The Early Warning System
All Sectors
Most firms learn about customer harm through complaint volumes — a lagging indicator that fires only after harm has occurred at scale. By the time the complaint cluster forms, the supervisory letter arrives, or the redress scheme is announced, the damage is already in the customers' accounts and the firm's reporting line. The signals were almost always present earlier — in claims patterns, journey abandonment, cancellation themes, contact-centre drivers, cohort outcome divergence — but watched in separate tools by separate teams. PRIN 2A.9 obliges firms to monitor outcomes and act; the design problem is the connecting infrastructure.
The structural move is to design an integrated signal-detection layer — drawing complaints, claims, journey, cohort, frontline, and external data into a single watchable view, with thresholds set against control limits, named action protocols when signals fire, and an audit trail that distinguishes signals seen-and-acted-on from signals seen-and-noted:
Integrated signal layer across data sourcesThe system draws on at least four independent source streams rather than relying on complaints alone: complaints and root cause output, claims or outcome data, operational journey and cohort data (abandonment, cancellation, drawdown initiation, retention exits, decline rates by employment type or vulnerability characteristic), and an external surface (supervisory communications, FOS decisions, peer enforcement, market themes). The integration is not a new monolithic platform — it is a defined data layer that draws from the existing systems on a settled cadence and presents them in a single watchable view. The FCA's December 2024 complaints/RCA review identified, as a marker of good practice, dashboards that linked complaint volumes, complaint outcomes, FOS data, quality assurance, and vulnerable-customer outcomes back to the four Duty outcomes — and identified, as a recurring weakness, MI captured but not granular enough to surface cohort-level harm. The design test: can the firm point to at least one signal in the past year that was visible in the integrated view earlier than it was visible in any single source — and can it name the source streams that converged on it?
Control-limit thresholds with named action protocolsThresholds are set against statistical control limits derived from the indicator's own historical variation, not against budget, absolute volume, or peer benchmark alone. Special-cause variation — a sustained shift outside the control limit, a run of consecutive periods on one side of the mean, an abrupt step-change — triggers an alert; common-cause variation does not. Each indicator carries an action protocol the firm has defined in advance: who owns the alert, what decision rights they hold, the escalation path if the threshold is breached for a defined period, the governance body to which the alert is routed, and the requirement to record action taken and re-test the signal post-action. Statistical process control is the methodological spine; the FCA has not prescribed it but has been explicit, through the cash savings update and the complaints/RCA review, that monitoring without acting on the data is not monitoring. The design test: for any indicator on the watchlist, can the firm produce its threshold rationale, its protocol, and the audit trail of the last alert it generated — including what action followed and whether the signal moved as a result?
Cohort-level watching, not aggregate-onlyThe signal layer disaggregates by cohort — vulnerability characteristic, target-market segment, channel, tenure, demographic, distribution route — rather than watching firm-wide aggregates that average pockets of poor outcomes against pockets of good ones. The cash savings front-book/back-book gap, the older drivers paying high GAP premiums on low-value vehicles, the long-tenure clients paying ongoing-advice charges against undelivered reviews, and the SVR mortgage cohort against widening rate spreads were all structurally observable in firms' own data — and were missed because the firm-wide number was moving the right way. Vulnerable cohort outcomes carry an explicit indicator stream, consistent with the December 2024 review's finding that complaints MI was rarely granular enough to identify outcomes for vulnerable customer groups. The cohorts watched are the cohorts named in the foreseeable harm exercise (Pattern 24) — the harms imagined at launch become the indicators watched in production. The design test: does the watchlist include cohort-specific indicators for each vulnerable population the firm has identified, and can the firm point to a case where a cohort signal fired before any aggregate signal would have done?
The signal layer integrates at least four independent data sources — complaints, claims or outcome data, operational journey or cohort data, and external surface (supervisory communications, FOS decisions, peer issues) — into a single watchable view, with disaggregation by cohort and outcome rather than firm-wide aggregate alone.
Thresholds are set against control limits derived from historical variation rather than against budget or absolute volume — and the firm can show, for any indicator, the basis for its threshold, its false-positive rate, and the cadence at which thresholds are reviewed.
Every fired signal carries a defined action protocol — named owner, decision deadline, escalation path, action taken, and post-action re-test — and the firm can produce, for any signal in the past twelve months, the audit trail of what happened next.
The signal layer connects forward into the feedback loop (Pattern 27) and upward into the annual reckoning (Pattern 25) — fired signals feed product and process change with measured impact, and the reckoning reviews how many signals fired, how many produced change, and what the firm learned.
A retail bank, reviewing its first-year Consumer Duty MI against the FCA's September 2024 cash savings update, identified that its early warning system had watched the wrong signal. Front-book rates had been published, back-book rates had been published, and complaint volumes had been within tolerance — but the gap between the two, watched as a rate over time and disaggregated by tenure cohort, had not been an indicator on any dashboard. The redesign integrated four data sources into a single watchable view: the rolling front-book/back-book rate spread by product tranche, the savings-balance migration trajectory from non-interest-bearing to fixed-term accounts, the contact-centre call-driver theme extraction, and the complaint free-text NLP output. Control limits were set against twelve months of historical variation rather than against budget. Three thresholds fired in the first quarter of operation: a tranche where the rate spread exceeded the upper control limit for four consecutive weeks; a cohort of long-tenure savers whose balance migration had stalled below population rates; and a call-driver theme on rate-change comprehension that had risen above its UCL. Each fired through a defined protocol — owner named, decision deadline set, action recorded, post-action signal re-tested. Two of the three produced product or communication changes before complaint volumes reflected the underlying issue.
A wealth manager, after the FCA's October 2024 ongoing advice review, accepted that its annual review-completion metric had been the only indicator watched on advice delivery — and that the metric had been moving in the right direction while the underlying harm was forming. The redesign treated review completion as a lagging indicator and built a signal layer around four leading ones: the rolling proportion of clients whose last documented review was more than fifteen, eighteen, and twenty-four months past due; the charge-to-review-evidence ratio by adviser cohort; the suitability-questionnaire-update trajectory; and the post-review survey theme extraction. Each was watched against control limits derived from internal historical variation, not against industry benchmark. The system was integrated with the platform's complaints data and FOS decisions feed, so an external signal touching ongoing-advice value could surface against the internal indicators automatically. In the first six months, the firm's cohort-overdue indicator fired against threshold for one adviser book — surfacing seventeen per cent of clients in that cohort with no review evidence in twenty-four months, the same proportion the FCA's industry review had named at sector level. The protocol triggered fee suspension, an outreach plan, and a structured remediation. The Pattern 25 reckoning the following year cited the early warning as the evidence that the firm's signal layer was working.
- Common failure modes
The most common failure mode is the dashboard-without-protocol: signals are aggregated into a view, the view is reviewed monthly, but no one is named to act when a threshold breaches and no decision rights are defined. The FCA's December 2024 complaints/RCA review found this pattern directly — data was captured and dashboarded, but action-taking was the consistent gap. A second is the volumes trap: indicators track absolute counts rather than rates against control limits, so growth in the customer base masks deteriorating cohort outcomes. A third is the noise-fatigue spiral: thresholds set too sensitively fire constantly, the firm desensitises, and a real signal arrives in a stream of false ones. A fourth is single-source dependence: a complaint cluster is treated as the trigger, but the cluster is itself a lagging indicator of what claims, journey, or contact-centre data already showed. A fifth is the aggregate-blindness failure: indicators watched at firm-wide level miss cohort-specific harm — the cash savings front-book/back-book gap, the older drivers on GAP, the long-tenure ongoing-advice cohort — because the firm-wide number is moving the right way. A sixth is the orphan-signal: a trigger fires, an alert is raised, no one owns the cohort it points to, and the signal is logged rather than worked.