Sentiment slide turns operational

By March, Australian households with consumer debt were reporting tighter cash positions and rising anxiety as prices and geopolitical volatility compressed budgets. The macro tape has begun to corroborate that mood: quarterly GDP growth in Q2 2025 was well below the 0.6% recorded in the December quarter and came in under market expectations of 0.4%. The signal is straightforward—demand is softening and income buffers are thinner.

Cost shocks compound the picture. In northern Australia, the Australian Government Actuary found insurers were paying about $1.40 in claims for every $1 in premiums collected in North Queensland, pushing premiums higher and eroding disposable income across affected postcodes. That pressure reverberates into arrears risk and retention for lenders and utilities. Meanwhile, discovery remains highly concentrated: the ACCC notes Google still commands nearly 94% of Australian general search, shaping how stressed customers find hardship support, refinance options and scam warnings.

Against this backdrop, a mid‑tier Australian lender (anonymised for sensitivity) moved from watchful waiting to proactive defence, building a cash‑flow stress early‑warning system and re‑designing hardship at meaningful scale.

Treat sentiment as an operating metric

The executive team reframed consumer anxiety as a measurable input to credit operations, applying a simple triad—exposure, transmission, actionability—to prioritise moves:

• Exposure: Identify cohorts most vulnerable to shock transmission, including customers in high‑premium regions, variable‑rate mortgage holders, and younger digital‑only borrowers susceptible to relationship and investment scams flagged by Scamwatch.
• Transmission: Map how macro softness and non‑discretionary cost rises convert into missed payments, policy lapses and product downgrades.
• Actionability: Focus on interventions that can be deployed in 90–120 days with measurable loss‑given‑default (LGD) and retention effects.

Two choices anchored the strategy. First, stand up responsible AI models to detect early distress, guided by the Australian Government’s AI assurance framework, which states that addressing specific AI challenges via responsible practices is a “critical enabler for AI innovation.” Second, use search and owned channels more deliberately, acknowledging the reality of a 94% search market share funnel.

Responsible AI meets frontline workflow

1) Data and model design
Engineering assembled a restricted, privacy‑by‑design feature set from transaction categorisation (e.g., rising insurance, utilities, grocery outlays), payment behaviours (grace‑period usage, partial payments), and digital signals (failed payments, session drop‑offs). Models included:

• Propensity‑to‑distress gradient boosting for 30–90 day arrears risk.
• Affordability drift classifiers to flag cost‑of‑living step‑ups (e.g., insurance premium inflation in northern regions).
• Scam exposure heuristics reflecting Scamwatch’s insights on manipulative relationship dynamics—longer, higher‑frequency transfers to new payees, late‑night spikes and narrative keywords in payment descriptions.

Models were routed through an internal AI assurance gate mirroring government guidance: traceability of features, explainability for adverse‑action contexts, bias testing by cohort and model risk documentation.

2) Channel and journey redesign
Given search concentration, the lender re‑architected SEO and paid search to land stressed users directly on a simplified hardship hub (two‑clicks‑to‑help), with structured self‑serve options: payment holidays, rate resets, and referrals to insurers for premium smoothing. For northern customers, content was localised to reflect insurance affordability programs and natural hazard risk resources.

3) Controls and governance
Model decisions were tiered by confidence, with human‑in‑the‑loop for interventions affecting credit limits or hardship outcomes. A red‑teaming function probed data drift (e.g., sudden shifts in grocery baskets), and a privacy review enforced purpose limitation. The approach aligned with the national AI assurance principles—fairness, accountability, and transparency—as enablers of safe deployment.

4) Partnerships
Insurance partners shared postcode‑level premium trajectories (aggregated and consented) to enrich affordability drift signals. Internal security integrated scam‑response flows: pausing suspicious transfers and pushing Scamwatch education prompts before high‑risk payments.

Signals, scale, and channel economics

This was a rapid stabilisation sprint rather than a growth play, and the team measured what they could prove quickly:

• Macro confirmation: Q2 GDP underperformed expectations (<0.4%), validating the decision to shift to defence.
• Cost‑pressure linkage: The North Queensland claims ratio (~$1.40 paid per $1 premium) anchored postcode‑level affordability flags and prioritisation logic.
• Channel leverage: With Google holding ~94% of Australian search, the lender concentrated SEM/SEO on hardship discovery; click‑to‑help frictions were cut from 6 steps to 2, and hardship page visitation rose materially following deployment.
• Model coverage and governance: The bank achieved portfolio‑wide monitoring with documented AI assurance artefacts, reducing decision variance and enabling auditable interventions consistent with government guidance.

While broader arrears outcomes will take quarters to crystallise, operational indicators—faster hardship routing, earlier contact, more accurate segmentation—improved in line with the stated objectives of the sprint.

Business impact and competitive advantage

Early detection is cheaper than late cure. Bringing forward customer contact and calibrating treatment based on explainable AI reduces roll rates, customer acquisition costs (by improving retention) and call‑centre load. For insurers in high‑risk geographies, feeding premium trajectories into affordability signals supports proactive payment smoothing and mitigates lapse risk. Firms that master responsible AI now can scale faster as conditions worsen, with governance already banked—a real moat when competitors are still arguing about model risk.

Technical deep dive: Why the models matter

The value is in feature engineering and assurance. Cost‑of‑living features must separate durable inflation (e.g., insurance premia) from seasonal noise. Affordability drift models emphasise change detection over absolute levels, reducing bias against low‑income but stable customers. Scam heuristics operationalise behavioural markers consistent with Scamwatch’s guidance on manipulative relationship dynamics, triggering just‑in‑time friction and education rather than blunt denials. All of this sits within an assurance scaffold—documented data lineage, testable explanations and cohort fairness checks—consistent with the Australian Government’s framework that positions responsible practice as the lever for safe innovation.

Market context and future outlook

Institutional investors gathered at State Street’s 2025 Markets Research Retreat to parse volatility—another sign that the macro debate has moved from transitory to structural. If growth remains below trend and cost shocks persist, expect three shifts: tighter affordability underwriting, increased collaboration between lenders and insurers in exposed regions, and more prominent hardship experiences in the digital storefront (surfaced via dominant search channels). Australia’s AI commercialisation gap has been noted in national ecosystem assessments; operational use‑cases like this—narrow, governed, and economically material—are the bridge from labs to P&L.

Lessons for leaders

• Treat sentiment as data: Don’t wait for arrears to spike; use macro misses and cost‑pressure telemetry as triggers for operational pivots.
• Build with assurance, not afterthoughts: The AI assurance framework is a blueprint. Bake in traceability, explainability and fairness so models survive audit and scale.
• Follow the customer’s path of least resistance: With search so concentrated, hardship discovery is a channel strategy problem as much as a risk one.
• Localise affordability intelligence: Regional cost idiosyncrasies—like North Queensland insurance premia—should directly inform treatment strategies.
• Scam resilience is part of credit risk: Behavioural signals and pre‑transaction education reduce losses and protect vulnerable customers, especially under stress.