L1 domains and L2 tiles describe the bank’s talent-acquisition operating model. The Harvey ball — not the tile’s presence — shows this suite’s coverage.
Implemented in this solutionImplemented, with gapsAvailable by other Tenyards SolutionsDecision support onlyEnabling foundation onlyIn the roadmap
Select any L1 domain to open its detailed coverage.
Who augments the TA team
Agent workforce
Five agents and four AI services. Agents read before they speak and every consequential action — a stage decision, a bulk rejection, completing a hire, closing a requisition — pauses for a person. AI services grade, score and interpret; people decide.
The single accountable controller for a requisition’s workflow: reads the pipeline, persists explainable fit scores, asks the Authenticity & Fraud analyst for a verdict before anything advances, and queues consequential decisions for a person.
Owns
Workflow sequencingFraud gate before matchingFit-score persistence
Contributes to
Requisition lifecycle & capacity control · AI talent discovery · Talent rediscovery · Integrity & fraud detection · Explainable fit scoring · Shortlisting & stage decisions · Hire & requisition completion · Recruiter copilot · Agent action approvals · Human oversight of AI decisions · Agent runtime & evaluation
How it works
Uses list_requisitions, get_requisition, list_applicants, list_potential_candidates, assess_authenticity, persist_fit_score, invite_screening, organize_interviews_for_job · pauses for approval on record_decision, reject_candidates_for_job, complete_hiring_for_job, close_requisition · code in app/agents/supervisor.ts
An isolated integrity specialist: weighs screening evidence and candidate history, and records evidence-backed fraud signals. It flags for human review — it never rejects and never scores fit.
Owns
Integrity verdictsEvidence-backed fraud signals
Contributes to
Integrity & fraud detection · Human oversight of AI decisions · Agent runtime & evaluation
The candidate-facing concierge: answers questions, matches skills and location to open roles, captures talent-community leads, and — for a signed-in candidate only — reads their own applications and books or moves their own interviews.
Owns
Candidate Q&ARole matching for candidatesSelf-service interview booking
The recruiter copilot: answers pipeline questions from live data, invites candidates to screening, organises interviews and produces a branded PDF status report. Bulk rejections, hires and closures pause for approval.
Interview logistics: reads the recruiter’s calendar and the requisition’s interviews, then schedules or reschedules with no double-booking and an optimistic version lock.
Turns a CV PDF into structured skills, experience and education (PDF to text, then the LLM model), maps the skills onto the catalogue so the CV drives the fit score, with a heuristic fallback parser.
Selection · Started per candidate or for a whole requisition
Grades a candidate A–D against one requisition with a score, rationale, strengths, gaps and per-skill evidence — factoring in screening results and told to ignore protected attributes. Runs in chunks of six for mass evaluation.
Owns
AI grade A–D
Contributes to
AI requisition-match grading · Mass evaluation · Bias controls
Screening · Runs when a candidate submits a screening
Transcribes video and voice answers, scores each answer on relevance, depth, communication and skill evidence, writes an overall verdict and flags synthetic or AI-generated answers — never facial or emotion analysis.
Turns a plain-words query into skills, location and years of experience (LLM Model), then ranks candidates by the deterministic fit model.
Owns
Query understanding
Contributes to
Natural-language talent search
How it works
Uses LLM Model, fit model · code in app/src/index.tsx
No agent matches this view
Clear the search or choose another organisation function.
How value moves
Talent value chains
Each chain crosses functional ownership. The capability map below supplies the reusable abilities behind these stages.
●Core talent●Enterprise control
Enterprise control
Plan to post
Turn an approved business need into a live, budgeted requisition.
01Identify need›
02Approve headcount›
03Define role & band›
04Assign owner›
05Publish
Core talent
Attract to apply
Bring the right people to the bank and into the funnel.
01Build brand›
02Discover talent›
03Engage›
04Apply›
05Capture profile
Core talent
Apply to shortlist
Turn applicants into a defensible, integrity-checked shortlist.
01Interpret CV›
02Screen›
03Check integrity›
04Grade & match›
05Shortlist
Core talent
Interview to offer
Assess the shortlist fairly and make an offer within capacity and budget.
01Schedule›
02Interview›
03Score›
04Decide›
05Offer
Core talent
Offer to productive hire
Convert an accepted offer into a vetted employee ready on day one.
01Accept›
02Pre-employment checks›
03HRIS handoff›
04Onboard›
05Day one
Enterprise control
Insight to action
Turn a pipeline signal into a traceable, authorised action.
01Sense pipeline›
02Detect risk›
03Analyse with agents›
04Approve›
05Execute & audit
What the TA function can do
Business capability map
The L0 enterprise is decomposed into L1 talent domains and L2 capabilities. Status describes this suite’s coverage today; the agent strip is a separate assistance overlay. Press any capability for its mechanics, the rules it enforces and what it deliberately does not do.
Showing all 60 capabilities
Implemented in this solutionImplemented, with gapsAvailable by other Tenyards SolutionsDecision support onlyEnabling foundation onlyIn the roadmap
L1.01
Workforce planning & requisitions
Decide what the bank hires, where, and within which budget.
Clear the search or choose a different coverage layer.
In the roadmap
Requisitions arrive as records; nothing in this App models future demand, attrition or a headcount budget by business line.
Key features
Requisition openings and budget
Each requisition carries openings and a budget, the only planning numbers the App holds.
What it does not do
No scenario planning, attrition forecasting or org-design modelling.
Enabling foundation only
Requisitions carry a draft → open → on-hold → closed status and a status note, but every requisition is loaded as data: no screen creates one, and nothing approves a draft.
Key features
Draft status with a reason
A requisition can sit in draft with a status note (e.g. pending compensation sign-off).
Department, recruiter and hiring manager
Every requisition is owned by a recruiter and a hiring manager and sits in a department and cost centre.
What it does not do
No requisition create or edit screen.
No multi-level approval (finance, HRBP, business) of a draft.
Recruiters filter and prioritise requisitions, move them through their status lifecycle and work each in an eight-tab workbench. Closing confirms first and dispositions every active candidate; offers and hires are capped by openings and budget.
Each requisition has a description and a weighted skills rubric (required skills and 1–5 weights) that the fit model, screening generator and AI grader all read. Both are loaded as data.
Key features
Weighted skills rubric
Required and optional skills with weights 1–5 drive the explainable fit score.
A public careers site with keyword search, filters for department, location, work mode, level and contract type, sorting, pagination and a job page with similar roles.
Key features
Faceted job search
Keyword, department, location, work mode, level, type and sort.
Aria answers questions, matches a candidate’s skills and location to open roles with a deterministic weighted overlap, captures talent-community leads and, for a signed-in candidate, reads their applications and books or moves their own interviews.
Key features
Role matching
Deterministic weighted skill overlap — the same model the recruiter sees.
Own-interview booking
Book or move an interview on the candidate’s own application only.
Rules it enforces
Candidate identity comes from the signed-in session, never from the model (D-005).
Ownership of every application and interview is checked in code.
Anonymous visitors get roles, FAQs and lead capture only.
What it does not do
Does not invite to screening or make any selection decision.
No job-board syndication, programmatic advertising or spend tracking.
Implemented, with gaps
Discovery proposes four to seven profiles per requisition from the enabled sourcing channels, scores them with the fit model and stages them as proposals; an accepted proposal becomes a candidate and application at the Leads stage.
Sourced leads wait in a review queue. Accepting creates a recruiter task, a cold email or an email to the representing agency; removing drops them silently. Until approved, a sourced lead cannot be screened or advanced.
Key features
Three engagement routes
Recruiter task, direct cold email, or via the headhunter.
Sourced gate
Unapproved leads can only be rejected and are kept out of screening.
Rules it enforces
Needs the sourcing-approve permission.
The shared decision function refuses to advance an unapproved sourced lead.
A free-text query is turned into skills, location and minimum years by the LLM model; results are ranked by the deterministic fit model, so the ranking stays explainable.
A signed LinkedIn-apply webhook creates a candidate and application at Leads, fit-scores it, runs an authenticity heuristic and is duplicate-safe. A simulator in configuration posts realistic payloads.
Key features
Signed webhook
Fails closed when its shared secret is not set.
What it does not do
No live LinkedIn, job-board or agency-portal connection.
Sign-up, sign-in, password reset and sign-out with salted PBKDF2 hashing; a profile with phones, addresses, right to work, SMS and WhatsApp consent, LinkedIn and Drive links; CV and cover-letter upload to private storage.
Key features
Secure credentials
PBKDF2 with 100k iterations; one-hour reset tokens.
Consent capture
SMS consent and WhatsApp opt-in recorded on the profile.
The Apply button and Aria both capture the candidate into the talent community; applications on a requisition come from sourcing, inbound channels or seed data.
A recruiter triggers parsing on a candidate: the PDF is converted to text and the LLM model returns structured JSON, with a heuristic fallback. Parsed skills are mapped onto the catalogue so the CV drives the fit score.
Every decision sends a templated notice by email when the email integration is on; messages are kept in an outbox. WhatsApp sends are logged without a gateway.
Key features
Templated notices
Templates per decision, editable in configuration.
Each requisition has a master question set; a draft is generated from the role’s top skills and the organisation-wide default questions and is then edited by the recruiter.
Key features
Draft from the rubric
Template questions built from the top weighted skills.
What it does not do
The draft is template-based, not written by a language model.
Recruiters attach a screening at the Applied stage, singly or as a batch first cut. Candidates answer on a private link; answers are transcribed and scored on four competencies with an overall verdict.
A flagged screening raises a high-severity signal automatically; the Authenticity & Fraud analyst records evidence-backed signals. Analysts review, dismiss, confirm or reopen; confirming rejects the candidate’s live applications.
100 × Σ(weight × proficiency/5) ÷ Σweight, minus 12 per missing required skill. The derivation is shown per candidate; a requisition can be re-scored in one action.
Key features
Per-line explanation
Each skill’s contribution on hover.
Re-score
Recompute a whole requisition after a rubric or CV change.
The LLM model grades the parsed CV against the requisition and its screening results, returning a grade, score, recommendation and written assessment shown beside the deterministic fit score.
Advance, reject, hold and hire all go through one decision function that checks the sourced gate and the fraud gate, writes the decision log and sends the candidate notice.
Every decision writes a row with action, from and to stage, decider, rationale and fraud-gate state; agent events are kept alongside. Each requisition shows its activity tab.
Selected candidates are organised into weekday slots with a video link; the candidate confirms, asks for three alternatives or declines on a tokenised page. The Interview Organizer agent schedules with no double-booking and an optimistic version lock.
Selecting for offer checks openings, budget and the fraud gate and proposes the band midpoint; the offer is then rolled out by email, generated letter or externally.
A decline is recorded by the recruiter; there is no candidate accept or e-signature.
What it does not do
No counter-offer tracking or e-signature.
Implemented in this solution
Hiring checks capacity and the fraud gate and notifies the hiring manager. Completing the requisition rejects the remaining pipeline with an apology and closes it. Withdraw, decline, revert and cancel-hire are supported.
Key features
Capacity-capped hire
Never more hires than openings.
Complete requisition
Close and notify the remaining pipeline.
Rules it enforces
An agent may only complete hiring through an approval-gated action.
The cockpit shows KPIs, the stage funnel and coverage by recruiter, hiring manager and country; the pipeline is a worklist with warn and stalled ageing from configuration and a next action per row.
Key features
SLA ageing
Warn and stalled thresholds set in pipeline settings.
Next action per row
Links straight into the step to take.
Rules it enforces
Recruiters and hiring managers see only their own requisitions.
Otto answers pipeline questions from live data, invites to screening, organises interviews and produces reports; the Supervisor and Interview Organizer are available in the in-app assistant.
Key features
Reads before it speaks
Every answer about state comes from a read tool.
Refuses what it has no tool for
e.g. offer roll-out stays with a person.
Rules it enforces
Reads go through named domain functions, never generic SQL (D-003).
Decisions, bulk rejections, hiring completion and requisition closure requested by an agent become pending approvals; resolving one resumes the agent with the outcome.
Tasks raised by sourcing and agents appear on the recruiter’s page and are completed with a note; every message is kept in the outbox with its channel and status.
The executive view has four lenses — capacity and velocity, capability and quality, financial and channel, pipeline health and risk — computed from live requisitions and applications.
Key features
Four lenses
One page for the CHRO and Head of TA.
Rules it enforces
Visible to TA admin, TA director, HR IT and CHRO by default.
A four-fifths adverse-impact read-out by source and country; protected characteristics are not collected, so true diversity monitoring needs consented data.
Eight personas, fourteen view and thirteen action capabilities in an editable matrix with protected cells; recruiters and hiring managers are pinned to their own requisitions.
Key features
Editable permission matrix
TA admin always keeps full access; HR IT cannot be locked out.
Advance, reject and hire are recorded human decisions; the fraud gate blocks advancing a flagged candidate; agent decisions pause for approval. The governance page reports human-oversight percentage, rationale coverage and gate blocks.
Key features
Oversight metrics
Human-decided share, rationale coverage, gate blocks and overrides.
The grading and screening prompts are told to ignore protected attributes; the deterministic fit model reads skills only; adverse impact is monitored on proxies.
Email is sent through the platform; LinkedIn apply arrives on a signed webhook. Slack, calendar, HRIS, finance and an outbound ATS webhook appear as switches only.
A catalogue of signed operations — status reports, screening, scheduling and more — callable by the platform and other agents.
Key features
Operation catalogue
Discoverable list plus nine operations.
Implemented in this solution
Five agents run as application code (D-002) on one model, emit typed events to an event store and carry evaluation cases for their primary flows, approval pauses and guardrails.
Key features
Evaluation cases
Primary flow, guardrail refusal and approval pause per agent.
Capabilities describe what a bank’s talent-acquisition function may need to do; their presence is not a statement that this suite supplies them. Value chains show how those abilities combine to produce an outcome. L0 is the enterprise, L1 the talent domains and L2 the capabilities within them. Where a capability has been taken to process level, L3 names its parts and L4 the processes that deliver them, each drawn as steps in the people, agent and system lanes; its Harvey ball is then worked out from those steps and never shows more than they support. Every claim is written from this suite’s own implementation — its screens, its agents and its data — with what it does not do stated beside what it does. The hierarchy is an adaptation for this product, not a verbatim reproduction of any external framework.