Choose a CRM Validation Sample Without Overlooking High-Risk Records
Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.
A migration team can review hundreds of records and still miss the opportunity assigned to the wrong owner, the account linked to another customer's contacts, or the opt-out value reversed during import. The problem is not simply a small sample. It is choosing records without considering how failures happen or what they would disrupt.
A useful CRM migration validation sampling plan combines three layers: mechanical checks across the entire imported population, manual review of high-impact and transformation-heavy records, and a separate random selection. Before reviewers begin, define which failures require a wider review and which defects block release.
The result is not a guarantee that every record is correct. It is a documented allocation of validation effort, with clear evidence, known gaps, and a response when the evidence turns against the migration.
1. Define the population and the decisions validation must support
Start with a fixed migration scope, not a spreadsheet of convenient records. Record the source extract version, destination import batch, included objects, exclusions, and the time boundary for source changes. Otherwise, a reviewer may mistake a legitimate update for a migration defect—or compare two different populations without noticing.
Define the unit of review as well. An account review might include its contacts, opportunities, ownership, and activity associations. That is different from checking one account row. Count selected primary records separately from related records inspected so that the reported sample size remains meaningful.
Validation should support explicit decisions:
- Can the imported data be released for ordinary use?
- Must a particular cohort remain unavailable or restricted pending repair?
- Does a mapping or transformation need correction and revalidation?
- Is the evidence insufficient to make a release decision?
Write down expected results before inspection. For a merged record, this includes the intended surviving identifier, field-precedence rules, and expected relationships. Without those expectations, reviewers may approve a plausible destination value that is nevertheless wrong.
2. Check the whole population for mechanically detectable failures
Do not spend scarce manual review time discovering failures that a reconciliation query or export comparison can identify across every record. Use available exports, scripts, or reporting tools to check the full population. Confirm what the workspace and current plan support before relying on a particular capability.
A recommended mechanical check set includes:
- Population reconciliation: Account for every source record as imported, intentionally excluded, merged, rejected, or otherwise resolved. Compare by object and batch, not only by a grand total.
- Identity reconciliation: Check that each expected source identifier maps to the intended destination record. Matching counts alone can hide one omission and one duplicate.
- Required values: Find missing identifiers, mandatory fields, owners, or other values required by the agreed mapping.
- Relationships: Detect missing referenced records, orphaned relationships, and unexpected relationship counts.
- Transformation rules: Compare mapped statuses, converted dates, normalized values, and other deterministic outputs against their documented rules.
- Unexpected changes: Flag values that should have been preserved but changed, including suppression or permission-related fields where these are in scope.
Distinguish a reconciliation failure from an approved exclusion. A deliberately excluded archived record should not count as an unexplained loss. Equally, an exclusion needs a recorded reason; it should not become a convenient label for an import failure.
Mechanical checks also have limits. A relationship can point to an existing account and still point to the wrong account. A status can be valid but semantically incorrect. Those are important reasons to retain manual review.
3. Allocate manual review by consequence and transformation risk
Build three selection lists before opening destination records. Keep their selection reasons visible throughout the review.
High-impact records
Choose records whose corruption would cause disproportionate disruption. Candidates include active opportunities with imminent next steps, shared accounts with complex ownership, unresolved service escalations, and records whose communication restrictions must be preserved. Let business owners identify critical records, but require a specific reason rather than accepting an unbounded list of favourites.
If this cohort is manageable, inspect every record in it. If it is too large, divide it further by consequence and state explicitly which portions receive complete review and which remain sampled.
Transformation-heavy records
Select records exposed to complicated migration logic: merges, split fields, owner substitutions, status remapping, ambiguous date formats, long text handling, and relationship reconstruction. Cover each transformation rule and its important branches. Reviewing many ordinary records does not compensate for never inspecting a rare merge path.
Random selections
Draw a separate random sample from the remaining eligible records. This gives ordinary records a route into review that does not depend on someone anticipating their failure mode.
Freeze the eligible list and preserve the draw method or random seed where practical. Avoid choosing the first rows in an export, alphabetically convenient accounts, or records that reviewers already know. These are convenience selections, not random selections.
4. Filled hypothetical sampling plan
Fictional worked example: Every quantity, cohort, timing estimate, and result in this example is hypothetical. The plan illustrates a recommended operating design, not a statistical sample-size recommendation or an existing product control.
A team is validating an import of 10,000 primary records: 2,000 accounts, 7,000 contacts, and 1,000 opportunities. Source changes are paused for the comparison period. Related records are inspected as context but are not counted as additional primary selections.
Whole-population checks
- Coverage: All 10,000 expected records, plus the documented exclusion and merge ledger.
- Checks: Source-to-destination identity reconciliation, required fields, duplicate destination mappings, owner mappings, relationship integrity, and deterministic transformation comparisons.
- Owner: Migration analyst, with business-owner review of exceptions.
- Gate: No unexplained record loss, unresolved duplicate identity mapping, or unresolved critical relationship or restriction defect.
Manual group A: high-impact records
- Selection: All 60 records on the agreed critical list: 25 active opportunities, 20 sensitive account structures, and 15 contacts with communication-restriction histories.
- Review: Identity, owner, relevant dates and statuses, relationships, restriction values, and the next intended business action.
- Evidence: Source and destination identifiers, expected result, observed result, reviewer, and defect reference where applicable.
- Expansion trigger: Any critical defect blocks release of the affected scope and starts a review of every record sharing the suspected cause.
Manual group B: transformation-heavy records
- Eligible cohort: 1,800 records passed through one or more complex transformations. Twenty are already in group A.
- Selection: 120 additional, distinct records: 30 merge cases, 30 owner or status remappings, 30 date conversions, and 30 relationship reconstructions. Each record receives one primary selection category, even if several rules apply.
- Coverage rule: Include every documented rule branch and boundary condition. If 120 selections cannot cover them, add records rather than silently dropping branches.
- Expansion trigger: A confirmed transformation defect triggers a full-population mechanical check for that rule where possible, plus manual review of affected cases whose correctness cannot be determined mechanically.
Manual group C: random selections
- Selection: Randomly draw 100 distinct records from the 9,820 records remaining after groups A and B are removed.
- Review: Use the same core checklist, adding object-specific checks.
- Coverage check: Inspect the draw for missing objects, source systems, or import batches. Add targeted coverage records if needed; do not replace or relabel the original random selections.
- Expansion trigger: Any material defect requires cause analysis. Two defects sharing a cause trigger review of that affected cohort, but a single systematic defect can trigger the same action immediately.
The baseline manual workload is 280 distinct primary records, plus any coverage additions and failure-driven expansion. At a hypothetical average of six minutes per review, initial inspection requires about 28 reviewer-hours. Investigation, repairs, evidence preparation, and retesting need separate capacity.
5. Give reviewers a checklist that tests meaning
A reviewer needs more than a pass/fail column. Supply the source evidence, destination identifier, mapping expectations, and selection reason. Where the intended result differs from the source because of an approved transformation, show both the original value and the expected transformed value.
Use a compact checklist:
- Identity: Is this the intended person, account, or opportunity—not merely a similar name?
- Relationships: Are parent accounts, contacts, opportunities, and relevant activities attached correctly?
- Business meaning: Do owner, stage, status, amount, date, and restriction fields retain their intended meaning?
- Transformation: Did the applicable rule produce the specified result, including boundary cases?
- Usability: Can an authorised reviewer find and interpret the information needed for the intended workflow?
Validate workflow readiness through inspection or an authorised, controlled test. Do not send live messages, alter live opportunities, or trigger downstream actions merely to prove that a record looks usable.
Record an inaccessible record as not validated, not passed. Likewise, missing source evidence is an evidence gap. It should not disappear into an apparently clean completion rate.
6. Expand by suspected cause, not just by sample size
Adding another arbitrary batch after a failure can create more work without answering the important question: which records share the defect?
Classify defects by consequence and likely mechanism. A wrong account association, lost restriction, or materially incorrect opportunity value may require an immediate hold on the affected scope. A cosmetic formatting difference may not. Agree on those categories before review so that release pressure does not redefine severity.
For each confirmed defect:
- Preserve the failing identifiers and comparison evidence.
- Identify the candidate cohort using transformation rule, source system, import batch, object, or other relevant exposure.
- Check the entire cohort mechanically where a reliable test can express the failure.
- Manually inspect cases that require interpretation, including distinct branches and exceptions.
- If the cause or boundary remains uncertain, widen the hold and review rather than assuming the defect is isolated.
Hypothetical failure: A reviewed opportunity has a shifted closing date because one conversion branch handled an ambiguous source format incorrectly. The response is to identify every record exposed to that branch, test its expected date, and repair affected records. Reviewing another unrelated set of opportunities would not resolve the known risk.
7. Retest the repair and its surrounding effects
A corrected example proves only that the example was corrected. Before reopening the affected scope, rerun the failed check across the identified cohort and inspect the repaired manual cases.
Also check for collateral effects. A merge repair may correct one account while changing contact relationships. An owner-map correction may affect records that initially passed. Rerun the relevant population checks and select fresh manual cases from the repaired cohort, rather than relying exclusively on records used to develop the fix.
Keep validation runs separate. Record which import or repair version each result describes. Do not combine passes from an earlier import with passes from a materially changed import and present them as one clean result.
If the team cannot identify the affected cohort, cannot reproduce the defect reliably, or cannot restore a trustworthy comparison baseline, that uncertainty is itself a reason to withhold approval of the uncertain scope.
8. Report coverage and limits, then make the release decision
A useful validation summary reports population checks completed, distinct manual records reviewed, coverage additions, unresolved evidence gaps, defects by severity, affected cohorts, repairs, and retest results. Keep targeted and random results separate: targeted selections were deliberately chosen for risk and should not be treated as representative of the whole import.
State the limits plainly. A clean manual sample does not establish that every record is correct, provide a population error rate, or prove that untested workflows will work. A full-population mechanical check establishes only what its comparison logic can detect. Weak expectations or incomplete source evidence can weaken both layers.
Next action: Before the next validation session, freeze the population list, name the three selection owners, and fill in the first review roster. Confirm that every critical record has a disposition, every transformation branch has coverage, and every failure category has an expansion rule. Those observable conditions make the plan actionable before anyone starts marking records as passed.
Put the next conversation to work.
Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.
Explore Communications