Start with the decision, not the model
The first question is what decision or task the AI feeds, and what happens when it is wrong. A summary that a staff member checks is a different risk from a system that decides who gets a benefit. That answer sets how much automation is appropriate and where a person has to stay in the loop.
Our approach
- Deterministic first. If ordinary code can do the job, ordinary code does it. Models handle only the cases code cannot.
- The smallest model that works. Smaller models are cheaper, faster, and easier to evaluate.
- Evaluations before launch. We build a test set from real cases, score the feature against it, and keep scoring it after launch.
- Guardrails. Constrained outputs, careful handling of personal data, and clear behavior when the system is unsure.
- Human review where it matters. The review step is part of the workflow, not an afterthought.
- Cost and abuse limits. Public AI endpoints get rate limits, alarms, and a way to shut them off.
A working example: WelcomeOath
In our own product, WelcomeOath, learners answer civics questions out loud. Speech is transcribed and checked against every accepted answer by code. Only when a single-answer question misses does a small model, Amazon Nova Micro, judge whether the answer was close enough. Most answers never reach a model, which keeps grading fast, predictable, and inexpensive.
What we avoid
- Letting a model make a consequential decision with no one reviewing it.
- Sending sensitive records to a provider before checking its data terms.
- Launching without a way to measure whether the feature is getting better or worse.
Common questions
- How do you know an AI feature is ready to launch?
- We build an evaluation set from real cases before launch and score the feature against it. The score, the failure cases, and the guardrails are reviewed with your team, and the evaluation keeps running after launch.
- Where do humans stay in the loop?
- Wherever a wrong answer has real consequences. We design the review step into the workflow so people see what the system proposed, why, and how confident it was.
- Which AI provider and models do you use?
- The smallest model that meets the requirement, from a provider whose data terms fit your constraints. We check how each provider treats your data before choosing one.
- Can you review an AI proposal we already received?
- Yes. An independent review of a vendor's AI claims is a common starting point. See our second opinion service.