Trust Center

Responsible AI & Governance

AI assists and drafts; humans evaluate and decide. These are the non-negotiable governance principles engineered into GattAI’s code—enforced by architecture, not just aspirational policy.

Enforced Human Gates & Four-Eyes Governance

In human resources, artificial intelligence must never replace human empathy, contextual judgment, or moral accountability. In GattAI, human approval is not an optional workflow recommendation—it is a structural constraint built into our code. Speed has no value if it detaches judgment from decisions affecting careers and livelihoods.

No Autonomous Employment Decisions
The platform is architecturally incapable of executing binding employment actions—such as disqualifying candidates, finalizing appraisals, altering salaries, or initiating terminations—autonomously. Models draft and analyze; designated human leaders decide.
Unskippable Terminal Approval Gate
Every AI generation terminates at a mandatory human approval interface. No administrative setting, system configuration, subscription tier, or API flag can bypass this sign-off milestone.
Self-Approval Banned (Four-Eyes Principle)
The system rejects self-approval outright: a user who initiates an AI workflow (such as commissioning a performance appraisal or compensation review) cannot be the sole approver of that record. Sign-off must come from a secondary authenticated supervisor.
Granular Edit-Then-Approve Workflow
Human approvers retain full authority to accept, edit, annotate, or reject AI drafts. Every iteration—the initial model proposal, intermediate human edits, and the final approved artifact—is archived with complete attribution.

Permanent Artifact Provenance & Lineage

Any artifact created or assisted by an AI agent carries indelible provenance metadata for the entirety of its lifecycle across the system.

Persistent Provenance Labeling
Every AI-generated document—from a job description draft to an interview scorecard—displays the model family, execution timestamp, input data baseline, and model confidence indicators.
Human Edits Do Not Launder AI Origin
Subsequent human editing does not erase the record of AI generation. The platform displays both the original machine-generated draft and the subsequent human modifications side-by-side in audit views, preserving full historical transparency.
Legal & Dispute Traceability
If an employment decision is ever contested or reviewed by labor courts, internal ethics boards, or external regulators, the organization holds an irrefutable paper trail proving exactly what the AI proposed and what the human approved.

Anti-Hallucination & Evidence-Grounded Synthesis

The most dangerous failure mode of generative AI in enterprise HR is a confident, fabricated fact or score. GattAI’s pipelines are engineered to reject hallucination at every step rather than apologize for inaccuracies after the fact.

Missing Data Stays Missing
If an employee’s historical performance records lack quarterly milestone data, or if an applicant’s CV omits specific technical certifications, the system strictly forbids the model from guessing or assuming a baseline. The output explicitly reports: "Data Unavailable for Assessment."
Evidence-Grounded Scoring Rubrics
Every qualitative rating (1–3 competency rating) requires the model to extract and cite verifiable, direct evidence from candidate documents or approved performance cards. Unsupported assertions fail automated rubric validation.
Automated Fact-Checking Layer
Generated outputs pass through deterministic verification checks that compare generated claims against the underlying structured input data. Any hallucinated metrics or dates trigger an automated rejection and regeneration.

When input data is too sparse to support a rigorous assessment, the only honest response is an abbreviated output that clearly communicates the context deficit—and that is exactly what GattAI provides.

Honesty in the Interface & Deterministic Telemetry

We reject dark patterns and simulated animations. The GattAI interface communicates the authentic state of backend systems at actual speed and verifiable cost.

Real Background Tasks, Real Timelines
Deep multi-stage HR synthesis takes computational time. Progress bars reflect real Celery background worker tasks. We publish authentic production benchmarks: Job Description Generation (2–4 minutes), Multi-Source Performance Synthesis (9–15 minutes), and Batch Candidate Screening (3–6 minutes).
Transparent Step-by-Step Cost Accounting
Administrators can inspect the exact token consumption, execution latency, and financial cost incurred by every discrete pipeline step.
Disclosed Model Fallbacks
If an external provider experiences transient API throttling or downtime and the system routes to a verified secondary model, the failover event is logged explicitly in the run record—never substituted covertly.
Real execution benchmarks and transparent cost accounting
job description generation 2–4 min · ~8,500 tokens · $0.024performance review synthesis 9–15 min · ~32,000 tokens · $0.096candidate screening (per 20) 3–6 min · ~18,000 tokens · $0.054progress indicators bound to real Celery tasks · zero simulated delay

Candidate Rights & Algorithmic Contestability

Individuals whose career trajectories are touched by our systems have rights and standing within our platform architecture.

Affirmative AI-Use Disclosure
Job applicants receive explicit, transparent notification prior to participating in any AI-assisted screening or avatar interview process. Notification is presented prominently, not buried in small print.
Right to Request Human Review
Every candidate application includes an integrated mechanism allowing the applicant to request a full secondary review by a human talent specialist if they believe an automated assessment failed to capture their qualifications accurately.
Demographic Blindness by Design
Candidate screening models are architecturally insulated from protected demographic identifiers (gender, age, marital status, nationality, and photos), eliminating common vectors for unconscious human or machine bias.