Paid Media Optimization Agent

What this agent does

Watches the live paid structure every day — search, LinkedIn, and the rest of the rebuilt channel mix — against pacing, efficiency, and the plan, then hands Alexis a ranked list of moves with the evidence attached: what to shift, why, expected impact. It executes only what's approved. The daily grind of pulling performance, spotting drift, and drafting changes moves to the agent; the judgment and the trigger stay human.

Why this agent

• The rebuild deserves daily eyes — the transition lands Sep 1 (Outshine out Aug 31, Verto live on search + LinkedIn) and the whole paid thesis is discipline: right personas, no waste, value-fed bidding. Discipline is a daily practice, and daily is what agents are for.
• Most of the budget flows through paid — small efficiency moves compound faster here than anywhere else; the H2 memo's argument for fixing paid first is the same argument for watching it continuously.
• The conversion loop makes optimization real — value-attached conversions (CAPI + Marketo OCT) start feeding the platforms in the same window; someone has to act on what the loop reveals, weekly, without fail.
• Volume beats a human calendar — every campaign, every geo, every day is a coverage promise a person can't keep and an agent can't break.

Trigger & inputs

• Trigger — daily scheduled pull; threshold alerts in between (pacing drift, spend spikes, conversion drops, disapprovals).
• Reads — platform APIs (read scope), the warehouse over the MCP for downstream truth (MQL→SAO by campaign, not just platform conversions), the media plan targets, and the tracking standards so campaign naming maps to the taxonomy.
• Context — the geo-holdout brand test design and the awareness/performance split — so it never "optimizes" a test into invalidity.

What it produces

• A daily state-of-paid — pacing vs plan by channel and campaign type, exceptions first.
• Ranked recommendations — each with evidence, expected impact, and the exact change spelled out, queued for approval.
• Approved-change execution + log — what changed, when, on whose approval, with before/after tracked.

How it works

• Pull performance across platforms + the warehouse's downstream funnel data.
• Compare against plan, baselines, and thresholds; separate signal from noise.
• Draft the moves: budget shifts, bid changes, negative keywords, audience exclusions — ranked by expected impact.
• Route to Alexis for approval; execute approved items; log everything with before/after.

Guardrails & human-in-the-loop

• Hard gate on spend — no budget, bid, or structure change without explicit approval; this gate stays through at least V2.
• Never-do list — never launches or pauses campaigns on its own; never touches the geo-holdout or any running test's variables; never changes tracking/UTMs; never moves spend between awareness and performance buckets (that split is a strategy decision).
• Thresholds — anomaly beyond bounds → alert, don't act; platform disapprovals → flag to Alexis same day.
• Escalation — data disagreement between platform and warehouse → flag to Data & Tracking, act on neither until resolved.

Success metrics

• North star — efficiency on treated campaigns vs locked baseline: cost per SAO, pipeline per dollar. Baselines lock at the September rebuild (TBD until then).
• Quality bar — recommendation acceptance rate on Alexis's weekly review; rejected recs reviewed for why.
• Guardrail metric — zero unapproved changes; zero test contamination incidents.
• Hours returned — daily reporting + change-drafting hours off Alexis's and Verto's plate, on the program ledger.

Build plan

• V0 · Shadow mode (early Oct) — recommends against live campaigns; Verto's specialists make their own calls; recs compared to what humans did. Exit: recommendation quality judged useful on a few weekly cycles.
• V1 · Assisted (Oct–Nov) — Alexis approves from the queue; agent executes approved items with full logging. Exit: acceptance rate + zero-incident record.
• V2 · Bounded autonomy (V3 window, if ever) — pre-agreed micro-moves (e.g., negative keywords) inside caps, only after cost controls + governance land in Month 3. Spend gates likely permanent.

Dependencies & risks

• Conversion loop live — value-fed optimization needs the CAPI + Marketo OCT loop from Data & Tracking; without it the agent optimizes platform-reported proxies and says so.
• Rebuild stability first — optimizing during the September transition chaos would tune noise; V0 starts once the new structure has baseline weeks.
• Test protection is sacred — the geo-holdout answers the brand-bidding question; one careless "optimization" invalidates a quarter's evidence.
• Platform API access — read + change scopes via individual accounts per the security protocol. [owner to complete]