Segurança & Confiança is an insurance brokerage headquartered in Lisbon operating across Portugal. With more than 28 (real case — data altered under NDA),000 active clients and a portfolio spread across auto, home, health, life and commercial insurance, it manages over 45,000 policies and processes around 2,400 support requests per month. The problem that led them to contact us was simple to describe and hard to solve: the customer care team, six people, was drowning in repetitive questions. 60% of the volume was questions about documentation, deadlines and coverage that could have been answered instantly — but took on average 4 hours to reach the client.
The Scenario Before: Team Drowning in Repetitive Questions
The customer care manager, Rita Silva, did a one-month ticket analysis. The results were revealing. Of the 2,412 tickets received in November 2025: 1,447 (60%) were about six recurring questions — "What documents do I need for my deductible?", "Does my policy cover X?", "When is my next renewal?", "How do I file a claim?", "How much does it cost to change my policy?", "How do I renew with a discount?". Average global response time was 4h17m. NPS stood at 34 — far from the 55 leadership considered acceptable.
Rita faced a financial dilemma. To bring response time down to acceptable levels (≤1 hour in business hours, ≤4 hours outside) she would need to hire 3 additional people — approximately €90,000/year in total costs. But 60% of those people's time would be spent answering the same things. The work itself degraded team morale: two team members had left in the last year, citing burnout.
A partial alternative — using generic "click here for FAQ" chatbots — had been tried years before and had failed. Clients rejected them because the FAQ was rigid and did not answer questions in natural language. The abandonment rate of the old chatbot was 89% — people preferred waiting 4 hours for a human response.
The Strategy: Specialised AI Chatbot, Not Generic
Phase 1: Ingesting Specific Knowledge
The difference between a useful AI chatbot and a frustrating one is specialisation. We started by ingesting into the model's knowledge base: all the general conditions of policies sold by the brokerage, the internal response scripts (with company tone), 3,000 real examples of questions and answers from previous tickets (anonymised), and the insurance technical glossary in the language a retail client uses.
Ingestion took 2 weeks and involved validation by Rita and two senior sales reps — to ensure the bot did not "invent" information nor provide answers that could create contractual issues. We also configured explicit guardrails: the bot never confirms specific coverages without human validation, never processes claims without involving the team, and never gives advice that could be considered unauthorised brokerage.
Phase 2: Integration with Internal Systems
A chatbot that only responds with generic information is limited. Our bot was integrated with the brokerage's policy management system via API, with controlled permissions. When an authenticated client asks "when does my policy expire?", the bot queries in real time and responds with the specific date, amount and renewal options.
We also integrated with the ticketing system (Zendesk) and WhatsApp Business — the brokerage already had a WhatsApp presence but without automation. Now, the same bot operates across the four channels (website, email, WhatsApp, in-app chat) with total consistency.
Phase 3: Intelligent Escalation to Human
The key to a well-designed chatbot is knowing when not to answer. We configured explicit escalation rules:
• Active claims: any mention of "accident", "claim", "collision", "theft", "damage" — immediate handover to human with context. The bot informs the client: "I'll transfer you to a colleague specialised in claims. They'll be with you in minutes."
• Complaints: detection of sustained negative sentiment (more than two conversation turns with dissatisfaction language) triggers escalation.
• High-value buying decisions: requests for new commercial insurance, patrimonial coverage over €100,000, or complex life products are immediately directed to a human sales rep.
• Three attempts without resolution: if the bot cannot resolve after 3 turns, it escalates automatically rather than insisting.
Does your team answer the same questions over and over?
AI chatbot specialised in your business, with guardrails. Resolves 70% of tickets. 30-min diagnostic.
View AI Chatbot →Phase 4: Measurement, Tuning and Iteration
During the first 6 weeks, Rita and our team reviewed 100 random conversations weekly — to identify weak responses, prompt or knowledge base improvement opportunities, and categories of questions that needed additional training. This iteration ritual was critical. A chatbot that isn't tuned degrades over time; a chatbot that is tuned improves every week.
Results After 60 Days
The metrics spoke loudly:
• Average response time: from 4h17m to 28 seconds — a 99% reduction. The 30% of tickets still requiring humans receive responses on average in 45 minutes (previously 4h).
• Tickets resolved without humans: 70% — slightly above the initial 65% target. The brokerage did not hire additional people and could redirect one team member to proactive management of VIP clients.
• NPS (Net Promoter Score): from 34 to 62 — a 28-point rise in 60 days. The improvement is greater among clients who interacted mainly with the chatbot (average NPS 68) than among those who interacted mostly with humans (NPS 57) — which is counterintuitive and important: clients prefer fast responses to "human" responses.
• Total support volume: grew 8% (clients ask more questions because they know the answer is fast), but unit cost per ticket fell 71%.
• Availability: service became 24/7 with consistent quality. 34% of conversations with the chatbot happen outside 9am–6pm — served at no incremental cost.
• Human team: now dedicated exclusively to complex cases. Internal satisfaction (measured in quarterly internal survey) rose from 3.2/5 to 4.5/5. Zero departures in the last 6 months.
Lessons for Companies with Heavy Support Load
The first lesson is that specialisation beats sophistication. A chatbot based on a generic LLM, without specific context, is a frustration. The same chatbot with 3,000 real examples of business tickets and integration with internal systems becomes a valuable collaborator. The knowledge-base curation work is what separates 30% automatic resolution (failure) from 70% automatic resolution (success).
The second lesson is that guardrails are not optional. A chatbot without clear limits on what it can and cannot do ends up giving advice the company cannot sustain, or that may have legal implications. In regulated sectors (insurance, health, finance), explicitly defining what the bot does not do is as important as defining what it does.
Conclusion
Segurança & Confiança not only solved the immediate response-time problem but unlocked a different tier of client relationship. NPS of 62 is unprecedented in the brokerage's history. The human team, freed from repetitive work, started doing proactive management — follow-up calls, annual portfolio reviews, personalised recommendations. AI did not replace the humanity of service; it freed it. This is the right way to think about AI for customer care: not as a replacement, but as the filter that lets humans do what only humans do well.
