Platform Foundation Customer Profiles

One customer. Many touchpoints. One understanding.

The same person shows up as a booking, a loyalty number, an app login and a support call. Resolve them into one trusted profile, carefully: with evidence for every link, and the ability to undo any merge.

  • Precision before coverage
  • Every merge explainable and reversible
  • Consent travels with the profile
Small figurines connected by lines to one larger figure in the centre
4 records, 1 personLinked on loyalty ID + verified email

Match evidence

  1. Loyalty ID exactstrong
  2. Email verifiedstrong
  3. Shared phoneweak
Duplicate profiles Fewer records for the same person
Match precision Correct merges on a labelled sample
Addressable audience More customers you can recognise and reach
Wrong-person sends Fewer messages to a merged stranger

Why change

A wrong merge is worse than no merge.

Identity is where personalisation quietly goes wrong. Too cautious and you see three strangers. Too eager and you send one person another person's booking.

One person, many records

Anonymous and known identities never meet, so history is split.

Shared details mislead

A family phone, a shared inbox or one booking for four passengers is not one person.

Merges can't be explained

Nobody can say why two records were joined, or separate them again.

Consent gets lost in the join

Permissions from one system are not carried to the unified profile.

What it does

Resolve carefully. Explain everything.

A match score is evidence, not permission to merge. Policy decides what links automatically, what goes to a person and what must never be joined.

01 Hybrid identity resolution

Trusted links first. Probabilities where they help. Policy on top.

Verified identifiers link deterministically. Messy historical records are scored probabilistically, with optional machine learning for the hard cases. Then a policy layer applies the rules a score cannot know: households, shared bookings, verified conflicts.

  • Linked, candidate, rejected and cannot-link are all first-class outcomes.
  • Person is not contact: traveller, booker, account holder, household and device are modelled separately.
  • Reversible merges, with the original source records kept intact.
Match review · candidate pairNeeds a person
AttributeRecord ARecord BEvidence
NamePriya S. MenonPriya MenonStrong
Date of birth14 Feb 199114 Feb 1991Strong
Emailpriya.m@…menon.family@…Different
Phone+91 98… 412+91 98… 412Shared?
Same bookingPNR K7Q2LMPNR K7Q2LMCaution
0.93match probability
Reviewpolicy decision
2similar cases today
Same booking + shared phone can mean two family membersIllustrative

Fictional records. A high score alone never triggers an automatic merge.

02 The unified profile

Everything you know about a customer, current to the second.

One profile with a timeline of events, computed traits and the identifiers behind it. Marketers see the story; systems read it through an API in milliseconds; both see the same thing.

  • Real-time traits: lifecycle stage, value, propensity, next likely need.
  • Zero- and first-party preferences, captured once and honoured everywhere.
  • Ask about a customer in plain language, and see which records the answer came from.
Priya Menon · profileGolden record
Goldloyalty tier
₹1.8L12-month value
Highchurn risk this month
Searched Kochi → Dubai, 2 adultsApp · 12 minutes agoNow
Called about baggage allowanceContact centre · resolved2 d
Opened fare alert, did not clickEmail5 d
Flew Kochi → Bengaluru, paid for extra legroomBooking system3 wk
4 source records · 6 identifiers · lineage on every fieldIllustrative

03 Consent and governance

Permission is part of the profile, not a separate system.

Each attribute carries its source, its purposes and its retention period. Every read is checked against them. When a customer withdraws consent or asks to be forgotten, the change reaches every downstream system.

  • Purpose-level consent by channel and by use.
  • Deletion and correction requests propagated and evidenced.
  • Data steward workbench for review queues, merge audits and quality.
Consent · Priya MenonUpdated 3 days ago
WhatsApp · service updatesGranted in app · 12 MarAllowed
WhatsApp · promotionsWithdrawn · 3 days ago · synced to 5 systemsBlocked
Email · promotionsGranted at bookingAllowed
Ad platforms · audience matchingNot askedNot allowed
Use in personalisation modelsGranted · review due in 9 monthsAllowed
Withdrawal propagated in 41 secondsIllustrative

Agents at work

Agents do the matching. Stewards make the calls.

Identity agents normalise, find candidates, score them and prepare the evidence. High-confidence, policy-eligible pairs link automatically. Everything ambiguous goes to a steward, and their decisions teach the system.

  • Evidence prepared for each review, so a steward decides in seconds, not minutes.
  • Match quality monitored, with precision and recall measured on labelled samples.
  • No raw personal data in logs, or in model monitoring.
Agent activity  Live
Nightly reconciliation of historical records1.2M records · 48k candidate pairsAutomated
39,100 high-confidence pairs linkedPolicy-eligible, evidence storedAutomated
7,300 ambiguous pairs queuedShared phone or shared bookingSteward review
Merge reversed: two siblings, one emailSteward decision, profiles restoredApproved
Auto-merge attempted on conflicting birth datesVerified conflict overrides scoreBlocked
Every step logged with inputs, reason and outcomeIllustrative

You decide what the agents may do here.

AutomatedClear-cut identity

Deterministic links on verified identifiers, trait computation and consent propagation.

Human-approvedAnything ambiguous

Probabilistic matches below the auto-link bar, unmerges, and changes to matching rules or thresholds.

RestrictedHard stops

Merging across verified conflicts, linking on weak signals alone and using data beyond its consented purpose.

What's inside

Identity, profiles and consent as one foundation.

Shared by every other capability, and available to the rest of your stack through a stable interface.

Identity graph

Deterministic identifiers with merge safeguards.

Probabilistic matching

Statistical linkage for incomplete and inconsistent records.

Normalisation

Names, emails, phones and addresses standardised, originals preserved.

Configurable match policy

Confidence thresholds, merge and unmerge rules.

Unified profile and timeline

Events, attributes and identifiers in one view.

Computed traits

Lifecycle, propensity and value, updated as events arrive.

Households and relationships

People, accounts, bookers and devices kept distinct.

Resolution explainability

Record-level evidence and a full merge audit.

Steward workbench

Review queues, golden-record governance and corrections.

Consent and purpose

Enforced on every read, withdrawal propagated downstream.

Deletion and retention

Right-to-erasure and correction with evidence.

Profile API and adapter

One contract, whether the profile lives in our store or your CDP.

How it runs

From scattered records to a trusted profile.

Each step is a named, logged stage. Nothing autonomous is switched on until its data, permissions, monitoring and rollback have been checked.

  1. Ingest source identity records
  2. Normalise, keep the originals
  3. Find candidate matches
  4. Score the evidence
  5. Apply policy: link, review or block
  6. Build the unified profile
  7. Monitor precision and freshness

How you'll measure it

Fewer profiles is easy. Correct ones are the goal.

We agree the numerator, denominator, cohort and success gate with you before anything goes live, and compare against a baseline or a holdout, never against a hunch.

  • Identity precisionCorrect merges divided by reviewed merges, on a labelled sample.
  • Candidate recallTrue matching pairs found, divided by true matching pairs that exist.
  • Duplicate profile reductionDuplicate eligible profiles after versus before, without incorrect merges.
  • Event-to-profile visibilityTime from an identity event to the profile reflecting it.
  • Consent propagationSends consistent with the latest valid consent.
  • Audit completenessAutomated and human identity decisions recorded with version and reason.

Questions

Things people ask before they start.

Deterministic or probabilistic matching: which do you use?

Both. Verified identifiers such as a loyalty number link deterministically. Probabilistic and machine-learned scoring handles incomplete historical records. A policy layer then decides what may link automatically.

What happens when the system gets a merge wrong?

It can be undone. Source records are never overwritten, every link stores its evidence, and an unmerge restores the original profiles and corrects downstream audiences.

How is this different from the Data Platform?

The Data Platform connects sources, builds the context graph and activates audiences. Customer Profiles is the identity and consent foundation underneath: deciding who is who, and what may be done with what you know.

Can it sit on top of our existing CDP?

Yes. An adapter exposes your current profiles through the same interface, so other capabilities work without a migration. You can add our identity resolution only where yours has gaps.

Composable by design

Useful alone. Sharper when connected.

Profiles are a foundation. Every capability above them gets better as identity gets cleaner.

Keeps your existing stack

Works with your booking, commerce, CRM, loyalty and contact-centre systems. Bring your own Customer 360, or use ours.

Runs where your data lives

Small language models trained on marketing data are deployed inside your cloud or data centre, so customer data and the models reasoning over it stay under your control.

Getting started

Start with two sources and a labelled sample.

Pick the two systems that disagree most about who your customers are. Agree identity and consent policy, measure precision on a labelled set, then switch on linking.

  1. Choose sources, agree identity and consent policy
  2. Run matching, measure on a labelled sample
  3. Review ambiguous pairs, tune thresholds
  4. Go live, add sources and set SLAs