Advertising AI agent · MMM & incrementality
Marketing-Mix Modelling & Incrementality AI Agent
Assemble the spend and outcome series, fit and refit the model, design the geo or holdout experiment that tests it, and turn the result into a budget recommendation the analyst accepts, interval attached.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Analyst approval
- Pre-decision
- After launch
- Agent Care
What this agent does
Models the spend, not the decision to move it
In
- 01
Spend and outcome series arrive from ad platforms, the CRM and finance systems.
- 02
Series are aligned to one calendar, one currency and one channel taxonomy.
Reason
- 03
The model is fit and refit against the aligned series, one channel at a time.
- 04
Each channel's prior is logged with its source — an experiment, an assumption or a default.
- 05
A geo or holdout test is designed with its minimum detectable effect and power agreed first.
Decide
- 06
The model's estimate is checked against the experiment, and any gap is surfaced, not resolved.
- 07
Every recommendation is routed to the named analyst before a number leaves the model.
Out
- 08
The specification, the priors, the experiment and every edit are retained against the run.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
The agent fits the model and reads the experiment; the analyst decides which specification to accept, and no modelled figure leaves the page without its interval.
Example workflow
One channel estimate, series to approval
- 1Series receivedSpend and outcome series, a new experiment result or a scheduled refit
- 2Inputs assembledHistoric series, the prior log, calibration history and channel taxonomy, each with its source
- 3Estimate draftedChannel contributions, credible intervals, experiment status and confidence
- 4Controls appliedSpecification checks, prior-source checks, calibration-gap checks and confidence threshold
No human action required
Stages 1 to 4 run unaided, and no figure is recommended at any of them — the agent is modelling, and the analyst's lane opens at the confidence gate.
Goes to the analyst to approve.
Adds a second modeller review first.
Analyst approval
The estimate is held with its interval, its calibration status and the confidence.
Approve · Revise · Escalate
Approved — released to planning
- 6Reporting systems updatedOnly where write access and approval policy allow it
- 7Outcome evaluatedInterval accuracy, calibration gap, edits and post-decision corrections
Edits
Every analyst edit is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Reallocating budget between channels.
- Publishing a modelled figure to the market.
- Setting a channel's ROI prior on its own.
- Declaring a test result before it is powered.
Automation boundary
Agent acts unaided
- ✓Assemble the spend and outcome series and fit the model.
- ✓Design the geo or holdout experiment and its power.
- ✓Attribute each channel's interval to its prior and source for the named owner.
- ✓Flag the gap when the model and the experiment disagree, and hold it for the analyst.
Any output leaves the model only inside the boundaries agreed at implementation, never as fact.
- Presenting a modelled return as revenue already earned.
- Reconciling attribution, incrementality and MMM into one figure.
- Approving the budget shift the recommendation implies.
- Changes to specification, priors or calibration rules.
Example output
One estimate, annotated
Everything the agent produces is attached to the run it was drawn from.
MMM output · single channel
Illustrative example- Channel
- Brand search, always-on
- Finding
- Large modelled contribution, no calibrating holdout run
- Estimate status
- Interval untested
- Model run
- This quarter's model run
- Confidence
- 68%
- Attribution
- Analyst name on file
As received
Taken from this quarter's model run and the series behind it — nothing on this side is asserted as fact by the agent.
Inputs checked
- Spend and KPI series
- Prior source logged
- Holdout test status
Why it's flagged
The contribution is real in the model and a person still decides.
What the score decides
Below the configured threshold the flag picks up a second modeller review before it.
Action
ApproveReviseEscalateValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every channel
From the assembled series
Fit against the series
Draw on the assembled series, the logged priors and the calibration history.
A model is not a measurement
Brand search and retargeting are the classic case: an observational read can show a strong return while a controlled test shows the outcome would have happened anyway — what the experiment exists to catch.
Label the question each number
Attribution, incrementality and MMM measure different things; every output states which one produced it, and the three are never blended into a single figure.
Build an evidence trail
The estimate, the interval around it and the analyst who accepted it stay on the plan.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Media & spend platforms
- Google Ads
- Meta Ads
- TikTok
- Amazon Ads
- DV360
Business & finance data
- ERP
- POS systems
- Revenue and conversion feeds
Modelling tooling
- Meridian
- Robyn
- Custom Bayesian MMM stacks
Agent
Marketing-mix modelling & incrementality
- Fits the model
- Designs the test
- Holds for the analyst
Experimentation & warehouse
- GeoLift
- Meridian GeoX
- BigQuery
- Snowflake
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the plan
Every control wraps the one within it. What the set does not catch is named in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether an estimate may stand
- L3
- Puts the release in the analyst's hands
- L4 – L5
- Keep the estimate and the interval behind it
- L6
- Reverts to descriptive reporting when signals degrade
How Nestack evaluates it
Evaluate the modelling workflow — not only the final number.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the number the analyst sees
- E1Final-output evaluationDid the channel estimate ship with its credible interval?
- E2Step-level evaluationDid the agent use the current series, priors and calibration record?
- E3Tool evaluationDid it read the correct series version and channel ID?
- E4Confidence calibrationDo low-confidence estimates actually diverge more from the experiment?
- E5Slice evaluationHow does interval width change across specific channel types?
- E6Business outcomeHow many estimates needed an analyst edit or a later correction?
Floor — the outcome the plan answers for
Failure modes
Where each failure originates in the agent
Seven failure modes, placed at the stage each one originates.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- MM-03
Stale series pull
The spend or outcome series read is a cached pull, not this period's data.
Stage gathers
Spend and outcome series, priors and calibration history
02 · Reasoning
2 modes
- MM-04
Prior source blurred
A default or assumed prior is shown as if an experiment set it.
- MM-06
Mismatched calibration basis
An experiment's window or population doesn't match the model's, but calibrates it anyway.
Stage proposes
Channel estimates, intervals and confidence
03 · Tool / write
2 modes
- MM-02
Low-confidence release
A recommendation reaches the analyst ahead of its confidence gate.
- MM-05
Modelled return shown as real
A modelled contribution is presented as revenue already delivered.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- MM-01
Interval dropped
A channel estimate reaches the analyst with no credible interval attached.
Stage returns
The estimate the analyst approves
05 · Change / Version
1 mode
- MM-07
Silent specification drift
A prior, adstock or specification change widens what the agent will report as fitted.
Stage tracks
Model, priors, specification and calibration version
- Sev-1 · plan acted on outside the boundary
- Sev-2 · a wrong estimate reaches the analyst
- Sev-3 · series degrades, run routes to review
Affected slices
One channel can carry most of the risk
A single model fit assigns every channel one number; some channels move far more than others when the specification changes. Nestack reports estimate instability by channel, not only in aggregate.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Channels with little spend variation | 7.5% | 3.9× | Review | |
| Brand search and retargeting | 5.7% | 3.0× | Review | |
| Channels changed mid-period | 3.3% | 1.7× | Watch | |
| Large channels with varied spend | 1.9% | 0.7× | Normal |
Bar: estimate-instability lift vs. the large-varied-spend baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
The loop closes on a case, not a story
The loop shuts when the wrong estimate is a regression case, not when it has been explained. That suite is what the next model run is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Interval width or edit rate rises in a channel slice.
- 02
Diagnose
The point estimate everyone quoted is checked against the series, the priors and the specification until one cause explains the swing.
- 03
Improve
Whatever changes ships against a version, with the runs that prompted it attached.
- 04
Verify
Nothing ships until the affected specification cases pass a second time.
- 05
Learn
The suite grows by one case; so does the calibration record.
Learn → Detect
The return edge. The next run starts against a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, modelling workflow, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Analytics discovery and boundary scoping and boundary definition.Week 1
- 02Media, KPI and warehouse source assessment.Week 1, Week 2
- 03Prior-source and calibration rule mapping and rule mapping.Week 1, Week 2
- 04Series ingestion and taxonomy mapping.Week 2
- 05Model-fitting logic and interval binding.Week 2, Week 3
- 06Confidence scoring and gap routing.Week 3
- 07Analyst approval workflow.Week 3, Week 4
- 08Warehouse and experiment-tool integration.Week 5
- 09Specification and prior cases.Week 4, Week 5
- 10Guardrails and acceptance controls.Week 4
- 11Run-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne channel, one market | ProductionProduction data and warehouse access | AdvancedMultiple markets / brands |
|---|---|---|---|
| Introduced at Pilot | |||
| Model fit to your data | ✓ | ✓ | ✓ |
| Analyst approval | ✓ | ✓ | ✓ |
| Calibration baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by channel | — | ✓ | ✓ |
| Escalation workflow in your tools | — | ✓ | ✓ |
| Approved output actions | — | ✓ | ✓ |
| Warehouse integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-market calibration rules | — | — | ✓ |
| Multi-stage analyst approvals | — | — | ✓ |
| High channel volume | — | — | ✓ |
| Multi-market model controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, transaction volume, approval controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your spend, KPI and warehouse access
→Series ingestion and taxonomy mapping
Week 102Representative historical series and any past tests
→Model-fitting baseline and interval attribution
Week 203Your prior-source and calibration rules
→Prior-source and calibration-boundary mapping
Week 104Access to relevant APIs, feeds or exports
→Media, KPI and warehouse assessment, then integration setup
Week 205Estimates you would not want budgeted on
→Experiment cases and failure-mode testing
Week 406What no estimate may be read as
→Confidence scoring, gap routing, guardrails and approval controls
Week 307Named analysts to review recommendations
→Analyst approval workflow, then pilot and production validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Bands follow the real work rather than the plan, which is why evaluation and pilot share week 5.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Analytics discovery, calibration mapping and the boundary
W2Source integration and the model-fitting baseline
W3Modelling workflow, confidence logic and approval controls
W4Evaluation suite, guardrails and failure-mode testing
W5Warehouse integration, pilot runs and targeted corrections
W6One planning cycle modelled under the analytics lead, then handover
- Reading the band
- A bar covers only the weeks its work is named in — the week 5 overlap is real, not padding.
- At the end of W6
- Once the cycle validates, Agent Care owns the running agent.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Advertising AI agent
Build a marketing-mix agent around your measurement stack.
Show us your spend and outcome data, your prior sources and who signs off on a number. The judgement that turns an experiment into a calibrated prior stays with the analyst, not the agent.
- Typical delivery
- 4–6 weeks
- Deployment
- Your spend and KPI data
- Analyst approval
- Pre-decision
- After launch
- Agent Care available
Nestack Agents · Marketing-mix modelling & incrementality
AGT-AM-15 · Agent Care available after launch