Several features in Jadiroon use AI. This page lists which ones, what the AI sees, what model is used, and how the workspace's quota is tracked.
Which features use AI
Six places in v0.5:
- Job description generation — when you generate a JD on the role wizard. The model receives the title, function, your seed description, and the required competencies; it produces the JD body.
- Interview rubric suggestion — when you click Suggest questions on the rubric editor. The model receives the role's required competencies and their taxonomy definitions; it proposes questions tied to each competency.
- CV screening narrative — when a candidate applies. The model receives the parsed CV text and the role's required competencies; it produces a short narrative summarising fit. The deterministic match score is computed separately, not by AI.
- Assessment narrative reports — after an assessment is completed. The model receives the deterministic scores and the assessment metadata; it produces a 3–5 paragraph narrative interpretation.
- Career path narratives — on the career-paths surface. The model produces the why this is a fit narrative for each top role match.
- Learning recommendation ranking — when ranking the catalogue against an employee's gap profile. The model receives the gap profile and the catalogue's metadata; it ranks the candidates.
Which features do not use AI
To pre-empt assumptions:
- Competency profile updates. The blend is deterministic — fidelity-weighted average with recency decay. No AI inference.
- Match scores for career paths and succession. Deterministic formula.
- Auto-advance threshold check. Simple comparison.
- Pipeline stage transitions and emails. Templated, not generated.
- Notifications. Triggered, not generated.
- Authentication, authorisation, audit logs. All conventional code.
The principle: AI handles language tasks where it adds value. The platform's load-bearing logic — what counts as a measurement, how competencies aggregate, who sees what — is conventional code.
What model
The current production 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 full transparency posture, including data residency for AI calls, is on the /trust page. A summary appears in Section IX — AI transparency.
Quota
Each workspace has an AI quota — the total token-equivalent budget for all AI calls per billing period. The default in v0.5 is generous; most workspaces do not approach the cap.
The quota is visible at /{slug}/app/settings/ai-usage:
- Total budget for the period.
- Consumption to date with the per-feature breakdown.
- Projected end-of-period consumption based on current rate.
- Days remaining in the period.
If you are approaching the cap, the platform sends an AI_QUOTA_WARNING notification at 80% and 95% consumption.
What happens at the cap
If the quota is exhausted:
- JD generation returns a clear AI quota exceeded message; you can save the role without the AI-generated body and write the JD yourself.
- Rubric suggestion is disabled until the quota resets; you can author rubric questions manually.
- CV screening narratives are skipped — applications still succeed, the deterministic match score still computes, the narrative is just not produced.
- Assessment narratives are queued and produced when the quota resets. The score is still available; the narrative is delayed.
- Career path narratives are skipped on the surface; the match scores still display.
- Learning recommendation ranking falls back to a deterministic match (less smart than the LLM ranking but functional).
In every case, the platform's core function continues. AI is additive; quota exhaustion does not break the platform.
Caching
Generations are cached against their inputs. Regenerating a JD with the same title, description, and competencies returns the cached output (and does not consume quota). Changing one input produces a new generation.
This means heavy use of regeneration during prototyping (you tweak the wording on a draft role multiple times) consumes proportionally to your input changes, not to your generate clicks.
Next
Workspace settings — the catch-all for per-workspace configuration — is the final page in this section.