How the AI works
The model does the reading. The structure makes it fair.
Slate is AI-native because reading unstructured applications against a role, with reasons, is only possible with modern language models. Here is what that actually means, in plain terms, including the limits.
From a pile of applications to a fair read
Language models do the reading
An application is prose in a hundred different layouts. Reading it against a role and saying, with reasons, how it lines up is exactly what modern language models are for. No brittle keyword rules that drop good people who phrased things differently.
Structured, and constrained to be fair
The read is not a vibe. Slate works the role's must-haves one at a time, cites evidence for each, and is instructed to keep name, age, gender, school and employer prestige out of the judgment. The structure is what makes it consistent and auditable.
Model-agnostic by design
Slate is not tied to one model or one vendor. It routes each task across leading models, picking the one that reads a given application best for the cost. As models improve, the read gets sharper without you changing anything.
It does not exist without AI
Take the language models out and Slate is a folder of PDFs. The reading, the evidence, the fair summary, all of it is the model working over unstructured applications. Slate is AI-native because there is no version of this that is not.
What the model is actually asked to do
The read is not a vibe or a hidden score. It is a structured task with clear constraints, which is what makes it consistent and auditable.
Reads against the role
Slate is given the role and its must-haves, then reads each application in full against that standard. Not a keyword match, a read, the way a careful person would do it.
Works the rubric
It goes through the must-haves one at a time and marks each evidenced, partial, or unclear, based only on what the application shows. The same rubric for every candidate.
Cites the evidence
Every reason is required to quote a verbatim line from that application. If something is not evidenced, it says so, rather than assuming it.
Keeps the person out of it
The screen is instructed to ignore name, age, gender, race, nationality, and the prestige of a school or employer. It judges the work, not who the person is.
Tuned for consistency
Slate runs the read at a low temperature and returns structured output, so two similar applications get similar reads rather than a different answer each time.
Honest when unsure
A thin application gets a read that says so. Absence of evidence is a gap to verify with the candidate, never a silent mark against them.
A thin layer that routes across models
Slate is not tied to one model or one vendor. A thin model layer picks a model per task from a small registry, so the read can trade quality against cost and is never locked to a single provider. The default is a fast, capable frontier model, which is the right tool for structured reading and evidence extraction.
This matters for two reasons. First, the read gets better on its own as models improve, without you changing anything. Second, if a provider has an outage or changes its terms, Slate can route around it. The reading is the product. Which model does the reading is an implementation detail we tune.
What it does not do, and where we are honest
Language models can carry bias from the data they were trained on. We do not claim to have removed it. We reduce it by constraining the screen to job-relevant evidence, applying one rubric to everyone, and making every reason auditable so a person can check the work.
Slate does not make hiring decisions, and it never auto-rejects. It produces a fair first read and surfaces the reasons and the gaps, and a person on your team decides. When the model is unsure, it says so. When the evidence is absent, that is a question to ask the candidate, not a verdict.
You can read more about the fairness stance and how your data is handled.
See the read for yourself.
Start free, paste a role and a few applications, and watch the evidence-based screen run.