Operationalizing analytical signals back into CRM
Computing a metric or a segment is the easy part. Making it arrive safely where people act—and stop driving action when it is no longer true—is the design.
The case in brief
Current reality
Performance and segments are computed in the analytical layer, but people act in the CRM—and without an activation contract, values arrive stale, editable, duplicated or long after they stopped being true.
What must become true
The CRM receives only the governed signals needed for operational action—each with its subject, value or state, as-of date, version, definition, provenance and current or removed status—so updates, removals and reruns are safe and the published state is reconciled against the expected analytical state.
Design question
What should be written back into the operational system, at what grain—and how does a changed or removed signal stop driving action safely?
01The reality · current state
What the data does today.
Account performance, pipeline coverage and a monthly performance segment—On Track, Underperforming, At Risk or No Sale—are computed in the data platform. Sellers and managers work in the CRM and should act there. The first attempt copied every KPI detail into CRM fields that users could overwrite, with no date and no link to the definition; segments were added but never removed, and a rerun after a failure wrote some accounts twice.
- Sellers overwrite derived values they disagree with
- Nobody can tell whether a value is from today or from last month
- Some accounts carry two segments after a rerun; others carry none
- Accounts that left the population keep their last segment for months
- A partially failed writeback goes unnoticed for weeks
- Teams recompute the metric in CRM reports because they do not trust the copy
02The semantic problem
Which meanings are mixed today.
Without an activation contract, every team picks one of the same bad options.
- Recompute in the CRMA second definition of the metric, drifting from the first.
- Query live on every screenValues that cannot drive lists, assignment or automation—and fail when analytics is down.
- Copy the detailAnalytical history in operational fields, without provenance or a freshness promise.
- Add, never removeOld segments keep driving action after the condition changed.
03Key dimension · Activation contract, removals & reconciliation
A signal has a lifecycle—and an explicit exit
What lands in the CRM is small, read-only and self-describing. Every state of the signal carries its version, freshness and provenance, and the lifecycle only closes when what was expected is what the CRM holds.
- 01ComputedDerived from the governed modelCarries Definition version
- 02PublishedReleased as the full expected state for the periodCarries Run version · as-of date
- 03AppliedUpserted in the CRM by key—safe to repeatCarries Delivery state
- 04CurrentRead-only on the account, available for actionCarries Freshness · provenance link
- 05Changed or removedA new value, a new segment or “not eligible”—never a silent blankCarries New version
- 06ReconciledExpected compared with actual; differences have ownersCarries Reconciliation result
The activation contract
| Field | Answers | Rule |
|---|---|---|
| Subject | Which account does it apply to? | The governed business key—never a name |
| Signal | Which value or commercial state? | Exactly one current state per eligible account |
| Grain | At which operational level does it make sense? | Account × signal × period |
| As of | When was it true? | Shown beside the value; stale after its promise |
| Version | Which definition and run produced it? | A recalculation publishes a new version; the CRM keeps the latest |
| Provenance | Where can the evidence be traced? | Links to the definition, owner and sources |
| Actionability | Which downstream process may react? | Declared per signal; the reaction belongs to the automation case |
| Exit | What happens when it is no longer true? | An explicit removal or “not eligible”, applied like any other change |
Expected state against actual state
Eligible accounts for the period, as published
| Key | Expected | In the CRM | Outcome |
|---|---|---|---|
ACC-A12 | On Track | On Track | MatchBoth sides agree. |
ACC-B07 | At Risk | Underperforming | DifferenceWrong state: a rerun applied an older version after the newer one. |
ACC-C33 | Underperforming | — | MissingThe writeback batch stopped half-way. |
ACC-D19 | Not eligible | No Sale | UnexpectedThe account left the population; its removal was never written. |
ACC-E02 | At Risk · v7 | At Risk · v6 | DifferenceStale version: the value is right, but it was produced by an older definition. |
ACC-F58 | No Sale | No Sale | MatchBoth sides agree. |
- Match
Same account, same state, same version.
- Owner
- Nobody
- Action
- Record the check.
- Difference
Wrong state or stale version.
- Owner
- Data platform team
- Action
- Re-apply the current version; the upsert by key is idempotent.
- Missing
Published, not present in the CRM.
- Owner
- Integration operations
- Action
- Replay the missing keys from the run.
- Unexpected
Present in the CRM, not in the published population.
- Owner
- Data platform team
- Action
- Write the removal so it stops driving action.
An analytical signal in the CRM is a governed projection—not a second source of analytical truth. Removal is a first-class state change, and reconciliation is part of activation, not cleanup.
04Grain & contracts
What one row or message means.
Meaning, grain, key, freshness and owner—written down before anything is mapped.
Account signal
The published value or state of one signal for one account and period.
- Grain
- Account × signal × period
- Key
Analytical customer key + signal + period + version- Freshness
- Daily; stale after its promise
- Owner
- Sales operations (definition) · data team (computation)
Segment membership
The current segment of one eligible account for one period.
- Grain
- Account × period
- Key
Account + period + run version- Freshness
- Monthly, rerunnable
- Owner
- Sales operations
Membership change
An entry, change or removal to apply downstream.
- Grain
- Account × change
- Key
Account + period + change type- Freshness
- Within a day of the run
- Owner
- Data platform team
- Chosen grain
- Account × signal × period: the level at which a seller or manager can act.
- Too fine
- Transaction or line detail in the CRM is analytical history, not a signal—and it goes stale.
- Too coarse
- Region or team values cannot drive an account’s list view, assignment or follow-up.
06Population, version & freshness
Publish the full expected state—not only the additions.
Treating only new entries as events makes removals, reclassification and recovery invisible. Publishing the complete expected state for the period makes every one of them explicit.
| Rule | Why |
|---|---|
| The eligible population is explicit | An account with no segment and an account out of scope are different facts |
| Exactly one current state per eligible account | Two memberships—or none—are detectable failures, not edge cases |
| Every run publishes the full expected state | Removals and recoveries are visible as changes, not as absences |
| Changes are applied by run version | A rerun is idempotent; an older version never overwrites a newer one |
| Every value carries its as-of date and freshness promise | A stale value is visibly stale, and new actions are suppressed |
07Candidate architectures
Credible options, judged against these premises.
Query live from the CRM
Users who only look at the number
Cost: Cannot drive list views, assignment or automation; depends on analytics being upCopy every KPI detail and add segments as they appear
Small, stable populations that never shrink
Cost: Volume, staleness, duplicates and removals nobody seesGoverned signals at account grain, full expected state, idempotent apply, reconcile
Signals that should change an operational action
Cost: Needs an activation contract, run versions and a reconciliation stepDesign decision
Publish a small set of governed, read-only analytical signals at an operationally meaningful grain; treat changes and removals as explicit state transitions; apply them idempotently; and reconcile the complete expected state against what the operational system actually holds.
08Resulting architecture
Where the data comes from, where it goes—and how.
Every hand-over is named with its mode. The mode follows the business tolerance for waiting, inconsistency and loss.
Data platformIntegration layerCRM
- 01 · Data platformGoverned modelHands over by Batch
Performance at account × month—see the performance data model.
- 02 · Data platformSignal computationHands over by Batch
Metric values and exactly one segment per eligible account.
- 03 · Data platformExpected stateHands over by API
The full population for the period, with explicit removals.
- 04 · Integration → CRMCRM writebackHands over by Event
Upsert by account, signal and version; read-only fields.
- 05 · CRMAvailable for action
Current signals on the account—what happens next belongs to the automation case.
09Failure & quality
Design the failure path before the happy path.
- Writeback partially succeedsReconciliation: missingReplay the missing keys; upsert makes it safeIntegration operations
- A rerun after a failureSame period, new run versionApply by version; older versions are ignoredData platform team
- An account leaves the populationReconciliation: unexpectedWrite the removal explicitly; downstream reactions close through the automation caseData platform team
- An eligible account gets no segmentPopulation count against segmented countFail the run before publishingSales operations
- Values are older than their promiseAs-of date past the freshness promiseShow “stale” in the CRM and suppress new actionsData platform team
Data quality is a property of the contract
| Dimension | Owner | Detected by | When it fails |
|---|---|---|---|
| Completeness | Data platform team | Population count against segmented count | The run is not published |
| Identity | Master data | Signals whose subject has no current account | Held back and routed to master data |
| Freshness | Data platform team | As-of date against the promise | Shown as stale; no new actions |
| Reconciliation | Integration operations | Expected against actual after every run | Differences routed to their owners |
10Governance
Who keeps the meaning stable over time.
| What | Owner | Cadence | Evidence |
|---|---|---|---|
| Which signals reach the CRM | Sales operations | With every new signal request | Signals published and the actions they drive |
| Freshness promise per signal | Data platform owner | Reviewed quarterly | Stale signals by period |
| Reconciliation differences | Integration operations | After every run | Open differences and their age |
11The second layer
Questions that change the design.
What lands
- Should the value be copied or queried live?
- What is the writeback grain—and what does the user need beside the number?
- Which downstream process may react to it?
Change & removal
- What happens when the metric is recalculated?
- How is “no longer eligible” represented?
- How is a rerun made safe?
Ownership
- Who owns a failed writeback?
- Can the CRM change the analytical value?
- Who owns each class of reconciliation difference?
12Decisions & outputs
What the work produces.
- 01Activation contract
- 02Signal lifecycle
- 03Eligibility & exactly-one-state rules
- 04Removal semantics
- 05Run versioning & rerun rule
- 06Expected-vs-actual reconciliation
- 07Lineage link