In an insurer’s contact center the customer waits on the line while the agent digs through policy terms, tariffs, procedures and several systems. The answer is somewhere, it is just hard to find fast and be sure it is current. For a head of customer service that means one thing: the knowledge exists, but it does not reach the conversation on time. That gap is what an AI assistant in insurance, built on RAG combined with a language model, closes.
In our experience, the hard part is not standing up a model but embedding it in the real service process and in the company’s sources of truth. For UNIQA, one of the largest insurers in Poland, we built an AI assistant for the contact center on the PLLuM model that suggests ready answers to agents and rewrites records into benefit-led language, running on-prem.
Why do agents waste time searching for answers?
The problem rarely sits with people. Knowledge in insurance is scattered, vast and changing: policy terms differ across products and years, tariffs get updated, and procedures live in separate documents. Instead of talking to the customer, the agent jumps between tabs and assembles an answer from several places at once.
The consequences are familiar to anyone who runs a service operation: handling time grows, the risk of inconsistent answers rises, and new hires take long to get up to speed. The same case can end differently depending on who handled it. This is not a motivation problem but one of access to knowledge at the moment of the conversation.
What is an AI assistant (RAG + LLM) in insurance?
It is a solution that combines retrieval from the company’s own documents (RAG) with answer generation by a language model (LLM). An LLM on its own writes well, but it does not know your policy terms and can make things up. RAG fixes that by giving the model specific fragments of real documents to base its answer on.
In practice, an agent’s question (or the content of a customer’s case) triggers a search for the most relevant fragments across the base of policy terms, tariffs and procedures, and the model phrases an answer grounded in those sources. So the agent gets not generic talk but an answer backed by a document you can point to. For UNIQA we built this on the PLLuM model, suited to the Polish language of insurance service.
How does retrieval from policy terms and tariffs, grounded in sources, work?
For the assistant to be trustworthy, what matters is the quality of the retrieval layer, not the model alone. The documents have to be prepared so the model reaches the right fragment:
- splitting documents into meaningful fragments (policy terms, tariffs, procedures) that preserve context,
- vector indexing, which finds content by meaning, not by keyword,
- grounding answers in sources, linking generated content to a specific document,
- updating the knowledge base when tariffs or terms change, without rebuilding everything.
As a result the assistant answers from current documentation, and the team knows where each answer comes from. This layer decides whether the solution is a demo or a production tool.
How do you connect the assistant to CRM and service channels (omnichannel)?
The assistant only gains value when it lives where the agent works, not in a separate window. So we design it as part of the existing service environment:
- an AI Suggestions Panel with ready answers directly in the agent’s interface,
- automatic summaries of records in benefit-led language the customer understands,
- integration with CRM and the call center, so the assistant sees the case context,
- a single “source of truth” in an omnichannel model, consistent regardless of the contact channel.
On the UNIQA project we wired the assistant into the CRM and the call center system so it became a shared point of reference across channels. The agent no longer hunts for answers in several places but receives one in the context of the case.
How do you handle data security and environment (on-prem / private cloud)?
In insurance the data is sensitive and the regulatory bar is high, so the architecture has to account for that from day one. We run the assistant where the company’s data stays under its control: on-premises or in a private cloud, without sending case content to public model APIs.
This gives the organisation two things at once: customer data never leaves the controlled environment, which simplifies conversations with security and compliance, and the company keeps control over the model and the knowledge base. For UNIQA we ran the assistant precisely in an on-prem / private-cloud model, which was a condition for going to production.
Why is the human in the loop essential?
Because the goal is not to replace the agent but to free them from tedious searching. The assistant drafts a proposed answer and summarises the case, but the human decides what to send to the customer, combining the speed of the machine with the judgment and empathy of the person.
For a head of service it also has a practical side: the human in the loop guards against error and builds the team’s trust in the tool. Agents adopt it faster when they see it supports their decisions rather than displacing them, and the knowledge of the best agents becomes available to the whole team instead of staying in a few heads.
How do you measure the effect (AHT, consistency)?
Deploying an AI assistant is a business project, so we measure it with numbers. The main directions we watch together with the service team:
- AHT (average handling time) as the primary efficiency metric,
- cost per handled case, which drops as agents close cases faster,
- answer consistency, meaning the same, correct answer regardless of the agent,
- time to onboard a new hire, because knowledge is available from day one.
The goal of the UNIQA deployment was to reduce AHT and the cost of service while keeping quality, which is why we set the metrics at the start rather than bolt them on at the end. That is the easiest way to show the return on investment to decision-makers.
Which skills do we provide for an AI project in insurance?
A deployment like this needs a team that understands the model, the service operation and the integrations at once. We shape it around the specific scope, not “just in case”:
- LLM/RAG engineers responsible for the model, the retrieval layer and answer quality,
- MLOps specialists keeping the solution in production (deployments, monitoring, updates),
- integration engineers wiring the assistant into CRM, the call center and the document base,
- data and security architects looking after the on-prem / private-cloud environment.
Our AI/Data competence is real, and we deliver GenAI, RAG and LLM projects in production, not just as demos. That is the difference between a showcase and a solution that supports real customer conversations every day.
How do we do it at RITS?
We do it as a partner that takes AI all the way to production, not one that stops at a proof of concept. Our project for one of the largest insurers in Poland shows it:
- for UNIQA we built an AI assistant for the contact center on the PLLuM model (RAG + LLM), with an AI Suggestions Panel in the agent’s interface,
- we added automatic summaries of records in benefit-led language and tied it all into an omnichannel “source of truth” integrated with CRM and the call center,
- we ran the solution on-prem / in a private cloud, so customer data stayed under the company’s control,
- we moved from a pilot in 8 weeks to the scale of the whole organisation, aiming to reduce AHT and the cost of service.
Most often this is an AI deployment that supports existing service processes and, where skills need topping up fast, we add our team in a chosen engagement model.
Where to start with deploying an AI assistant?
Start from one concrete service scenario (for example the most common questions about policy scope) and from the question of which documents must feed the knowledge base. That defines the pilot scope, the environment requirements and the metrics that will show the effect, before you invest in scale. We covered the differences between engagement models in our article on IT engagement models.
Want to deploy an AI assistant (RAG + LLM) in your contact center? Let’s talk - we’ll match the pilot scope and the architecture to your organisation.