Ask people at a research site what AI looks like, and many will picture a chatbot. You type a question, and it gives you a paragraph in return. (Handy when an amendment needs decoding at the end of a long day.)
Spend a day alongside a study coordinator, though, and another side of the work comes into view. Much of it isn't writing. It's making small, repeated decisions. Is this referral worth a call? Does the patient meet the age cutoff? Did the visit happen within the allowed window? Does the dose in the EDC match the visit note?
These questions often have a limited set of possible answers, and they come up dozens of times a day. That's where decision models may help. To understand where they fit, it helps to compare them with the chatbots many teams already know.
Two approaches, two different jobs
Fast chatbots, such as Claude Haiku, DeepSeek Flash, or Gemini Flash, are useful when a task involves language. They can read, summarize, explain, and draft, generating a response in natural language.
A decision model takes a different approach. Rather than drafting a response, it selects from options defined in advance, such as “eligible,” “not eligible,” or “needs review.” It may also return a score or confidence estimate. Jev, released by TypeSafe AI in September 2026, is one example of this approach.
The distinction is about the kind of output each system produces, not a guarantee that one is more accurate or safer than the other. A structured answer can still be wrong.
| Comparison | Decision model (e.g., Jev) | Fast chatbot (e.g., Claude Haiku) |
|---|---|---|
| Typical output | A choice from predefined options, sometimes with a score or probability | Written text, or structured data when requested |
| Output format | Designed to return one of the specified options | Can occasionally return unusable output; about 1.4% of Haiku responses in one test |
| Accuracy in one reported comparison | About 66% (Jev) | About 94% (Haiku) |
| Speed | About a third of a second per medical question in the reported test | Varies with the length and complexity of the response |
| Cost | Fractions of a cent per query, as reported | Low, but typically depends on input and output volume |
| Potential fit | High-volume tasks with a defined set of choices | Reading, summarizing, drafting, and tasks requiring nuance |
These figures come from a specific comparison, not a universal ranking. Performance will depend on the task, the data, and how each system is evaluated.
Different strengths, one workflow
It's more useful to see these systems as tools for different jobs than as direct competitors. A decision model may be suited to an initial check against inclusion criteria, flagging diary entries that may mention an adverse event, or identifying a possible mismatch between an EDC dose and a visit note. A chatbot may be better suited to extracting details from a long report into a table, drafting an initial SAE narrative, or simplifying protocol language for review.
In either case, human oversight remains essential. A chatbot's draft needs review before it is used, and a decision model's flags need checking. That should include sampling cases it did not flag, too; otherwise, a system can appear to perform well while quietly missing important cases.