Skip to content

Privacy and preparation

Portfolio data overlap audit: when two companies hold the same customer records

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Detect duplicate customer records across portfolio companies by inventorying each company's licensable scope, matching on identifiers inside a controlled environment with blocking and probabilistic linkage, and then reviewing every overlap cluster. For each cluster, trace which entity collected the record under which notice, carry opt-outs across, check exclusivity already granted, and record one outcome: license from one company, license from both with consent, or exclude.

Checklist

Portfolio data overlap audit

This page covers portfolio data overlap audit: when two companies hold the same customer records.Illustrative

Key takeaways

  • Overlap is normal in a portfolio: shared customers, add-on migrations and shared CRM instances all put one person in two datasets.
  • Authority follows the transaction documents and the collecting entity's notice, not whichever system holds the cleanest copy.
  • Matching records to find overlap is itself processing; hashed identifiers are still personal data, so keep the crosswalk inside a controlled environment and delete it after the audit.
  • An opt-out or restriction recorded in either company should travel to every copy of that customer before anything is licensed.
  • Count each real-world customer once when describing volume, and never grant two exclusive licenses over the same people.

How do you detect duplicate customer records across portfolio companies before licensing?

Run the overlap audit in four passes: define each company's candidate scope, match records across companies in a controlled environment, review each overlap cluster for authority, privacy and existing commitments, and record a single outcome per cluster. Detection is the easy part. The decisions after detection are where most of the work and the risk sit.

The audit matters because a portfolio rarely has clean borders. A sponsor that bought a platform company and three add-ons may find the same commercial customer in four CRMs, each with different contact details, different notices and different opt-out histories. If two of those companies each license their records separately, the same people end up in two datasets under two sets of terms.

This page is general information for operators and deal teams. The privacy and contract questions in it depend on each company's documents, so counsel reviews the specific facts before any record is licensed.

Why does the same customer appear in two portfolio companies?

Most overlap comes from ordinary business, not from mistakes. Knowing the source of an overlap tells you which documents decide it, so classify each cluster by cause before you judge it.

Why does the same customer appear in two portfolio companies?
Overlap causeWhat it looks likeDocuments that decide itTypical first question
Shared customerA business buys from two sister companies independentlyEach company's own customer terms and noticesDid each company collect the record itself?
Add-on migrationAcquired company's customers copied into the platform CRMPurchase agreement, transition plan, migration logWas the record moved, or copied and left in both?
Shared system instanceSeveral brands use one HubSpot or Salesforce orgSystem ownership, data processing terms between entitiesWhich entity is the controller for each brand's records?
Referral or reseller flowOne company passes leads to anotherReferral or reseller agreement, lead source fieldsWhat did the customer agree to when referred?
Carve-out or divestitureA sold business left records behind or took copiesSeparation agreement, transition services agreementWhich side kept rights in the retained copy?
Group-level servicesShared billing, support or call center across brandsIntercompany service agreementsWas the data shared only for internal administration?

Which entity has authority over each overlapping record?

Authority over an overlapping record belongs to the entity that collected it, as modified by the transaction documents, and is limited by the notice and terms the customer saw at collection. The system that holds the cleanest copy does not settle it, and neither does common ownership on its own.

Privacy law treats related companies in specific, limited ways. Under the California Consumer Privacy Act, an entity that controls or is controlled by a covered business, shares common branding with it and receives consumers' personal information from it is treated as a business too; control is defined by voting shares, control of director elections or controlling influence. The same statute excludes from selling a transfer of personal information as an asset in a merger, acquisition or bankruptcy where the acquirer takes control of the business, provided the information is then used consistently with the statute. Neither rule converts a customer list into something one sister company may license on behalf of another.

In the EU, recital 48 of the GDPR recognizes that companies in a group may have a legitimate interest in transmitting personal data within the group for internal administrative purposes, including client and employee data. Licensing records to an outside AI developer is not internal administration, so that recital does not cover it. In the U.S., the FTC has warned that adopting more permissive practices, such as sharing data with third parties or using it for AI training, through a quiet retroactive change to terms or a privacy policy may be unfair or deceptive. A portfolio cannot fix a restrictive notice at one company by pointing to a permissive notice at another.

Keep four questions apart for every cluster: who can access the record, what privacy permission covers it, who owns the business that holds it, and what contractual licensing rights exist. De-identification can reduce privacy exposure, but it does not create a right to license that was never there.

How do you match records without creating a new privacy problem?

Match inside an environment that one accountable party controls, use only the identifiers needed, and delete the crosswalk once outcomes are recorded. Finding overlap requires comparing personal data across companies, and that comparison is processing in its own right.

Hashing identifiers before matching reduces casual exposure but does not anonymize them. The FTC has explained that the same input always produces the same hash, so a hashed email or phone number can still single out a person. Treat hashed match keys as personal data with the same access controls as the originals.

Exact matches on tax ID, normalized email domain or account number catch the easy overlaps. For the rest, probabilistic record linkage compares several fields at once and estimates match likelihood. Splink, an open-source package from the UK Ministry of Justice, implements the Fellegi-Sunter model for records that lack a shared identifier and notes that it performs best with several columns that are not highly correlated, such as name, address and phone rather than city and postcode. Its documentation also explains why blocking rules matter: comparing every record in a 1 million record table against every other one would generate around 500 billion pairs.

  • Normalize first: legal entity names, addresses, phone formats and email case.
  • Block on stable fields such as postal code plus first letters of the company name, then score within blocks.
  • Set two thresholds: auto-match above one, human review between them, no match below.
  • Sample reviewed pairs to estimate false matches before treating clusters as final.
  • Store only cluster ids and the outcome; delete raw pair scores after sign-off.

Which ten checks settle each overlap cluster?

Use this review once per overlap cluster, or once per group of clusters with the same cause. Each check names the evidence that settles it and the outcomes allowed. A cluster passes only when every check has a recorded answer.

Which ten checks settle each overlap cluster?
CheckQuestionEvidenceAllowed outcomes
1. Match confidenceIs this the same real-world customer?Match score, reviewer decision, sampled error rateConfirmed, rejected, needs review
2. Overlap causeWhy is the customer in both companies?Lead source, migration log, system ownershipOne cause from the cause table
3. Collecting entityWhich company collected each copy, and when?Created-by fields, contract dates, invoicesNamed entity per copy
4. Transaction effectDid a deal move rights in either copy?Purchase, separation or transition agreementsMoved, retained, shared, unclear
5. Notice at collectionWhat did the customer see when each copy was collected?Archived privacy notice and terms by datePermits, silent, restricts
6. Opt-outs and restrictionsHas the customer opted out or objected anywhere?Suppression lists, preference centers, DSAR logsApplied to all copies, or none recorded
7. Customer contract limitsDoes a customer agreement limit data use?MSA data clauses, confidentiality termsNo limit, limited, prohibited
8. Existing licensesHas either company already licensed this customer's records?License register, exclusivity termsNone, non-exclusive, exclusive
9. CountingHow will this customer be counted in volume figures?Cluster id in the scope summaryCounted once, under one company
10. OutcomeWhat happens to the records?Sign-off by both company owners and counselLicense from one, license from both, exclude

Worked example: two service companies, one customer base

This example is illustrative; the companies and counts are invented to show the method. A sponsor owns two commercial HVAC service companies in neighboring metro areas. Company A has about 40,000 customer accounts in its field service system. Company B, acquired later as an add-on, has about 25,000. Both are considering licensing de-identified work order history.

Exact matching on billing address and account phone links about 1,800 accounts. Probabilistic linkage on name, address and phone, blocked by postal code, proposes another 1,400 pairs; reviewers confirm about 1,100 of them after sampling. The audit therefore has roughly 2,900 overlap clusters, all illustrative.

Most clusters are shared customers: property managers who used both companies for different buildings. For these, each company collected its own work orders under its own terms, so each may decide on its own records, but the property manager is counted once in any combined volume statement and appears under one exclusivity commitment at most.

The exception is a group of about 600 accounts that Company A copied from Company B during an early integration attempt. The purchase agreement transferred those customers to the platform, but Company B's old notice limited use to service delivery. Counsel advises excluding the copied records from the first license and revisiting them only after a notice and consent review. One opt-out found in Company B's preference center is applied to the matching Company A records as well.

Why does a duplicate matter to the dataset itself?

Duplicates distort the dataset as well as the rights picture: they inflate stated volume, overweight some customers and can leak between training and test splits. A buyer who splits by record rather than by customer can end up testing a model on the same customer it trained on.

Research on language model training data shows why this is taken seriously. Lee and colleagues found near-duplicate examples in widely used training sets and overlap between training and validation data, and reported that deduplication reduced memorized output and made evaluation more accurate. Business records are not web text, but the same mechanism applies when two companies' work orders describe the same jobs at the same sites.

Joining two companies' records about one customer also adds attributes, and more attributes make a person easier to re-identify. That is a reason to keep overlapping customers in one company's delivery, or to apply de-identification after the overlap is resolved rather than separately in each company.

Which findings should pause or shrink the license scope?

Pause the overlapping records, not the whole program, when a check cannot be answered from documents. Narrow the scope when the problem is confined to one cause or one period, and bring in counsel when authority or notice is genuinely unclear.

  • The purchase or separation agreement is silent on customer data, or the copy you hold was taken outside it.
  • The notice at collection for one copy restricts use, and the records cannot be separated from the permitted copy.
  • Either company has granted an exclusive license that could cover the same customers.
  • An opt-out, deletion request or objection exists in one company and cannot be reliably matched to the other.
  • Match quality is too uncertain to say which records overlap, for example because one company lacks addresses.
  • Records include sensitive categories, such as health or financial account details, that change the privacy analysis.

When should the overlap review run in a licensing assessment?

Run the overlap review after each company's scope is drafted and before any agreement is signed, so that volume figures, exclusivity and exclusions are settled once. Each portfolio company can check whether its own records might qualify through SourceX's short assessment, which asks for no data. Because each company approves its own scope and terms and signs before SourceX receives anything, a recorded overlap outcome can be written into that scope rather than repaired afterwards.

Frequently asked questions

Can a holding company license customer records from all its subsidiaries as one dataset?

Only to the extent each subsidiary has the right to license its own records and the transaction documents allow pooling. Common ownership alone does not give one entity rights in another's customer data. Each overlap still needs an authority, notice and opt-out check.

Is matching customer records across companies itself a privacy issue?

It can be, because comparing personal data across entities is processing. Keep matching inside a controlled environment, use the minimum identifiers, treat hashed keys as personal data, and delete the crosswalk once outcomes are recorded.

If one company's privacy notice allows broader use, can the other company rely on it?

No. Each copy of a record is governed by the notice and terms in force when that company collected it. A broader notice at one sister company does not extend to records collected by another under narrower terms.

Should an opt-out at one portfolio company apply to records at another?

When the audit confirms the same person, the safer practice is to apply the restriction to every copy before licensing. Whether the law requires it depends on the facts and jurisdictions, which counsel should confirm.

How should overlapping customers be counted when describing dataset volume?

Once, under one company, using the cluster id from the audit. Counting the same customer twice overstates volume and can lead to conflicting exclusivity commitments over the same people.

Sources

  • The CCPA treats controlled, commonly branded entities that receive a business's personal information as businesses, and excludes merger and acquisition asset transfers from selling when the data is used consistently with the statute. Source
  • The FTC warned that adopting more permissive data practices, including AI training, through a quiet retroactive policy change may be unfair or deceptive. Source
  • GDPR recital 48 recognizes a possible legitimate interest in transmitting personal data within a group of undertakings for internal administrative purposes. Source
  • The FTC explains that hashing does not anonymize data because the same input always produces the same hash. Source
  • Splink implements Fellegi-Sunter probabilistic record linkage for records without unique identifiers and uses blocking rules because full pairwise comparison grows very large. Source
  • Deduplicating training data reduced memorized output, and train-test overlap affected standard validation sets. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies