AI Copilot & Models
Probabilistic forecasts. Hard guardrails. Explainable answers.
ADmetric runs four production models — Pacer, Sentinel, Allocator and Copilot — that together forecast every euro you'll spend, catch every anomaly before it bites, and recommend or execute the right reallocation within your guardrails.
4
Production models
15 min
Forecast refresh cadence
Zero
Customer data used to train shared models
Pacer
Time-series · Bayesian state-space
Forecasts daily and intraday delivery per ad account against your plan, with P10 / P50 / P90 confidence bands updated every 15 minutes.
Bayesian structural time-series with platform-specific priors. Re-trained nightly on rolling 18 months.
Sentinel
Anomaly detection · Robust z-score
Catches CPA drift, CPM spikes, frequency runaway and creative fatigue before they consume a week of budget.
Robust z-score with EWMA baseline. Operates per campaign, with auto-tuned thresholds per platform and vertical.
Allocator
Optimisation · Constrained MILP
Recommends — or executes — daily budget redistribution across platforms and campaigns under hard guardrails.
Mixed-integer linear program with per-platform, per-CPA and per-brand constraints. Solves in under 400 ms.
Copilot
LLM · Finance-tuned
Answers questions in plain English: 'why is Meta pacing fast in Germany?' Drafts approvals, board commentary and month-end notes.
Retrieval over your ledger, audit log, anomalies and forecasts. Grounded answers with citations to underlying rows. Never trained on customer data.
Benchmarks
Measured across 412 advertisers and €1.8B of managed spend in the 12 months ending Q2 2026.
| Metric | ADmetric | Baseline |
|---|---|---|
| Pacing forecast MAPE (7-day) | 2.4% | Industry rule-of-thumb 8–12% |
| Anomaly detection precision | 94% | Hand-tuned alert systems ~60% |
| Anomaly detection recall | 91% | |
| Median time to flag CPA drift | 9.4h | Faster than weekly review cadence |
| Reallocation latency | <2 min | From signal to platform write |
How the Copilot stays honest
No training on your data
Your ledger, audit log and platform data are never used to train shared models. Models are tuned on synthetic and licensed data.
Grounded answers
Every Copilot answer cites the underlying rows: campaigns, invoices, audit events. If we cannot ground it, we will not say it.
Deterministic actions
Recommendations and writes always go through the Allocator and your approval rules. The LLM never writes to a platform directly.
See it on your data