Filter That Saves Months
- Is this a task a human is already doing manually, repeatedly, in a pattern that’s learnable from data you actually have?
- Is the cost of an occasional wrong answer tolerable, or does this need to be right 100% of the time?
- Could a simpler rule-based system solve 80% of this without the complexity and cost of a model?
If the answer to that third question is yes, that’s usually the right place to start — AI earns its place later, once the simpler version proves the problem is real and worth the added complexity.
Where It’s Worked for Our Clients
The AI features that have actually stuck for our clients are narrow and unglamorous: auto-categorizing support tickets, summarizing long documents before a human reviews them, flagging anomalies in data a person would otherwise scan manually. None of them replaced a person’s judgment — they removed the boring 80% of the task so the judgment could go toward the interesting 20%.