Seven features in v1.0 use AI. This page is the user-facing summary of what each one does, what it sees, and where the deeper transparency posture lives.
Where AI is used
- Job description generation (
jd-gen) — when you generate a JD on the role wizard. - CV extraction (
cv-extract) — when a candidate uploads a CV; structured fields and competency tags are parsed out. - CV screening narrative (
screening-narrative) — after CV extraction; the deterministic match score is non-AI. - Behavioural narrative report (
behavioral-narrative) — after a behavioural assessment is completed. - Career-path narratives (
career-path) — on the career-paths surface, per person/target-role pair. - Simulation generation (
simulation-gen) — when authoring a custom job simulation; AI scaffolds the situation and branching options for HR review.
The exact prompt scaffolding for each feature lives in lib/ai/prompts/ in the codebase. Each file is a one-to-one mapping from a feature to its prompt template.
Which model
The current model is a leading large language model, consumed via API. The platform does not deploy its own model weights and does not fine-tune. The provider name is disclosed to enterprise clients during procurement due diligence.
The model version is recorded with every generation in the audit log, so any narrative or generated text can be traced to a specific model version after the fact.
What the model sees
For each feature, only the inputs needed:
- JD generation sees: the title, function, your seed description, and the required competencies you selected. No employee data.
- CV extraction sees: the parsed raw text of the uploaded CV and the workspace's competency catalogue. No data on existing employees or other candidates.
- CV screening narrative sees: the extracted CV fields and the role's required competencies. No data on other candidates or existing employees.
- Behavioural narrative sees: the deterministic per-symbol colour scores and the behavioural template metadata. It does not see the assessee's full profile, name, or historical narratives.
- Career-path narrative sees: the gap snapshot for the specific person/role pair. No data on other people.
- Simulation generation sees: the seed scenario and the target competencies. No data on candidates or employees — output is reviewed by HR before being published into the simulation library.
In every case, the model sees the minimum context needed for the task, no more.
What the model does not see
To pre-empt assumptions:
- No PII transmission beyond what is in the input. Names go in only when the input requires them (CV screening); they do not go in elsewhere.
- No cross-workspace data. The model sees one workspace's input at a time.
- No retraining on your data. The model provider does not retain content for training under our contractual zero-retention agreement.
Auditability
Every AI generation is recorded with:
- The feature that triggered it.
- The model version.
- The hashed prompt (so we can verify what was sent without storing the readable prompt long-term).
- The output that was returned.
- The user who triggered the generation.
- The timestamp.
This gives auditable history for every AI-produced piece of content. If a JD or a narrative is ever the subject of a question (a candidate disputes a screening narrative, an audit asks where a number came from), we can trace it.
Where AI is deliberately not used
To make the limits explicit:
- Competency profile updates are deterministic — fidelity-weighted average with recency decay.
- Match scores for career paths and succession are deterministic.
- Learning recommendation ranking is deterministic — gap-weighted catalogue lookup with no LLM in the loop.
- Auto-advance threshold checks are simple comparisons.
- Pipeline stage transitions and emails are templated, not generated.
- Notifications are triggered, not generated.
- Authentication, authorisation, audit logs are conventional code.
The principle: AI handles language tasks where it adds value. The platform's load-bearing logic is conventional code so that it is auditable and reproducible.
Quota
Each workspace has an AI quota. See AI usage in your workspace for the detail. The short version: the default is generous; if you exceed it, AI features degrade gracefully (the platform does not break, it just falls back to manual or deterministic alternatives).
Where to read more
The authoritative AI transparency posture is on the /trust page. It includes the full data-flow diagrams, the contractual posture with the model provider, and the upcoming migration plan to a regional managed model service in Saudi Arabia.