Talk Society
for Business
Trust & compliance
Grace scores. A person decides. Here is exactly how.
Everything on this page is how the product is built, not a promise about outcomes. If a lawyer on your side wants the detail, this is the page to send them.
How Grace scores an interview
Evidence, or nothing.
-
Every point quotes the applicant
Each scored dimension must cite words the applicant actually said, and the quote is checked against the transcript. A quote that isn't there scores zero. Nothing is invented.
-
The score is added up, not awarded
The headline number is computed from the dimension scores. The model that ran the conversation is not asked to pick a number, and its suggested recommendation is capped by the score.
-
A skipped question is our fault, not theirs
If Grace never asked about something, that dimension is left out of the average and named on the card. A card with a gap can never read “advance” on its own.
-
Charm is not evidence
Warmth, fluency, enthusiasm and rapport with the interviewer are written into the rubric as things that do not count. A quiet applicant with specifics beats a delightful one without.
-
You see how she ran the call
How much she talked, whether she said who she was, whether she asked about recording before collecting anything. A measure of the interviewer, shown under every call.
-
Nothing about age, sex, race or health
No dimension asks about them and none may use a proxy for them. Anything an applicant volunteers about their health is not followed up and never used in the scorecard.
Decisions
There is no automatic rejection. There is no setting for one.
Grace ranks, summarises and recommends. Advancing, holding or declining an applicant is a button a named person on your team presses, and the record shows who pressed it and when. The only thing that ends a call early is a knock-out question you wrote — a licence, an availability, a distance — which is your stated requirement, not a judgement.
An applicant who would rather speak to a person can say so at any point; Grace ends the call politely and you are told. That call is not scored and not billed.
Disclosure and recording
She says she's an AI, then asks before she records.
In the first minute of every call Grace says who she is — an AI voice, built by Talk Society, interviewing for your company — and asks whether she may record. The line records from the moment it connects, so her question and the answer are on it; she waits for the answer before she asks anything else, and if the answer is no she ends the call and the recording and everything from it are deleted. Applicants are texted a written notice after the call.
Applicants ring Grace; she does not cold-call anyone.
Data
Kept for 180 days by default. Deleted on request.
Recordings, transcripts and scorecards are kept for 180 days and then deleted, unless the hiring company sets a different period for its own legal reasons (some U.S. states require employers to keep hiring records longer). Anyone we have spoken to can ask for their data to be deleted, and the deletion reaches recordings, transcripts, scorecards and Grace's memory of them.
Applicants often mention other people — a former employer, a family they cared for. Those third parties' names are removed from every summary and note before a person reads them.
Calls are carried by Twilio; the conversation runs on OpenAI's voice models; scorecards and summaries are written by Anthropic's models. Talk Society is the processor for the hiring company, which is the employer and the controller of its applicants' data. A data processing agreement is provided before launch.
To ask for your data, or to have it deleted, email hello@talksociety.app.
The law, in plain words
What the rules ask, and what the product does.
| Rule | What it asks | What Grace does |
|---|---|---|
| New York City Local Law 144 | Employers using an automated tool to substantially assist a hiring decision in NYC need an annual independent bias audit and must notify candidates. | Grace advises; a person decides. Every scorecard and transcript is readable over the API for an auditor. We do not collect race or sex data, and we make no claim that the scoring is free of bias. |
| Illinois (HB 3773, from January 2026) | Tell applicants when AI is used in hiring; no zip-code proxies for protected characteristics. | Disclosed in the first minute of every call and in writing afterwards. Nothing geographic enters the scoring. |
| California automated-decision rules | Notice before use, a way to opt out or take an alternative route, access and appeal, and longer record retention for employers. | Notice templates; the ask-for-a-person path; per-company retention up to four years; exports on request. |
| Federal law on AI voice calls (TCPA) | An AI voice is an “artificial” voice; calling someone with one needs their prior express consent, and the call should say it is AI. | Applicants call Grace. Any call she places is consent-gated and opens with her identifying as an AI. |
| Call-recording consent | Several states require every party's consent to record a call. | Grace asks, and waits for a yes, before anything else is collected; a no ends the call and deletes it. The console shows, for every call, at what point she asked. |
This is a description of how the product works, not legal advice. Your obligations depend on where you and your applicants are, and your own counsel should confirm them.
Questions for us