The old way to screen at scale is a rules engine. Match these keywords, require this degree, filter out anyone missing the phrase. It is fast, and it is quietly unfair, because it rewards people who wrote their application to match the filter and drops people who did the work but described it differently. Adding more rules does not fix this. It just makes the unfairness more elaborate.
The hard part is reading, in context
A fair first read is not string matching. It is reading a paragraph and understanding that managing a book of accounts and owning renewals for a set of customers are the same thing described two ways. It is noticing that a teacher who kept ninety percent of families engaged has evidenced the retention instinct the role needs, even with no SaaS on the resume. For a long time only a person could read like that, which is why fair screening did not scale.
Modern language models can do this reading. That is the specific capability that makes Slate possible, and it is why the model is the engine, not a feature bolted onto a filter.
- They read unstructured applications the way a person would, in context, not as keywords to match.
- They can check each application against a role's requirements and cite the exact evidence for each.
- They can write the read in plain language, so a reviewer sees the reasons, not a mysterious score.
AI-native does not mean unaccountable
Being AI-native is what makes the read possible. It is not a license to hide behind the model. Slate is constrained to screen on job-relevant evidence, to keep personal characteristics out of the read, and to attach a verbatim quote to every reason so a human can check its work. Models can carry bias from their training data, and pretending otherwise would be the dishonest move. The structure around the model is what makes it fair.
You cannot bolt fairness onto a keyword filter. The reading, against the role, with the reasons, is the whole job.Daniel Okafor
Model-agnostic, on purpose
AI-native does not mean tied to one model. Slate routes each task across leading models, choosing the one that reads a given application best for the cost. The reading is the product. Which model does it is an implementation detail we tune, and one that gets better on its own as the models do. Fair screening had to wait for models that could read. Now that they can, this is the obvious thing to build, and there is no version of it that is not AI-native.