AI governance

AI that moves money has to be accountable.

ADmetric AI is built on a simple promise: every model decision is explainable, evaluated, overridable, and never trained on your data. This is how we live up to it.

Human-in-command, always

Every AI action that moves money is gated by a policy or an explicit approver. Autopilot can recommend; only people and signed policies execute.

Explainability by default

Each AI suggestion ships with the inputs, the model, the version, and the counterfactual — visible in the activity timeline and exportable to your auditor.

No training on your data

Customer financial data is never used to train foundation models. Prompts and completions are stored in your tenant, encrypted, and deleted on schedule.

Least-privilege model access

Models run in regional inference with no internet egress and no access to other tenants. Prompts are scoped to the requesting workspace.

Evaluated, not vibes-checked

Every release ships with a model card: pacing MAPE, anomaly precision/recall, currency drift, hallucination rate on finance Q&A, and bias probes.

Right to a human

Any user can dispute an AI decision in one click and route to a human reviewer with full context. SLA: 4 business hours.

Model cards

The models in production

ModelTypePurposeEvaluation
Pacing Forecast v4.2Probabilistic time-series (gradient-boosted + quantile heads)P10/P50/P90 spend trajectories per budget.MAPE 2.1% at 7d horizon · backtested 18 months.
Anomaly Sentinel v3.1Multi-variate isolation + supervised classifierDetects spend, FX, fee and pacing anomalies.Precision 0.94 · Recall 0.88 · false-positive budget 1.5%.
Copilot Q&A v2.0Retrieval-augmented LLM with finance groundingAnswers questions over your transactions, budgets, contracts.Citations 100% · ungrounded refusal rate 99.6%.
Policy Recommender v1.3Rule-mining + LLM rewriterSuggests guardrails based on your spend patterns.Human acceptance 71% · zero unsafe accepted.

Full model cards (datasets, fairness probes, known limits, deprecation policy) available under NDA on /procurement.

Frameworks

Frameworks we map to

EU AI Act

High-risk classification self-assessment, technical documentation, post-market monitoring.

NIST AI RMF 1.0

Govern, map, measure, manage — annual third-party assessment.

ISO/IEC 42001

AI management system; certification target 2026 H2.

OECD AI Principles

Inclusive growth, human-centred values, transparency, robustness, accountability.

The override switch

One toggle in your workspace freezes every AI-initiated action across every platform within 12 seconds. The state is logged, signed, and replicated to your audit log. You can stay in monitoring mode forever — Active mode is opt-in, per brand, per policy.