+150 XP

Banking · Operationalizing consumer-protection and vulnerability safeguards in campaigns

The flag that arrives four hours too late

A customer rings the contact centre on Tuesday at 08:40. She says she has just lost her job and asks to pause her payment, and the agent sets a hardship flag in the servicing system. At 09:00 she receives an email offering a credit limit increase. The audience was assembled from a warehouse snapshot taken at midnight and handed to the email platform before anyone was awake.

No one wrote a bad rule. The rule existed and did not fire. Consumer protection breaks at the seams: between the system that learns something about a customer and the system that decides who gets the send. This lesson covers three ways to close those seams:

  • which targeting categories are off limits for credit
  • which flags have to survive the trip through the data pipeline
  • how to wire suppression so it executes at send time instead of living in a policy document

What has to be encoded, and where

Restricted targeting categories for credit

Since 2019, credit offers on Facebook and Instagram have had to run under a Special Ad Category. In early 2025 Meta widened that category and renamed it "Financial products and services". It now covers credit cards, loans, insurance, investment and banking products more broadly, not just credit. Inside it you lose age, gender and ZIP-code targeting. The tightest geographic radius is 15 miles, and most detailed interest and behaviour options vanish from the picker.

The history:

  • March 2019: the US Department of Housing and Urban Development brought a charge over ad tools that let advertisers exclude users by proxies for protected classes.
  • 2022: Meta settled with the Department of Justice. It retired Special Ad Audiences (the lookalike substitute it had built) and agreed to a delivery-variance system for housing ads, with credit and employment named as extensions.

Treat the platform restriction as the floor. It constrains what you can type into an ad account. It does nothing about three other channels: the audience file you upload, the list you push to your ESP, and the model that scores your own base. Your restricted-variable register has to cover those too.

The register should name the proxies as well as the protected classes:

  • ZIP or postcode at fine granularity
  • preferred language
  • device age
  • benefits receipt
  • first-name frequency tables
  • branch catchment

A retail bank can defend targeting a personal loan on estimated income and existing product holdings. It cannot defend a credit-builder card audience whose only distinguishing variable is Spanish-language app locale.

Working rule: every credit audience is built from a declared variable list. Anything outside the register needs a written business reason attached to the audience definition itself, not to a deck that gets lost.

Vulnerability flags

The FCA's guidance on vulnerable customers (FG21/1) groups the drivers into four: health, life events, resilience and capability. Since July 2023 the Consumer Duty has put this on a firmer footing. Firms must show that customers with characteristics of vulnerability get outcomes as good as everyone else's. That makes a leaked credit email to a customer in hardship a reportable outcome failure, not just a lapse in good practice.

The FCA's Financial Lives research puts roughly half of UK adults in scope of at least one characteristic at any time. That tells you the operational shape of the problem. Vulnerability is a state that switches on and off, not a segment you build once.

In the US there is no equivalent single duty. The abusive prong of the UDAAP standard in the Consumer Financial Protection Act reaches the same conduct: taking unreasonable advantage of someone's inability to protect their own interests. The CFPB has cut back supervision and withdrawn much of its guidance since 2025, but the statute is unchanged. State attorneys general can enforce it as well. The examination procedures are here: CFPB on UDAAP.

Four sources feed the flag, and they arrive at different speeds:

  • Servicing systems: hardship arrangements, forbearance, bereavement notification, collections status.
  • Customer-set product controls, such as the gambling block HSBC and other UK banks expose in the mobile app. A customer who has switched that on has told you something, and a limit-increase or overdraft campaign should read it.
  • Behavioural signals from the account: persistent overdraft use, repeated failed direct debits, minimum-payment-only cycles.
  • Third-party instruments: power of attorney, debt-management plan notification, court orders.

One design decision matters more than the taxonomy. Marketing systems should receive a boolean plus an effective date, never the reason. "Suppress from credit promotions, set 2026-03-04" is enough to run a campaign. Pushing "bereavement" or "gambling block" into a hashed audience upload sends special-category-adjacent data to an ad platform for no operational gain. Suppress locally, then send the residue.

Affordability flags

Affordability is the flag most often computed and least often connected. Banks model it for underwriting, then market from a separate list built by a growth team on different infrastructure. If the modelled figure does not reach the campaign layer, the suppression cannot happen.

The fix is unglamorous:

  • the affordability score gets a home in the same customer table the audience builder queries
  • it carries a refresh date
  • audiences that ignore it are rejected at build

Rules that fire, not rules that advise

Suppression runs last, and fails closed. Sequence matters:

  1. Build the audience.
  2. Apply exclusions.
  3. Re-check flags in the minutes before dispatch, not at build time.

If the flag service times out, the send does not go. Fail-open suppression is the default in most marketing stacks, and it is the wrong default here.

One identity key, agreed in advance. A hardship flag keyed to a customer ID will not remove a record from an email list keyed to an address. Neither will remove the hashed phone number sitting in a custom audience on a social platform. Most suppression leaks are join failures, not policy failures. Pick the key, resolve everything to it, and test the resolution rate rather than assuming it.

Freshness windows with teeth. Decide the maximum age of the data behind any credit audience. 24 hours is a reasonable ceiling; weekly snapshots are not. Then make the build fail when the source table is staler than that. Batch audience syncs in a CDP such as Segment are measured in hours, so a nightly cadence plus a queued send can put a full day between the flag and the inbox.

Reinstatement is a decision, not a timeout. Flags that never expire quietly shrink your addressable base and push teams to work around them. Flags that expire automatically re-target people mid-difficulty. Write the rule down per flag family:

Flag familyWhen suppression lifts
Collections statusAccount cure plus a cooling-off period
BereavementUsually a fixed window, then manual release
Customer-set gambling blockOnly when the customer clears it

No manufactured urgency on credit. Countdown timers and "2 left" mechanics push someone past the point where they weigh a borrowing decision. That is exactly the conduct the abusive standard describes, and the Consumer Duty treats it as foreseeable harm. Genuine, factual deadlines are fine and should be stated as such. How the rate and the material terms are then presented is covered in the fee-tables lesson.

A worked suppression rule: the affordability screen

You are promoting a card with a $10,000 limit. You need a filter that removes people the offer would predictably strain. Debt-to-income (monthly debt payments over gross monthly income) is the usual guardrail. What follows is a marketing suppression rule, not a lending decision.

Assume minimum monthly payment ~ 3% of a used balance.
Prospect at modelled DTI 45%, plausible $200/month on the new card
  -> post-offer DTI approaches the ~50% zone many lenders treat as high risk.

Rule set (evaluated in order, first match wins):
  1. hardship_flag OR collections_status  -> suppress, no override
  2. gambling_block_on                    -> suppress from credit + overdraft
  3. modelled_post_offer_DTI > 50%        -> suppress
  4. affordability_score_age > 30 days    -> suppress (stale, cannot assert)
  5. else                                 -> eligible

Log the rule number that removed each record. Counts by rule are the
artefact you can show later.

Rule 4 is the one teams argue about and the one that saves them. Missing data is not clean data. If you cannot assert affordability, you have not screened for it.

Wiring it so it cannot be skipped

Policy that depends on someone remembering is not a control. Four checks make the pipeline testable:

  1. Seed records. Maintain a small set of synthetic customers carrying each flag type, and inject them into every credit audience. If a seed reaches the seed inbox, the suppression layer is broken and the send is quarantined.
  2. Reconciliation. Compare three numbers: records built, records suppressed by rule, and records dispatched. They should tie. A gap of even 0.3% on a 400,000-record file is over a thousand people you cannot account for.
  3. Identity-resolution coverage. Report the share of the audience where the suppression key resolved. Anything unresolved is treated as suppressed, not as eligible.
  4. Partner reach-back. Suppression files pushed to affiliates, aggregators and ad platforms need a confirmed timestamp, because a list you sent three weeks ago is still being mailed.

The pre-launch checklist lesson specifies what gets retained and who signs it off. The point here is narrower: the counts have to exist before anyone can sign anything.

Knowledge check

1. A marketing team builds a mortgage campaign audience map that, while not explicitly referencing race, ends up excluding neighborhoods that are predominantly composed of minority residents. What fair-lending concept does this most directly illustrate?

2. Why does the lesson emphasize that 'your audience-selection logic is a fair-lending decision'?

3. The lesson notes that much of the harm in the Wells Fargo settlement traced back to how products were 'sold and communicated, not just how they were priced.' What broader principle does this illustrate for marketers?

4. A bank shows a high-cost credit-builder card to one ZIP code cluster and a premium rewards card to another, where the split correlates with race. Which practice does this best exemplify?

MULTIPLE CHOICE

5. Select ALL correct answers about how fair-lending law (ECOA and FHA) applies to marketing campaigns.

Select all the correct answers.

MULTIPLE CHOICE

6. Select ALL correct answers about the consumer-protection duties that shape banking campaigns.

Select all the correct answers.

Where teams actually get caught

The lookalike loop. Feed a seed list of "good customers" into a lookalike model and it reproduces whatever skew the existing base has, then amplifies it across acquisition. Meta's special ad category and the retirement of Special Ad Audiences removed part of this on one platform. The same mechanic survives in every in-house propensity model and in most programmatic setups.

For credit:

  • prefer broad reach with hard affordability suppression over narrow amplification
  • audit the seed list composition before the model ever runs

Reactivation of dormant distress. A "we miss you" campaign to lapsed customers picks up people who lapsed because they got into trouble. The suppression flag often sits in the collections system, which is frequently the one database the reactivation query never touches.

The adtech and affiliate blind spot. Partners writing their own copy ("guaranteed approval") create exposure that lands on you.

  • US: In August 2022 the CFPB issued an interpretive rule treating digital marketing providers as service providers. It withdrew that rule in May 2025, but the withdrawal does not shield the bank. Misleading partner copy remains a deceptive practice under the statute, and state regulators are increasingly active.
  • UK: The Consumer Duty and the financial promotions regime keep responsibility for approved promotions with the firm.

The controls are the same in both markets. Contract for creative control, receive the suppression confirmation, and sample live placements rather than approved mockups.

Key takeaways

  • Suppression that runs only at build time is decoration. Re-check flags minutes before dispatch, and fail closed when the check does not return.
  • Meta's "Financial products and services" special ad category removes age, gender and tight geography. Your own lists and models need an equivalent restricted-variable register.
  • Vulnerability flags should travel as a boolean plus an effective date. Suppress locally and never ship the reason to an ad platform. In the UK, the Consumer Duty makes leaks an outcome failure.
  • Most leaks are identity-resolution failures. Agree one key, measure the resolve rate, and treat unresolved records as suppressed.
  • Missing affordability data is a suppression trigger, not a pass. Log which rule removed each record so the counts exist when someone asks.