Outlook
Part of A monitoring framework for a marketing attribution outlook
Four attribution risk scenarios, from silent AI changes to supplier exit
Prepare attribution risk scenarios with observable triggers, affected records, accountable owners, containment, rollback proof and explicit expiry.
What to take away
- Build each attribution risk scenario as a rehearsal record with an observable trigger, affected records, an owner, containment, recovery evidence, rollback and expiry.
- Likelihood and loss estimates stay blank because no buyer incident history or defensible probability model was supplied.
- Write the manual baseline before buying a tool, following selecting attribution software after a manual baseline.
- Rehearse with synthetic or authorised fixtures, record defects, and reopen the register after any source, method, supplier or data-flow change.
- The source review closed on 6 September 2026 for an England operating context, with UK-wide legal records.
Four scenario records in one table
Each record carries the same seven fields. The table puts the four trigger types side by side, and the detail follows.
Four scenario records
| Scenario | Observable trigger | Owner | Containment | Expiry |
|---|---|---|---|---|
| Population and consent change | Tag, pixel or permission change | Measurement lead, privacy review | Pause claims on the broken series | New baseline published |
| Silent AI component change | Model or schema change with no version record | AI service owner, statistical reviewer | Stop automated use, keep logs | Monitoring and supplier notice verified |
| Credit reported as causal effect | Modelled revenue called incremental without a counterfactual | Evaluation reviewer | Withdraw causal wording | Decision no longer needs the claim |
| Supplier exit | Contract, export or retention change | Commercial lead, security and legal | Freeze changes, export what is allowed | Deletion and continuity duties evidenced |
Scenario one: population and consent changes
Trigger: a tag, pixel, script or permission change. Platform observations and reconciled journeys are both affected, along with any trend series spanning the break.
Leading indicators: a tag inventory difference or an abrupt missingness pattern. The owner is the measurement lead, with privacy and PECR review.
Containment: pause comparisons and budget claims that depend on the broken series. Rollback: restore the last authorised configuration where lawful and safe. Expiry: publish the new baseline and its limits first.
Observable population change response
- Trigger: tag, pixel, script, purpose or permission change
- Indicators: tag inventory difference, missingness, unmatched records
- Owner: measurement lead with privacy and PECR review
- Containment: pause comparison and budget claims
- Recovery: data map, configuration record, test fixture, break marker
- Rollback: restore last authorised configuration
- Expiry: publish new baseline and limitations
Observed current evidence, UK: The ICO's storage and access technologies guidance, finalised on 29 April 2026, makes a tag or permission change here a compliance decision the buyer must document.
Scenario two: a silent AI component change
Trigger: a model, prompt, API or schema change with no approved version record. Classification and explanation output can both shift.
Leading indicators include a failed reproduction or a shift in error categories, as testing AI attribution output sets out. The owner is the AI service owner with a statistical reviewer.
Containment: stop automated use and preserve inputs, outputs and logs. Rollback: return to a known approved version or the manual baseline stored outside the tool. Expiry: change monitoring and supplier notice routes are verified.
Silent AI change response
- Trigger: model, prompt, API or schema change without version record
- Indicators: reproduction fails, error categories move, wording inconsistent
- Owner: AI service owner and statistical reviewer
- Containment: stop automated use, preserve inputs, outputs, logs
- Recovery: repeated fixtures, error review, security approval, human route
- Rollback: known approved version or manual baseline
- Expiry: verify change monitoring and supplier notice routes
Observed current evidence, international: The NCSC's secure AI system development guidelines cover logging, monitoring, updates and incident management. For a buyer, those are the logs to demand at contract stage.
Scenario three: credit reported as causal effect
Trigger: a paper describes platform-assigned or modelled revenue as incremental without a credible counterfactual. Budget recommendations, ROI statements and public claims are affected.
Leading indicators: causal verbs with no comparison group, and uncertainty ranges that vanish between drafts. The owner is the evaluation reviewer.
Containment: withdraw causal wording and label the output as credit allocation. Expiry: close the record when the decision no longer needs causal language.
Observed current evidence, UK government evaluation: HM Treasury's QPIE guidance, updated in 2026, treats counterfactual quality as the limit on causal claims. The methods comparison keeps causal language held in reserve for that reason.
Scenario four: supplier exit blocks reproducibility
Trigger: contract, export, model access or retention terms change. Records at risk include authorised inputs, transformations, model versions and correction history.
Leading indicators include export requests that queue without reply. The owner is the commercial lead with security and legal reviewers.
Containment: freeze destructive changes, export what permissions allow, and preserve audit evidence. Rollback: use the documented manual process. Expiry: deletion and continuity duties evidenced.
Example of a completed scenario record
Illustrative record. Trigger: a model version change with no approved note. Leading indicator: error categories drift from the frozen fixture baseline. Owner: AI service owner. Containment: automated classification paused. Recovery proof: fixture rerun reproduces frozen labels. Rollback: revert to the approved version and the manual baseline. Expiry: next quarterly check.
Before you act
- Define an observable trigger.
- Assign a named owner.
- Specify containment and recovery evidence.
- Document rollback and expiry.
- Rehearse with synthetic fixtures, record defects, and reopen the register after any relevant change.
Common questions
Why is likelihood left blank for these scenarios?
Likelihood stays blank because no buyer incident history or defensible probability model was supplied, so no figure would be supported.
What should be done if an AI component changes silently?
Stop automated use and preserve inputs, outputs and logs under the retention rule. Return to a known approved version or the manual baseline, and retire the scenario only when monitoring is verified.
How can a buyer prepare for a supplier exit?
Freeze destructive changes, export what the contract and permissions allow, and preserve audit evidence. Recovery proof needs a restored dataset and a reproducible report.
Which UK rules and guidance apply?
The ICO guidance covers tags and permissions, the NCSC guidelines cover the AI lifecycle, and HM Treasury's QPIE guidance covers counterfactual quality.
