Why Banks Need to Move Past Keyword Algorithms for Complaints and Vulnerability Detection

Customer conversations rarely follow a simple script. People pause, soften their voice, or mention concerns without naming them outright. These signals appear throughout daily interactions across every channel, and they ultimately shape the truth of a customer’s experience.

Banks work within this reality every day. The FCA expects firms to recognise subtle signs of dissatisfaction and vulnerability wherever they appear. The FCA has made this expectation explicit through Consumer Duty. Firms must now demonstrate how they identify and respond to customer concerns across every touchpoint. In 2026, supervisors will scrutinise the evidence firms collect throughout the customer journey, not just at isolated moments.

Traditional keyword detection tools struggle to keep up with this heightened regulatory scrutiny. They were built for predictable patterns of speech, not the varied ways customers actually express concern. Fortunately, AI solutions designed for financial services offer a way forward. To understand why, we need to examine where keyword detection breaks down.

The Keyword Detection Problem

Many contact centres still rely on systems that scan for specific words: “struggling”, “complaint”, “hardship”. While these tools are familiar and inexpensive, they’re also fundamentally limited.

Customer language doesn’t work like a checklist. People express uncertainty differently based on personality, confidence and context. A customer facing financial pressure might never use the word “struggling”. They might say “things are tight this month” or “I suppose that’s fine” when they mean the opposite.

This creates two persistent issues. Systems trigger alerts on benign conversations, creating noise that buries real concerns. Meanwhile, genuine early warning signs slip through undetected because they don’t match the keyword list. Some banks now report false positive rates so high that QA teams spend more time dismissing alerts than investigating actual problems.

Consider these common examples:

• “Things are tight this month.” The customer signals pressure. The system logs a neutral comment.
• “I suppose that’s fine.” The customer sounds uncertain. The system records agreement.
• “Can we speak another time?” The customer hesitates. The system sees a scheduling request.

At small volumes, human reviewers might catch what automated systems miss. At scale, the gaps become systematic. And under Consumer Duty, those gaps matter.

What the FCA Now Expects

The regulatory landscape shifted significantly throughout 2024 and 2025. The FCA moved from general principles to specific expectations about how firms capture and interpret customer signals.

In December 2024, the FCA’s complaints handling feedback made clear that high-level categorisation isn’t sufficient. Firms need granular data showing outcomes for different customer groups. Three months later, the vulnerable customer review reinforced that vulnerability can emerge at any point and change over time. Banks can’t rely on static flags or one-time assessments.

The 2026 Consumer Duty supervision cycle raises the bar further. Supervisors will examine how firms collect and use information across all four Duty areas, with particular attention to friction points affecting vulnerable customers. And joint FCA-ICO guidance expected early in 2026 will set new standards for identifying and supporting vulnerable customers while meeting data protection requirements.

The FCA’s message is clear: keyword systems and small sample sizes don’t provide the depth of understanding regulators now require. Banks need to see the full picture, not fragments..

How AI-Powered Detection Works Differently

AI language models follow the full arc of a conversation. They pick up on tone, phrasing and intention in ways keyword systems cannot. When a customer says “the economy is worrying”, an AI model knows that’s different from “I’m worried about my mortgage”. It distinguishes between polite deflection and genuine agreement.

Aveni’s FinLLM was built specifically for this context. It learned the language of mortgages, affordability concerns, vulnerability indicators and regulatory obligations within UK financial services. It understands the rhythm of banking conversations and the cues that signal discomfort or hesitation.

Real-world testing has validated this approach. In work with one of UK’s leading banks, Aveni’s models achieved ninety percent correlation with expert human judgement during outcome testing. The model found meaningful signals because it was trained on the language banks actually use.

Aveni Detect applies these capabilities across entire cases. Calls, documents and evidence sit together so reviewers can see the full story. Compliance teams set up assessments in plain language, and the system handles the rest across every interaction. There is no need for keyword lists or technical rule translation.

AI doesn’t remove the need for human reviewers. What it provides is coverage, consistency and context so teams can focus on the interactions that require expertise or empathy.

The Machine Line of Defence™

Traditional assurance models were designed for a time when most regulated activity was carried out manually. First line teams handled the work. Second line reviewed a small sample. Third line audited a portion of that sample. The majority of interactions received no direct oversight.

That model made sense when contact volumes were manageable and regulatory expectations were lighter. Neither is true anymore.

Contact volumes and evidence requirements have changed. Banks now need a steady understanding of what is happening across the full estate, not only the small fraction captured through sampling.

Aveni’s Machine Line of Defence™ adds continuous oversight across interactions. AI monitors conversations as they happen, identifies early signs of concern and provides structured context for reviewers. Oversight becomes a live process rather than a retrospective exercise.

Practical Impact for Banking

AI-powered detection supports daily operations in ways that feel immediate for teams.

• Clearer visibility of vulnerability: Indicators of stress, personal hardship or health concerns surface consistently, even when expressed indirectly.
• Early dissatisfaction becomes visible: Subtle cues such as tone changes or hesitation are identified before they escalate into complaints.
• Forbearance conversations can be reviewed in full: Fairness, clarity and suitability become measurable across large teams.
• Evidence becomes complete: Audio, transcripts, analysis, assessments and reviewer notes sit in one place.
• Advisers receive practical support: Real-time coaching helps them strengthen conversations while the experience is fresh.

The Path Forward

Preparation for 2026 brings a sharper emphasis on how firms identify concerns, support customers and demonstrate outcomes across entire journeys. Keyword tools can’t deliver that level of insight. They were built for a simpler time.

AI provides a practical route forward, and many banks begin with focused areas such as vulnerability detection or complaints triage. These steps build confidence while improving the quality of oversight. The result is consistent interpretation, full coverage and reliable evidence that supports customers and reviewers.

Banks that continue to rely on keyword detection will find the gap widening between what their systems can show and what regulators expect. Firms that move early towards AI-powered detection place themselves in a stronger position. They gain clearer visibility of customer needs, better protection against missed signals and a more dependable evidence base for Consumer Duty.

This shift can be introduced carefully and with full control. It strengthens day-to-day supervision and helps teams focus on the cases that matter most.