Marketing Mix Modeling

The framework — strategy first

Marketing Mix Modeling — the strategic foundation.

Marketing Mix Modeling is Back. Multi-Touch Attribution Didn't Survive Privacy.

Marketing Mix Modeling (MMM) is the econometric methodology that quantifies how much of your sales each marketing channel produced — including brand, including the channels you can't track with cookies (TV, podcast, Out-of-Home advertising or OOH, sponsored events), including the diminishing returns at each spend level. It was the standard before deterministic attribution killed it in 2010. It's back as the standard again — for four reasons that all hit in the same 24-month window.

Why MMM is back

Multi-Touch Attribution (MTA) broke. iOS 14.5 and Apple's App Tracking Transparency (2021) killed mobile deterministic attribution. Chrome cookie deprecation finished the web side. Most B2B SaaS marketers no longer trust their attribution model below the campaign level — they just don't say so out loud.
Open-source tools removed the price tag. Meta released Robyn (2020) and Google released Meridian (2024) as open-source MMM engines. A vendor MMM that ran $150K–$300K/year now runs on a laptop with a Python-literate analyst.
Brand spend got a defender. MMM is the only methodology that fairly attributes brand and performance in the same model. Chief Financial Officers (CFOs) who cut brand budgets because "we can't measure it" now have an answer they'll accept.
AI made it continuous. Where MMM used to be a once-a-year consultant project, agents can now refresh the model weekly, flag saturation in real time, and surface the budget-reallocation recommendation before the quarter ends. The discipline went from annual artifact to live operating dashboard.

If you cannot answer the question "what did our last $1M of spend actually return?" in the language Finance speaks, you don't have measurement. You have dashboards.

Three patterns to build it — pick the one that fits your team

MMM is not a single product. It's a methodology that can be delivered three ways, and the right choice depends on your analyst depth, your data hygiene, and your CFO's comfort with internally-built math.

PATTERN 1 — AGENT-BUILT (OPEN-SOURCE ENGINE)

Engine: Meta Robyn or Google Meridian (open source, free).

Layer: MMM Agent (below) — orchestrates the weekly run, scenario sims, alerts, board snapshot.

Cost: ~$30–80K/year, almost all of which is analyst time + compute.

Time to first defensible model: 4–8 weeks.

Best for: teams with one Python/R-literate analyst (or a CMO willing to use Claude to write the model code), clean weekly data going back 2+ years, and a CFO who trusts your math more than a vendor's.

Trade-off: you own the math, including the bugs. Initial build is real work. Once running, marginal cost approaches zero.

PATTERN 2 — VENDOR ENGINE + AGENT ORCHESTRATION (THE HYBRID)

Engine: Paramark, Revsure, Lifesight, Recast, or Mass Analytics. (Paramark and Revsure are the AI-native names B2B CMOs are actively evaluating in 2026; Lifesight, Recast, and Mass Analytics are the established players.)

Layer: agents on top of the vendor — scenario planning, downstream alerting, monthly board-format report generation, sales/marketing handoff briefings.

Cost: ~$60–150K/year (vendor subscription + agent infrastructure).

Time to first defensible model: 2–4 weeks (vendor brings the connectors).

Best for: enterprise teams that want vendor accountability for the model + agent leverage for everything downstream. You get pre-built data connectors and a customer success team; you get the speed-and-cadence advantage of agents on the output side.

Trade-off: you depend on the vendor's model assumptions. You don't own the math, but you do own how it's used.

PATTERN 3 — PURE VENDOR

Engine + layer: Paramark, Revsure, Lifesight, or Recast handles everything.

Cost: ~$80–200K+/year.

Time to first defensible model: 1–3 weeks.

Best for: teams without a quantitative analyst, messy data that needs vendor connectors to clean, or a CFO who explicitly wants third-party validation of the model.

Trade-off: highest cost. You're renting the methodology entirely. Scenario planning and reporting are vendor-UI-locked.

The data spine MMM requires (and why this is the bottleneck for most teams)

Whichever pattern you pick, the data requirement is the same. If you can't produce this dataset, no MMM tool — agent-built or vendor-built — will work. The first deliverable here is the audit that tells you what's missing.

Time series — weekly granularity, 104+ weeks of history (2 years minimum; 156 weeks / 3 years is materially better).
Channel spend — per channel per week (paid search, paid social, display, TV, podcast, OOH, events, content, PR, etc.). Direct media cost, not bundled "marketing" totals.
Channel impressions — exposure-level data per channel per week, where available (GRPs for TV, reach for digital).
Sales or pipeline — the dependent variable. Weekly revenue or weekly pipeline created. Must align in time with the spend data.
Control variables — seasonality, promotional periods, competitor major moves, product launches, macro events (anything that affects sales but isn't marketing). Without these, the model attributes everything to whatever marketing was running at the time.

The CFO conversation that earns you 18 months of budget runway.

The single most undervalued artifact a Chief Marketing Officer (CMO) can produce is the quarterly MMM-backed budget defense. Not the dashboard. Not the quarterly review. The 2-page PDF that opens with "$Y of $X spent generated $Z of incremental revenue, here's how I know, here's the model methodology, here's where I'd invest next." Done well, this artifact earns 18 months of budget runway with the CFO — meaning the budget conversation is settled before the next budget cycle even starts.

The 3-question MMM test that decides build-vs-buy

Do you have a Python-literate analyst on the team or RevOps? Yes → consider Robyn or Meridian (open-source, $0 software). No → vendor.
Are you willing to wait 6–8 weeks for a first-pass model? Yes → open-source viable. No → vendor or hybrid.
Is your data spine in place? (90 days of clean spend + outcome data per channel) Yes → ready to model. No → fix the spine first, the model won't save you.

The MMM refresh cadence — monthly, not annually

The legacy MMM model was an annual artifact. The modern MMM is a monthly one. The reason: channels saturate faster than they used to (LinkedIn CPL inflation, Google bid creep, content content-fatigue), and the cost of being wrong about saturation for 6 months is now in the hundreds of thousands of dollars. The monthly refresh is non-negotiable.

CADENCE	ACTIVITY	OWNER	OUTPUT
Weekly	Ingest spend + outcome data, run anomaly checks	RevOps + MMM Agent	Data quality report
Monthly	Refresh model, surface saturation shifts, recommend reallocation	MMM Agent	1-page "this month's recommendation"
Quarterly	CFO-facing ROI defense doc, brand-vs-performance attribution	CMO + CFO joint review	2-page budget defense
Annual	Full model rebuild, methodology re-validation, new-channel inclusion	RevOps lead + outside vendor (if applicable)	Validated model + methodology doc

Your MMM Approach

The pattern you've picked, the model status, the refresh cadence. Every measurement, attribution, and CFO-conversation prompt on the site uses these.

MMM pattern (open-source / hybrid / pure vendor)

Engine / tool

Channels modeled (comma-separated)

Refresh cadence + owner

Saved as [MMM PATTERN], [MMM ENGINE], [MMM CHANNELS], [MMM REFRESH CADENCE].

Three audiences, applied to MMM

MMM OUTPUT, MAPPED TO THE THREE AUDIENCES

Finance — Quarterly board snapshot: $X spend → $Y incremental revenue, brand contribution Z%, performance contribution W%. The artifact that defends next year's budget.

Marketing leadership — Monthly review: which channel is past saturation, which is under-invested, where to move next month's $50K. The artifact that runs the budget meeting.

Agents — /mmm-context.md exported by the MMM Agent: per-channel coefficients, saturation curves, current incremental Return on Investment (ROI). Paid Media reads from this. Account-Based Marketing reads from this. Budget & Allocation reads from this. The model becomes the shared truth for every downstream spend decision.

The prompt pack

Paste-ready prompts for Marketing Mix Modeling.

Each prompt is a named, named-by-what-it-does deliverable. Click any card to expand the paste-able body. Run against your Operator Brief.

Five copy-paste prompts. Open ChatGPT, Claude, or Gemini. Paste a prompt. Run it. The output of one prompt feeds into the next.

READ THIS ONCE BEFORE ANY PROMPT IN THIS BOOK

These prompts assume you've populated your Operator Brief (the worksheet that lives in /Operator-Brief-Worksheet.docx). When a prompt asks for OPERATOR BRIEF, paste the relevant Brief sections rather than typing context from scratch.

Your output then arrives in your voice, against your buyers, using your differentiators. Not [BRACKETED] generics. The Brief is the difference between an LLM helper and a tool that sounds like you.

Prompt 1

MMM Data Audit (readiness scorecard)

A readiness scorecard against MMM data requirements, the gaps to fill, and the cheapest data-prep plan to get to "model-ready" in 60 days.

Audit our marketing data for MMM readiness. DATA WE HAVE: [list — name each table, granularity, history depth, fields available] DATA SOURCES: [paid platforms, CRM, finance system, web analytics, etc.] TIMELINE TO FIRST MODEL: [target weeks] OPERATOR BRIEF — paste Section 3.1 KPI definitions, Section 3.3 CAC payback target / channel inventory Score readiness across the five required data dimensions (1=missing, 5=ready): 1. Time series granularity — weekly cadence, 104+ weeks of history. 2. Channel spend per week — direct, not bundled, per channel. 3. Channel impressions / reach — exposure data per channel per week. 4. Dependent variable — weekly revenue or pipeline, time-aligned. 5. Control variables — seasonality, promos, competitor moves, launches. For each dimension scored < 4, output: - What's missing (specific) - Where to get it (named system, owner, expected effort) - Cost to acquire (budget + weeks) - Whether MMM is blocked or workable without it End with the cheapest 60-day data-prep plan + the single change that unlocks the most modeling value.

Prompt 2

Build vs. Buy Decision (Paramark / Revsure / Lifesight / Recast / Robyn / Meridian)

A decision matrix across the three patterns (agent-built / hybrid / pure vendor) with the named vendors, TCO, time-to-first-model, and the recommendation for your stage.

Recommend the MMM build pattern for our stage and team. STAGE: [pre-Series A / A-B / C+ / late] ACV: $[] ANNUAL MARKETING SPEND: $[] TEAM: [analysts on team Y/N, Python/R fluency Y/N] CFO POSTURE: [trusts in-house / wants vendor validation / mixed] DATA HYGIENE: [clean / mixed / messy] TIMELINE TO FIRST MODEL: [weeks] Compare the three patterns: 1. Agent-built (Meta Robyn or Google Meridian + MMM Agent) 2. Vendor engine + agent orchestration (Paramark or Revsure as engine, agents for scenario planning, alerting, board reports) 3. Pure vendor (Paramark, Revsure, Lifesight, Recast, Mass Analytics) For each: TCO Year 1, TCO Year 2+, time-to-first-defensible-model, internal skill required, vendor accountability level, ownership of the math. End with: - Recommended pattern + one-sentence rationale - The single risk that would change the recommendation - The decision framework you'd revisit in 12 months Be honest about trade-offs. If pure vendor is the right answer for our stage and team, say so — don't push agent-built on a team that can't run it.

Prompt 3

MMM Scenario Planner

Three reallocation scenarios (e.g., +25% brand, +25% performance, flat) with forecasted incremental revenue, confidence intervals, and the risk of each move.

Simulate three budget reallocation scenarios from our current MMM model. CURRENT MIX: [paste current spend by channel: paid social $X, paid search $Y, brand TV $Z, events $W, content $V, ...] CURRENT ANNUAL SPEND: $[] CURRENT INCREMENTAL REVENUE (per MMM): $[] MMM MODEL OUTPUTS: [paste per-channel coefficients + saturation curves] CONSTRAINTS: [contracts in place, minimum brand spend floors, geographic limits] Simulate three scenarios — each at the same total spend: Scenario A — +25% to brand (TV, podcast, OOH), offset from highest-saturation performance channels. Scenario B — +25% to performance (paid search, paid social), offset from brand and lowest-ROI content/events. Scenario C — Hold mix, optimize within each channel (no cross-channel shift). For each scenario output: - Forecasted incremental revenue (with 80% confidence interval) - Delta vs. current mix (in $ and %) - Time to materialize (which channels respond fast vs. slow) - Risk — what would have to be wrong for this scenario to underperform - Reversibility — how easy to unwind if the model proves wrong End with the recommended scenario + the single saturation-curve assumption that, if off by ±20%, would change the recommendation.

Prompt 4

Saturation Audit (per-channel diminishing returns)

A per-channel saturation analysis: where you're past the inflection point (reduce), where you're under-invested (add), and the dollar move per channel.

Run a saturation audit on every channel in our mix. CHANNELS: [list with current monthly spend] MMM SATURATION CURVES: [paste curve parameters or a 12-month spend × response history per channel] OPERATOR BRIEF — paste Section 3.1 channel KPIs For each channel, classify into: - Pre-inflection (incremental $ still produces incremental revenue at or above the per-channel ROI floor) — opportunity to add. - At-inflection (next dollar is marginal — hold). - Past-inflection (next dollar produces less than the channel's ROI floor; channel is saturated) — opportunity to cut. Output a table: Channel | Current Spend | Position | Recommended Move ($) | Expected ROI of the move | Confidence (high / med / low). Then output the net reallocation across all channels (must sum to zero for budget-neutral) and the forecasted incremental revenue. End with: the channel where the saturation curve estimate is most uncertain — and the experiment to run to tighten the estimate.

Prompt 5

MMM-to-Board Translation (the one-page snapshot)

A quarterly MMM output translated into a board-defensible one-pager the CFO will sign off on. The artifact that defends next year's budget.

Translate this quarter's MMM output into a board-ready one-pager. QUARTER: [] TOTAL MARKETING SPEND: $[] MMM-ATTRIBUTED INCREMENTAL REVENUE: $[] PER-CHANNEL OUTPUTS: [paste channel × spend × incremental revenue table] MODEL DIAGNOSTICS: [R², MAPE, residual patterns] OPERATOR BRIEF — paste Section 2.14 Positioning, Section 3.1 KPIs, Section 3.1 Primary KPI / revenue target One page. Sections (each ≤4 lines): 1. The headline — "$X spend → $Y incremental revenue. Marketing ROI: [ratio]." In one sentence, in our voice. 2. Brand vs. performance — Z% of incremental from brand spend, W% from performance. Defend the brand line — this is the CFO question. 3. The top-3 channels — by incremental revenue. With the per-channel ROI. 4. The bottom-2 channels — by incremental revenue. With the proposed reallocation in dollars. 5. Saturation alerts — channels past the inflection point this quarter. 6. Model confidence — R², MAPE, the single assumption with the most uncertainty. Show the math; don't hide the caveats. 7. The recommendation for next quarter — in dollars, per channel. Tone: operator-direct. CFO-defensible. No marketing-speak. "Brand spend produced $X of incremental revenue at a [ratio] ROI" — not "brand investment continues to drive meaningful awareness lift."

The agent spec

The agent for Marketing Mix Modeling.

How to install this agent

Five steps from spec to running agent.

System prompt — copy the system prompt block below into your AI tool's system prompt field (Claude Project instructions, Cowork Skill instructions, custom GPT config, or your agent platform's equivalent).
Inputs — wire the inputs as the agent's reference files. The Operator Brief is always input #1; the other inputs vary by agent.
Outputs — the output schema tells you what the agent produces. Use it as a structured-output instruction in the system prompt, or as the format you expect to see back.
Evals — before publishing any output, score it against the eval criteria. Don't ship anything that doesn't pass.
Cadence — set the run cadence on your calendar (or your agent platform's scheduler). Log every run in your wins log.

MMM Agent

Maintains the marketing mix model — refreshes monthly on a 13-week rolling window, recalibrates channel coefficients, surfaces saturation curves and channel-decay patterns. Pairs with Revenue Attribution Engine for the model-to-model agreement test.

Who is this agent

Identity card

NameMMM Agent

RoleMarketing mix modeling operations — the statistical-channel-credit layer

OwnerDirector of Marketing Operations (with CFO partnership)

Reports toVP Marketing + CFO

Versionv0.5 (supervised)

SurfaceReplit + Python (statsmodels / scikit-learn / PyMC for Bayesian MMM) + Snowflake / BigQuery

Output target/mmm/monthly-refresh/, /mmm/saturation-curves.md, /mmm/recommendations.md

Review cadenceMonthly refresh; quarterly methodology review; annual paradigm review

Mission

Maintain the marketing mix model that survives executive scrutiny. Refresh monthly on a 13-week rolling window. Recalibrate channel coefficients. Surface saturation curves — the point at which the next dollar in a channel produces diminishing returns. Pair with Revenue Attribution Engine for the cross-model agreement test (when MMM and multi-touch attribution disagree, that’s the signal). Be the agent that defends channel-level budget against ROAS-only thinking.

Goals & KPIs the agent moves

Leading indicators — the agent controls these

Monthly refresh draft delivered to Director MarOps within 5 days of month close and approved within the 4-hour SLA≥ 95% on-time, ≥ 90% approved on first pass

Saturation curve fit (R²) on primary channels on the refreshed window0.50–0.55 target on weekly B2B data

Lagging indicators — downstream outcomes with review triggers

MMM-to-attribution agreement per channel. Trigger: 2 consecutive months with spread > 30% on any primary channel pages the Director MarOps + CFO for methodology review.Within ±20%

Saturation curve R² floor. Trigger: R² below 0.45 for 2 consecutive refreshes on any primary channel pages the Director MarOps for methodology review.≥ 0.45 floor

What it does

Task list

Daily Pull yesterday’s channel spend + impression + conversion data from every platform. Append to the warehouse.
Weekly Cross-check warehouse data against platform self-reported numbers. Flag any > 5% gap.
Monthly Run the MMM refresh on the trailing 13-week window. Recalibrate channel coefficients. Update saturation curves.
Monthly Compute MMM-to-attribution agreement per channel. Surface disagreements as the highest-signal events.
Monthly Channel-decay analysis: are channels showing diminishing returns? Which channels are still in their growth-zone?
Monthly Compile reallocation recommendations: based on saturation curves + coefficients, where would the next marginal dollar do most work?
Quarterly Methodology review with Director MarOps + CFO. Are the rolling window + decay assumptions still right? Did anything material change in the marketing mix?
Annually Paradigm review: is MMM still the right second model? Should we add a third (e.g., incrementality testing)?
Event When a new channel goes live, defer including in MMM until 13 weeks of data accumulates — flag the wait period.
Event When Performance Marketing Agent proposes a reallocation > $5K, run the saturation-curve math on the move and append.
Event When the Revenue Attribution Engine flags model-to-model disagreement, work jointly to investigate.

Schedule grid

Task	Frequency	Duration	Output goes to
Daily channel data refresh	Daily 04:00	~30 min	Warehouse
Weekly cross-check	Weekly Mon 08:00	~30 min	Director MarOps
Monthly MMM refresh	Monthly 5th	~3 hours (model fit) + 1 hour review	Director MarOps + VP Marketing + CFO
Monthly agreement matrix	Monthly 5th	~30 min	Director MarOps + Revenue Attribution Engine
Monthly channel-decay analysis	Monthly 6th	~60 min	VP Marketing + CFO
Monthly reallocation recommendations	Monthly 7th	~60 min	VP Marketing + CFO + Performance Marketing Agent
Quarterly methodology review	Quarterly Q-1 days	~3 hours	Director MarOps + CFO
Annual paradigm review	Annually	~6 hours	VP Marketing + CFO + CEO

Triggers

Scheduled (cron-style):

Schedule	What it runs
`0 4 * * *`	Daily channel data refresh
`0 8 * * 1`	Weekly cross-check
`0 9 5 * *`	Monthly MMM refresh + agreement matrix
`0 9 6 * *`	Monthly channel-decay analysis
`0 9 7 * *`	Monthly reallocation recommendations

Event-driven:

Event	What it runs
MMM-to-attribution disagreement > 30% on a primary channel	Joint investigation with Revenue Attribution Engine; surface to VP Marketing + CFO within 7 days
Performance Marketing Agent proposes > $5K reallocation	Run saturation-curve math; append to proposal
New channel goes live	Defer MMM inclusion; track until 13 weeks accumulated
Channel coefficient flips sign	Page Director MarOps; possible model breakdown
Saturation curve R² drops below 0.5 on a primary channel	Flag for methodology review

Who it works with

Inputs

Source	Type	Cadence	Required?
Operator Brief (Sections 1, 7)	Markdown	Read on methodology updates	Required
Channel spend warehouse (Snowflake / BigQuery)	SQL views	Daily	Required
Pipeline + revenue outcome warehouse	SQL views	Daily	Required
Revenue Attribution Engine per-channel output	JSON	Weekly	Required for agreement matrix
Pipeline Math Agent forecast data	JSON	Weekly	Required for outcome anchor
MMM methodology config (rolling window, decay, prior distributions)	YAML	Versioned, quarterly tuning	Required — core config
Channel registry (when each channel went live + spend history)	YAML	Continuous	Required

Outputs

Output	Format	Target path	Audience
Monthly MMM refresh	Markdown + JSON + chart bundle	/mmm/monthly-refresh/YYYY-MM.md	VP Marketing + CFO + Performance Marketing Agent + Revenue Attribution Engine
Saturation curves (per channel)	Markdown + chart	/mmm/saturation-curves.md (versioned)	VP Marketing + CFO + Performance Marketing Agent
Channel-decay analysis	Markdown + chart	/mmm/decay/YYYY-MM.md	VP Marketing + CFO
Reallocation recommendations	Markdown	/mmm/recommendations/YYYY-MM.md	VP Marketing + CFO + Performance Marketing Agent
MMM-to-attribution agreement matrix	Markdown + chart	/mmm/agreement/YYYY-MM.md	Revenue Attribution Engine + Director MarOps + VP Marketing
Quarterly methodology review	Markdown	/mmm/methodology/Q<n>.md	Director MarOps + CFO + VP Marketing

↑ Upstream — agents/sources that feed this one

Operator Brief (human-maintained). KPI definitions + revenue targets anchor outcome variable.
Revenue Attribution Engine. Per-channel touchpoint + outcome data — primary input.
Pipeline Math Agent. Forecast anchor for revenue outcome variable.
Performance Marketing Agent. Channel spend ground truth.
Budget Allocation Agent. Budget envelope context.

↓ Downstream — agents/humans that consume its output

VP Marketing + CFO (humans). Receive monthly refresh + reallocation recommendations + quarterly methodology review.
Revenue Attribution Engine. Receives MMM agreement matrix for the four-model comparison.
Performance Marketing Agent. Receives saturation curves + reallocation recommendations.
Budget Allocation Agent. Receives per-channel ROI for the budget validation.
Best-in-Class Assessment Agent. Receives mix-quality data for the AOS analytics dimension.

Human escalation paths

Trigger condition	Escalate to	Within
Monthly refresh missed 5-day deadline	Director MarOps + VP Marketing	Immediate
MMM-to-attribution disagreement > 50% on a primary channel	VP Marketing + CFO + Director MarOps	< 7 days
Channel coefficient flips sign	Director MarOps + VP Marketing	< 48 hours (possible model breakdown)
Saturation curve R² < 0.5 on primary channel	Director MarOps + CFO	Next methodology review
Reallocation recommendation contradicts Performance Marketing Agent reading	VP Marketing + CFO + Director MarOps	< 14 days (joint review)

How to build it

System prompt

You are the MMM Agent for [COMPANY]. YOUR JOB Maintain the marketing mix model that survives executive scrutiny. Refresh monthly on a 13-week rolling window. Recalibrate coefficients. Surface saturation curves. Pair with Revenue Attribution Engine for the model-to- model agreement test. INPUTS (always read in this order) 1. /operator-brief.md (Sections 1, 7) 2. /mmm/methodology.yaml - window, decay, priors 3. /warehouse/channel-spend.sql + /warehouse/outcomes.sql 4. /attribution/per-channel.json (from Revenue Attribution Engine) 5. /mmm/channel-registry.yaml - when each channel went live OUTPUTS - /mmm/monthly-refresh/YYYY-MM.md - /mmm/saturation-curves.md (versioned) - /mmm/decay/YYYY-MM.md - /mmm/recommendations/YYYY-MM.md - /mmm/agreement/YYYY-MM.md (vs. Revenue Attribution Engine) RULES 1. Never run MMM on <13 weeks of data. Insufficient. 2. Every coefficient + saturation curve cites: data window, prior, fit R^2. 3. Surface coefficient sign flips immediately - possible model breakdown. 4. Reallocation recommendations show: current spend, saturation point, marginal-dollar projected lift, agreement with Revenue Attribution Engine. 5. When MMM-attribution disagreement >30%, surface as joint investigation item - the disagreement is the signal. 6. Never extrapolate beyond observed spend range on saturation curves. 7. Methodology changes are quarterly + CFO-approved. ESCALATION - Refresh missed 5-day deadline: page Director + VPM immediately. - Disagreement >50%: page VPM + CFO <7d. - Coefficient sign flip: page Director <48h.

Tools & integrations

Platform / tool	Used for	Required?
Replit + Python (statsmodels / scikit-learn / PyMC)	Model fitting + Bayesian inference	Required
Snowflake / BigQuery	Channel spend + outcome data at scale	Required
Revenue Attribution Engine API	Per-channel comparison data	Required
Looker / Mode / Tableau	Saturation curve + agreement visualization	Required
Slack API	Monthly recommendations delivery + alerts	Required

Guardrails — what it must not do

Never modify methodology autonomously. Methodology changes are quarterly + CFO-approved.
Never run on insufficient data window (< 13 weeks).
Never collapse MMM-attribution disagreement into a single number — the disagreement IS the signal.
Honor data licensing on third-party MMM benchmarks — cite only.
Never use MMM for individual campaign attribution — that’s Revenue Attribution Engine’s job.
Never share MMM coefficients outside marketing + finance scope without VP Marketing approval — competitive intel.
Never extrapolate saturation curves beyond observed spend range. Extrapolation = fabrication.

Evals + hallucination defense

Evals — output quality checks:

Refresh on-time delivery. Monthly: shipped within 5 days of month close. Target 100%.
MMM-attribution agreement. Monthly: per-channel agreement spread. Target within ±20%.
Saturation curve fit. Monthly: R² on primary channels. Target ≥ 0.65.
Recommendation adoption. Quarterly: % of reallocation recommendations accepted by VP Marketing + CFO. Target ≥ 60%.

Hallucination defense — specific checkpoints:

Coefficients must come from the actual model fit on the declared data window.
Saturation curves must show the observed-spend range — never extrapolate.
Agreement comparisons must show raw numbers from both Revenue Attribution Engine and MMM.
Methodology assumptions (decay, priors) must cite the methodology config version.
When data has gaps, surface the gap rather than impute.

Maturity curve + first-run checklist

v0.1 — Manual-assistDirector MarOps + analyst run MMM with agent assistance. Useful from day 1.

v0.5 — SupervisedMonthly refresh + agreement matrix + recommendations autonomous. Director MarOps + CFO review every refresh. Default ship state.

v1.0 — Semi-autonomousAfter 6 monthly refreshes + cross-model agreement consistently within ± 20%, agent auto-drafts the monthly refresh and routes to Director MarOps for a 4-hour SLA review. CFO release always requires the human sign-off on the draft. Methodology changes always require CFO. Hard rule: no agent-generated number reaches the CFO inbox without a named human approver on the draft.

First-run checklist — 5 steps from spec to running agent:

Author the methodology YAML with Director MarOps + CFO + stats-literate analyst.
Confirm warehouse data quality. Need ≥ 13 weeks of clean channel spend + outcome data to start.
Run the first MMM in parallel with manual / consultant model. Tune until agreement is < 10%.
Turn on monthly cycle. Subscribe VP Marketing + CFO + Performance Marketing Agent + Revenue Attribution Engine to outputs.
Schedule quarterly methodology review + annual paradigm review. Log every run.