Skip to main content
TalentInitiator

How TalentInitiator uses AI

This page lists every place a model is used, what it receives, what it cannot do, how we test it, and how candidates ask for a person instead.

How to read this page

This page explains how TalentInitiator applies these rules. It is not legal advice for your organization. Our team checks every source again by November 30, 2026, and sources marked “Pending legal review” have not been confirmed by counsel yet.

Models and where they run

Each AI feature is assigned a tier. In our cloud service every tier runs on Anthropic’s Claude models in the United States, under a contract that excludes training on what we send. The exact model version is recorded with every call.

TierModelsUsed forFeatures
Fast tierClaude HaikuReading documents, drafting messages, screening turns and decision explanations.6 features
Balanced tierClaude SonnetScoring-criteria drafts, French drafts, matching, interview questions and interview evaluation.5 features
Senior tierClaude OpusOnly for jobs an administrator marks as senior or regulated, and only for scoring-criteria drafts and interview evaluation.2 features

The demo uses an offline evaluator with prepared results. It sends nothing to a model provider, and its badge reads “Demo mode · offline evaluation”.

Read where information is stored 

What AI does and what people decide

AI can

  • Read job descriptions and résumés into structured fields.
  • Suggest scores against criteria a recruiter approved, with quotes as evidence.
  • Draft messages, French versions and interview questions for a person to edit.
  • Ask scripted screening questions and record the answers.
  • Write the plain-language text of a decision a person already made.

People decide

  • Which criteria candidates are scored on, approved before anyone is scored.
  • Every advance, hold and rejection, each with their own written reason.
  • Every outreach message, approved before it is sent.
  • Every request a candidate makes for a review by a person.

What AI never does

  • It never advances, holds or rejects a candidate.
  • It never analyzes faces, voices, tone of voice, emotions or personality.
  • It never sees names, contact details, photos or dates of birth when it scores.
  • It never writes or changes legal statements and notices; those come from templates.
  • It never learns from candidates’ information.

What is hidden from AI

Before a profile is scored, a redactor removes identifying details. It uses fixed rules, not a model, and recruiters can open “Hidden from AI” on any application to see what was removed. Interview transcripts go through the same step.

Removed before scoring

  • Name, and its variants
  • Photo
  • Email address
  • Phone number
  • Street address
  • Postal code
  • Date of birth
  • Age
  • Graduation years
  • Exact dates, turned into durations
  • Names of schools
  • Nationality, a first language, and languages the job does not ask for
  • Citizenship or work-permit status
  • Links to social profiles
  • Words that indicate gender
  • Location more precise than the city
  • Family status, such as marital status or children
  • Religion or creed
  • Health and physical details

For example, April 2019 to August 2023 becomes “4 years 4 months”.

Kept because the job needs it

  • Job titles and employers
  • How long each role lasted
  • Skills, certifications and their status
  • The languages the job asks for, with a level rather than “native”
  • City or region, when the job has a location requirement

Reading a résumé is the one step that sends the whole document, because the model has to read it. The contact details it returns are encrypted at rest and never sent to a model again.

Bias monitoring

Scoring criteria guard

Criteria that rely on these stand-ins for protected grounds are rejected before anyone is scored:

  • Age
  • How recently someone graduated
  • National or ethnic origin, or first language
  • Canadian experience
  • Citizenship
  • Family or marital status
  • Gaps in employment
  • The reputation of a school
  • Religion or creed
  • Sex or gender
  • Health, disability or physical ability
  • A name or an accent

Self-identification

Candidates may answer voluntary self-identification questions. Answers are stored separately, with no link to scores or to the people deciding.

Adverse-impact report

Selection rates per group are computed from human decisions, never from model scores. An impact ratio below 0.8 shows a warning, and groups smaller than 5 are flagged as too small to read.

Checks before release

An evaluation set re-runs redaction, scoring-criteria drafts and matching on fictional candidates. A single leaked identifier, or a single stand-in accepted as a criterion, fails the check.

Model cards

One card per AI feature: what it is for, what it receives, the guardrails around it and how we test it.

  • Job description reading

    Model tier
    Fast tier
    Purpose
    Reads a pasted job description and fills in the job’s fields: title, location, shift pattern, pay when it is stated, and requirements.
    What it receives
    The job description the customer wrote. No candidate information.
    Guardrails
    It never invents pay: when the description gives none, the field stays empty for the recruiter.
    How we test it
    The recruiter reviews every field before the job is saved.

  • Scoring criteria draft

    Model tier
    Balanced tierSenior tier for senior or regulated roles
    Purpose
    Proposes job-related criteria and weights for scoring applicants.
    What it receives
    The job’s details and the organization’s scoring policy. No candidate information.
    Guardrails
    A guard with fixed rules rejects criteria that rely on protected grounds or stand-ins for them, physical ability included: a job’s physical tasks stay in the posting and the screening chat, never in a score. The recruiter approves the criteria before anyone is scored, and each approved version is kept.
    How we test it
    Ten sample jobs: no stand-in accepted, and no criterion without a job-related justification.

  • French draft of a posting

    Model tier
    Balanced tier
    Purpose
    Drafts the French version of a posting for the recruiter to edit.
    What it receives
    The English text of the posting.
    Guardrails
    Legal statements are inserted from templates, never translated by the model. Both languages publish together.
    How we test it
    The recruiter edits and approves the French text before publishing.

  • Second look at unclear wording

    Model tier
    Fast tier
    Purpose
    Classifies sentences the rule-based linter could not settle, such as possible Canadian-experience wording.
    What it receives
    Only the unclear sentences of a posting.
    Guardrails
    Called only for sentences the rules mark as unclear. Each finding is a suggestion the recruiter accepts or ignores.
    How we test it
    The recruiter reviews each finding next to its sentence.

  • Résumé reading

    Model tier
    Fast tier
    Purpose
    Turns a résumé into structured fields: experience, education, certifications, skills and languages.
    What it receives
    The document the candidate submitted.
    Guardrails
    Contact details it returns are encrypted at rest and never sent to a model again. A résumé it cannot read goes to a person.
    How we test it
    Thirty résumés in English and French, including two-column and scanned layouts: at least 95% of required fields correct in English and 92% in French.

  • Match against the criteria

    Model tier
    Balanced tier
    Purpose
    Scores a redacted profile against each approved criterion and quotes the evidence.
    What it receives
    The redacted profile and the approved criteria.
    Guardrails
    Quotes must appear word for word in the profile or they are dropped. Any mention of a protected ground fails validation. The result is a recommendation; it cannot move a candidate.
    How we test it
    Thirty profiles: every result matches its schema, no quote that is not word for word is kept, and scores stay within one point across three runs.

  • Outreach message draft

    Model tier
    Fast tier
    Purpose
    Drafts a first message about a job, in the candidate’s language.
    What it receives
    Public facts about the job and the candidate’s first name.
    Guardrails
    A recruiter approves every message. The sender identification and unsubscribe footer are added from a template the draft cannot change.
    How we test it
    The recruiter reads and approves each message before it is sent.

  • Screening chat

    Model tier
    Fast tier
    Purpose
    Words the next pre-screening question and records the answer.
    What it receives
    The candidate’s message, the current question and the approved job facts.
    Guardrails
    Questions come from a fixed script, so no question about a protected ground can be asked. When the posting states physical tasks, the chat asks only whether the candidate can do them, with or without accommodation. A “no”, or any limit the candidate describes, goes to a recruiter, who talks with them about accommodation; only a plain yes is recorded as an answer. Off-script questions are answered only from approved job facts. A sign of distress or a request for a person hands the chat to a recruiter.
    How we test it
    Twenty turns in English and French: at least 90% of answers recorded correctly, and every distress case handed to a person.

  • Interview questions draft

    Model tier
    Balanced tier
    Purpose
    Proposes up to two questions per criterion, in English and French.
    What it receives
    The approved criteria and the role.
    Guardrails
    No question that invites a candidate to disclose a protected characteristic. The recruiter edits the questions.
    How we test it
    The recruiter approves the questions before any interview.

  • Interview evaluation

    Model tier
    Balanced tierSenior tier for senior or regulated roles
    Purpose
    Scores interview answers against the criteria, with quotes, strengths, concerns and limitations.
    What it receives
    The redacted text of the transcript and the approved criteria. Never audio or video.
    Guardrails
    Quotes are checked against the transcript. Nothing about demographics, tone or emotion is inferred. The result sits beside the recruiter’s decision form, never inside it.
    How we test it
    Ten transcripts: every result matches its schema, quotes are word for word, and each one states its limitations.

  • “Why this decision” text

    Model tier
    Fast tier
    Purpose
    Writes the plain-language explanation of a decision from what the recruiter recorded.
    What it receives
    The stored decision: its factors and the information used.
    Guardrails
    It cannot add a factor that is not in the recorded decision. Simple cases use the template alone, without a model.
    How we test it
    Every explanation is checked against the keys of the recorded decision.

Human-rights impact assessment

We assess the product with the Human Rights AI Impact Assessment published by the Ontario Human Rights Commission and the Law Commission of Ontario. This is the summary.

Edition 1, October 9, 2026. Prepared by the product team; not yet reviewed by counsel or an outside assessor.Pending legal review
Purpose
Help recruiters at staffing agencies and high-volume employers sort and screen applications faster, without letting software decide who gets a job.
People affected
Job applicants, including newcomers to Canada, people with disabilities, French speakers and people returning to work after a gap.

Risks and how the product addresses them

RiskWhat we do
Scoring criteria that stand in for age, origin or disability.The criteria guard rejects them, physical ability included, and a recruiter approves the criteria before scoring. A job’s physical tasks are asked in the screening chat, with or without accommodation.
Recruiters defer to the score.There is no “accept recommendation” button, every decision needs the recruiter’s own written reason, and a reason that only restates the recommendation is refused.
Identity shapes the score.Identifying details are removed before scoring, and recruiters can see what was hidden.
Unequal outcomes go unnoticed.The adverse-impact report on human decisions, with a warning below the threshold.
The interview format disadvantages people with disabilities.Written answers by default, accommodations on request, and an interviewer who is a person at any time.
Newcomers are screened out by Canadian-experience wording.The posting linter and the criteria guard both catch it.

Monitoring

The adverse-impact report, the evaluation set before each release, and a look at every request a candidate makes for a review by a person.

Sources

How a candidate asks for a person

  1. Open the link in the message the employer sent you, then choose “Why this decision”.
  2. Read the factors and the information used, and correct anything that is wrong.
  3. Choose “Request a review by a person” and say what you would like looked at again.
  4. An administrator or the privacy lead at the employer reviews the decision and answers you in your portal.

Earlier steps have a person too: “Talk to a person” appears on every turn of the screening chat, and you can ask for an interviewer who is a person instead of recording answers.

You can also write to the employer directly at any time. The portal is a shortcut, not the only way.

Training data

  • Candidate information is not used to train AI models. Anthropic’s commercial terms exclude training on the information we send.
  • We do not train models of our own on customer data.
  • Transcription runs with the provider’s option to use content for service improvement turned off.
  • The examples we test models with are fictional candidates written for that purpose.

The statement in job postings

Ontario requires postings to say when AI screens applicants. See the statement the product inserts, in English and French.

See the statement