Anomaly Watcher: An AI Agent That Explains Alerts and Pauses Sends

Its rail is part of the vendor-neutral AI agent rails method.

Rail at a glance

Two separate paths share the Anomaly Watcher's rail, and the table covers both. A rule service owns the pause and takes numbers alone, while the model session holds read tools and has no route to the pause call.

What it does and who uses it

The Anomaly Watcher answers two questions during a send: is this number bad enough to stop, and why might it have moved? Specialists read the explanations, and the head of marketing operations owns the rule.

Take an illustrative morning. A product-launch email starts going to 400,000 people at 09:00. By 09:20 the hard-bounce rate on the first 30,000 messages is 6%, against a written limit of 4%. The send clears the minimum volume and is not exempt, so the rule service pauses it and writes this record (illustrative values, not a product API):

{
  "event": "send.paused",
  "metric": "hard_bounce_rate",
  "value": 0.06,
  "threshold": 0.04,
  "threshold_version": "pause-rule-v7",
  "data_age_seconds": 120,
  "resumed_by": null
}

The model then reads the change log and offers two hypotheses. A list import at 08:40 added an old file, which fits hard bounces. A provider block would show different bounce codes, so it ranks lower. The specialist removes the import, and a named person resumes the send with a reason.

Detection tools and pause rules work at different speeds. Adobe Analytics anomaly detection uses a 95% prediction interval at day granularity 3, and Amplitude's Prophet-based Anomaly + Forecast does not support real-time charts 4 (both checked 1 November 2026). Detectors like these suit explaining a move better than stopping a send.

Where Amazon Simple Email Service (SES) is the sending layer, it reviews an account at a 5% bounce rate or a 0.1% complaint rate, and may pause sending at 10% or 0.5% 1. Those are account rates over a representative volume, so a per-send limit should sit well below them. Google's spam-rate lines, below 0.10% and never reaching 0.30%, apply over days 2, so they belong to the Deliverability Watcher.

Why this level

The Anomaly Watcher carries two levels because explaining and pausing carry opposite risks. A bad explanation misleads a reader who can check it. A pause delays a campaign, while a bad send cannot be recalled, so stopping early is the cheaper mistake.

Meta's Agents Rule of Two applies per path 5. The explanation session reads change notes and calendar entries, so it holds untrusted input, but no profiles and no state-changing tool. The pause path changes state but takes only numbers, which leaves no text for an instruction to hide in.

L4 normally expects auto-revert, and a pause deliberately has none, because a timer would restart a send while the cause is unknown. An immediate alert to the owner and the daily cap stand in for that safeguard. A breached false-pause budget or an open incident returns the rule to alert-only.

Three ways it fails, and the control for each

The Anomaly Watcher fails differently on each path, so the table says where each control sits.

How to evaluate it

Evaluate the Anomaly Watcher in two parts, because the rule path can be proved while the explanation path can only be graded.

For the rule path, write tests with a known answer and require every one to pass:

1. A value just below the threshold, and one exactly on it.
2. A breach on a send under the minimum volume, on an exempt send, and on a transactional send.
3. A breach with stale data behind it, and one after the daily cap is used up.

Then replay history. Past incidents measure time to contain, from the first bad metric to the stopped send. Quiet periods and seasonal peaks measure false pauses.

For explanations, a reviewer checks whether the true cause appeared among the hypotheses. Finish with a drill: disable the pause rule, and confirm that alerts still reach a person.

Where it runs

The Anomaly Watcher's action path uses rules only, so no model sits between a metric and a pause. A model writes the explanation, and the on-premise LLM guide covers the private option, preferred because change notes are commercially sensitive. PrivyCDP's AI-assisted review extends to unusual events, while the rail itself stays yours to define.

Related

• Marketing anomaly detection on rails: the use case, step by step.
• AI agents for marketing operations: the playbook this watcher belongs to.
• Rail card template: a place for your threshold and exempt list.

Other watchers and reviewers are listed under AI agent rails. To record your own pause rule, download the rail card template.