How to Implement Complaints AI at Scale: Lessons from Aviva’s Tristan Harper

How to Implement Complaints AI at Scale: Lessons from Aviva’s Tristan Harper

How does one of the UK’s biggest insurance brands use generative artificial intelligence (AI) to manage complaints handling? We sat down with Aviva’s Principal Data Scientist, Tristan Harper, and asked him how his company is overcoming AI implementation challenges, the critical importance of prompt engineering, and why transparency with staff and customers matters for AI adoption.

Here are some highlights from our conversation. For the full interview, download Aptean’s new ebook: Expert Insights on Using AI to Manage Customer Complaints.

Q: How does Aviva use AI in complaints management and where do humans add value?

A: We're using AI to establish what complaints are about, create timelines and draft responses—but always with human review. AI processes the data sources to extract key information, identify patterns and flag important elements like vulnerability indicators, using keyword searches and natural language processing.

What AI doesn't do is replace human judgment. Our specialist teams still investigate complaints, make nuanced decisions and provide the human empathy that is crucial when customers are upset.

Our goal is augmentation, not automation: AI manages data and admin so people can focus on context, empathy, and resolving complex issues.

Q: What AI challenges does Aviva face managing multi-channel customer interactions?

A: The sheer variety of data sources is both our biggest challenge and opportunity. Customers contact us via phone, live chat, email and even letters. Each requires different processing like audio transcription, optical character recognition (OCR) for scanned letters, or text analysis.

We're not just dealing with volume but also complexity. A pension complaint might reference decades-old legislation, while a motor insurance complaint needs completely different expertise. That's why we maintain specialised teams for each product area, with AI augmenting their deep knowledge rather than replacing it.

The key is a unified approach that works across channels, respects the unique requirements of each product type and maintaining the human expertise that complex complaints often require.

Q: How critical is prompt engineering for developing AI applications?

A: Poor prompts lead to poor outputs meaning you miss critical information in complaints. For example: imagine a customer calls about a pension withdrawal delay, but during the conversation they also mention they need to change their address because they’re going through a divorce.

If someone just prompts AI to summarise the call as briefly as possible, it might say: customer discussed four topics including pension access and address change. That completely misses the vulnerability indicator of divorce.

But with a well-engineered prompt that specifically asks to identify customer intent, areas of dissatisfaction and vulnerability markers, and creates a timeline with bullet points, you get an output that captures everything a complaints handler needs to properly manage that customer relationship.

At Aviva, we're standardising these prompts for common scenarios. If frontline staff all created their own prompts we’d get wildly inconsistent outputs. Standardisation ensures every complaint is analysed with the same thoroughness.

Q: What will real-time AI support unlock for Aviva’s complaints handling?

A: Real-time AI in complaints management is a game-changer. By year-end, we expect AI will listen to calls as they happen, not just reviewing them afterward, opening up several exciting possibilities.

First, real-time quality assurance. Instead of sampling calls for QA after the fact, every interaction is reviewed live. The AI prompts agents about missed vulnerability indicators or suggests follow-up questions while the customer is still on the line.

Eventually, I believe we'll see traditional QA become obsolete as AI takes over in real-time. But we're taking it step by step to ensure we get it right.

We'll also see automatic logging of key information as conversations happen, reducing administrative burden on agents. If we go back to the customer getting divorced scenario, imagine AI identifying that the customer mentioned divorce and automatically flagging it for vulnerability consideration, so the agent can focus on the conversation, not note-taking.

Q: What AI advice would Aviva give to other insurers with less mature AI strategies?

A: First, don't underestimate the human element. Technology is the easy part; getting people on board is where many AI initiatives fail. Be transparent about your intentions and show staff how AI helps them do their jobs better, that it doesn’t threaten their livelihoods.

Second, invest heavily in prompt engineering and standardisation. The difference between a basic prompt and a well-engineered one is the difference between useful insights and expensive noise.

Third, learn from others but move at your own pace. Banks blazed the trail for AI in customer service, but they encountered issues like chatbots that could be manipulated into saying inappropriate things. Learn from their mistakes while building something that fits your organisation's needs and risk appetite.

Finally, remember that in regulated industries, transparency isn't just nice to have, it's mandatory. Build compliance into your AI strategy from the start, not as an afterthought. Your customers and regulators will thank you for it.

Read more from Tristan and other industry leaders in Aptean’s new ebook: Expert Insights on Using AI to Manage Customer Complaints.