Part 05 · Flagship

5 Controlled MarTech Pilots

Every terminal below is an illustrative dry run using synthetic data. It is not captured platform output, a customer result, a performance forecast, or a cost estimate.

synthetic fixtures · read-only first · approval before writes

Baseline for every pilot: one tenant per trust boundary; task-scoped credentials from a secrets manager; limited network hosts; separate web-search/fetch domain rules; disabled or approval-gated write tools; cost, rate, time, turn, and retry caps; redacted event logs; and a tested stop/kill path. Reviewed 31 Aug 2026 against Anthropic's permission and environment guidance.

01
Daily Cross-Platform Ad Report

Start with synthetic exports, then a canary account using read-only scopes. Validate schemas and provenance before generating an alert draft; do not change campaigns or send messages automatically.

SYNTHETIC DRY RUN: Read-only report reviewer
✓ Loaded tenant_demo fixtures; no live credentials
✓ Validated schema, currency, timezone, and source manifest
✓ Reconciled totals against deterministic checks
⚠ Policy threshold triggered in campaign_fixture_07
→ Evidence attached; causal explanation not inferred
✓ Draft report generated with source timestamps
⚠ Slack send blocked pending analyst approval
Pilot records list cost, latency, reviewer corrections, and disposition
Pilot gate: reconciliation accuracy + acceptable alert precision
02
Creative Testing & Analysis

Use licensed or synthetic creatives and a human-authored rubric. The agent can tag observable features and propose test hypotheses; it cannot know which asset will win before a controlled experiment.

SYNTHETIC DRY RUN: Creative evidence assistant
✓ Rights manifest and synthetic asset set accepted
✓ Assets blinded as creative_A through creative_T
creative_A: benefit stated early; mobile text needs review
creative_B: testimonial claim requires substantiation
⚠ No predicted CTR or universal quality score produced
✓ Test hypotheses balanced across audience and placement
⚠ Launch blocked pending brand, legal, and media-owner approval
Pilot compares tags with human labels and later experiment data
Pilot gate: label agreement + zero unsupported performance claims
03
Competitor Intelligence

Research public, permitted sources on an approved domain list. Preserve URLs and timestamps, separate observations from inference, and route strategic recommendations to a human owner.

SYNTHETIC DRY RUN: Public-source monitor
✓ Allowed domains, robots/terms checklist, and rate cap loaded
✓ Retrieved approved fixture snapshots; no login bypass
Observation A: page copy changed [source + timestamp]
Observation B: public ad variant appeared [source + timestamp]
⚠ Business intent and performance cannot be inferred
✓ Brief separates facts, uncertainty, and analyst questions
⚠ No pricing, bidding, or campaign action executed
Pilot gate: citation coverage + lawful-source review
04
Landing Page Optimizer

Audit an owned or explicitly permitted page for observable accessibility, semantics, performance, and content issues. Recommendations are hypotheses; conversion impact requires a measured experiment.

SYNTHETIC DRY RUN: Owned-page audit
✓ Ownership and approved URL scope confirmed
✓ Deterministic accessibility/performance tools completed
⚠ Form label missing in fixture [selector + evidence]
⚠ Image payload exceeds team threshold [measurement]
Hypothesis: clarify CTA copy; validate with an approved test
⚠ No conversion-lift estimate or automatic edit produced
✓ Findings queued for designer and developer review
Pilot gate: reproducible evidence + reviewer acceptance
05
Email Sequence Builder

Draft messages from an approved brief and synthetic recipient profiles. Consent, suppression, localization, claims, accessibility, and send approval remain deterministic or human-controlled.

SYNTHETIC DRY RUN: Email drafting assistant
✓ Approved brief and synthetic personas loaded
✓ Brand, claims, locale, and accessibility rules applied
Draft A: welcome; no invented discount or performance metric
Draft B: product education; claims linked to approved source
✓ Rendering, links, and unsubscribe placeholder checked
⚠ Recipient selection and send tools are disabled
⚠ Legal, brand, and campaign-owner approval required
Pilot measures defect rate and editor effort, not predicted opens
Pilot gate: zero policy violations + acceptable edit burden

These are pilot patterns, not deployments.

A working demo proves only that one path ran once. Production candidacy requires representative, adversarial, prompt-injection, and regression evaluations; privacy and legal review; least privilege; approval for consequential tools; monitoring; incident response; and a reversible rollout.

Next: configure a bounded beta pilot and verify every API detail in current docs