Detection finds bad actions after they execute. Pre-action authorization stops them before they execute. For irreversible actions — wire transfers, benefit redirects, AI-voice-cloned approvals — detection alone is the wrong primitive.
Modern fraud detection — behavioral analytics, statistical anomaly models, BEC scoring, transaction monitoring — runs after the action submits. The signals are real: unusual destination, unusual time, unusual amount, atypical user agent. The downside is structural: by the time the alert fires, the wire has cleared.
That tradeoff worked when most fraud cleared slowly and recovery was possible. It does not work when the action is an instant ACH or a same-day wire to a beneficiary that goes silent within minutes.
Behavioral models assume the legitimate user is a stable signal — same IP ranges, same device, same approval cadence. AI-voice-cloned phone calls reproduce the legitimate user's signal exactly. Prompt-injected agent runtimes operate from the same authenticated session, the same scope, the same device. The "anomaly" the detection model is looking for is no longer there.
EP changes the question. The system doesn't ask "does this transaction look anomalous?" — it asks "did a named human authorize this exact destination, this exact amount, with a valid handshake?" The answer is binary, cryptographic, and resistant to the channel the attack arrived on.
| Dimension | Post-action fraud detection | EP pre-action authorization |
|---|---|---|
| Where the check runs | After the action executes | Before the action executes — gates execution |
| Signal source | Behavioral patterns, statistical models | Cryptographic handshake + named human signoff |
| False-positive cost | Legitimate transactions blocked or delayed | Adds a signoff step on Tier-2 actions only |
| False-negative cost | Funds gone; recovery rare | Action does not execute without valid handshake |
| Effectiveness on AI-voice / deepfake | Degrades — model-driven attacks evade behavior baselines | Independent of attack channel — binds the action, not the actor channel |
| Effectiveness on insider misuse | Limited — insider patterns look normal | Handshake binds authority chain at request time |
| Audit evidence | Alert + post-hoc investigation | Self-verifying trust receipt issued at the gate |
| Composes with | EP, MFA, audit logs | Detection (defense in depth) |
EP and detection compose. Detection is still the right control for Tier-0 reads, login risk scoring, fraud pattern discovery across the long tail, and downstream forensics. EP is the right control for the irreversible Tier-2 actions where post-hoc detection doesn't return your money.
A community bank running EP on wire releases keeps its existing transaction-monitoring stack. Most transactions never see EP — they're below the action-binding threshold. The wire-out-to-new-beneficiary action does. The handshake refuses to clear until a named officer signs off on the exact destination and amount.