In brief
RIS d.o.o., a Croatian business-software company with 10–49 employees, worked with EDIH Adria on a digital assistant for customer support. The system receives email requests, identifies their topic and priority, retrieves information from a curated knowledge base, and proposes a response draft to an agent. It uses a retrieval-augmented generation pattern: the language model is given relevant passages from FAQs, manuals, and past interactions instead of relying on model memory alone.
The published case study estimates a 50–60% reduction in manual workload, more than 70% faster handling of standard replies, and up to 40% fewer escalations. These figures are useful as directional targets, not guaranteed outcomes. The publication does not disclose the initial ticket volume, measurement period, or calculation method. Any other company should validate the effect on its own support queue.
The problem
Small support teams face a familiar trap. Routine questions keep repeating, yet an employee still has to open each message, identify the customer and product, find the current instruction, assemble a response, and decide whether second-line expertise is required. As volume grows, specialists spend time searching for the same information instead of resolving genuinely difficult cases.
According to EDIH, RIS was receiving many routine requests by email and phone. The pilot was not designed to remove people from the workflow. Its purpose was to reduce the mechanical work: triage, knowledge retrieval, and drafting. That boundary matters. A customer receives a response only after an employee has reviewed it, while unusual or disputed cases remain with a specialist.
What the demonstrator contained
The published design included several practical elements:
- a connection to a support mailbox through IMAP;
- classification and prioritisation of incoming messages;
- a curated knowledge base containing FAQs, manuals, and historical interactions;
- retrieval of relevant passages and generation of a response draft;
- an interface where an agent can review, edit, and approve the result;
- a dashboard for monitoring the request flow.
The demonstrator was built with Python and Streamlit. This does not make that stack mandatory for production. It is convenient for testing a hypothesis, while a live service also needs access control, observability, resilience, action logging, and integration with a ticketing system or CRM.
The primary source describes the architecture as secure but does not publish a detailed threat model. It also does not say that the language model was deployed on premises. The case therefore cannot be presented as proof of a fully local deployment.
A reference design for another SME
A safe pilot for a small or medium-sized company can be organised as a chain of seven components.
1. **Inbound channel.** A separate service reads one support queue through IMAP or a ticketing API. It has no sending permission during the first stage. Each message receives a stable identifier so that retries cannot create duplicate tickets or replies. With IMAP, stable UIDs are preferable to sequence numbers that may change.
2. **Pre-processing.** The system removes signatures and quoted history, checks attachment types, identifies the customer and product, and masks unnecessary personal data. Attachments must not automatically become trusted instructions for the model.
3. **Routing.** A classifier assigns a topic, urgency, and queue. Deterministic rules remain above the model: security incidents, payments, complaints, and personal-data requests, for example, go directly to a person.
4. **Knowledge retrieval.** RAG searches only a published, versioned knowledge base with document owners and access rules. The source and version are retained for every retrieved passage. If no adequate evidence is available, the system should abstain rather than produce a confident answer.
5. **Drafting.** The model receives the message, selected passages, and a response template. The operator sees the proposed text together with its sources, confidence signals, and escalation reasons. A short, verifiable draft is more valuable than a long, polished answer unsupported by company documentation.
6. **Human decision.** The agent accepts, edits, or rejects the proposal. A separate component sends the message only after explicit approval. Retries use an idempotency key so a network failure cannot result in two identical emails to a customer.
7. **Feedback.** The system records the events needed for improvement rather than indiscriminately storing every conversation: sources retrieved, changes made by the agent, acceptance or rejection, and escalation. These signals support quality measurement and knowledge-base maintenance.
Data and integrations required
The minimum input is not the entire corporate mailbox. It is one bounded queue and the documents employees already rely on:
- 30–50 approved FAQs and instructions, each with an owner and review date;
- 300–500 anonymised or otherwise permitted support requests for evaluation;
- categories and priority rules;
- a list of cases where automated drafting is prohibited;
- read-only access to an inbound mailbox or a test copy of tickets;
- a separate test environment for knowledge, models, and logs;
- a human approval mechanism before sending.
Historical answers should not be copied into the knowledge base without review. They may contain outdated terms, one-off exceptions, personal data, or wording that was appropriate only for a particular customer. They first need to be cleaned, grouped, and converted into approved articles.
When local deployment makes sense
An on-premises model or private cloud is useful when messages contain contractual terms, customer information, system configurations, or other sensitive data. The vector store, embedding service, generative model, and operational logs can then remain within a controlled perimeter. Message bodies and internal-document fragments do not have to leave it.
Local deployment does not, however, solve three central problems by itself: outdated knowledge, unjustified model confidence, and excessive integration permissions. The system still needs access control, attachment filtering, protection against instructions embedded in inbound content, knowledge versioning, and human approval. “Runs on our server” describes placement; it is not a certificate that an answer is correct.
How to measure the economics
The percentages reported for RIS should not be copied directly into a budget. Start with a baseline: request volume, average handling time, the share of routine tickets, employee cost, and the cost of escalation. After the pilot, measure accepted outcomes rather than generated text.
Useful metrics include:
- the share of drafts accepted unchanged and after minor edits;
- average and 95th-percentile time to the first acceptable response;
- correct routing and justified escalation rates;
- errors caused by stale or irrelevant sources;
- infrastructure and maintenance cost per accepted response;
- time agents actually recover for complex cases.
Consider a simple model. If a queue receives 1,000 routine emails per month, each takes eight minutes, and the assistant saves four minutes in only 60% of cases, it frees 40 hours. The business must subtract review time, knowledge maintenance, infrastructure, and correction of errors. If very few drafts are accepted, a cheaper model will not rescue the economics.
Risks worth testing
The main risk is not that “AI replaces support,” but that a plausible-looking draft contains a quiet error. An incorrect price, deadline, or instruction can pass more quickly precisely because the writing looks polished. A pilot should therefore include conflicting document versions, questions with no answer in the knowledge base, and deliberately malicious instructions embedded in attachments.
The second risk is leakage across customers. Retrieval must apply access rules before passages are selected, not after generation. The third is automating a broken process: if categories are undefined, instructions conflict, and escalations have no owner, the model will merely accelerate the disorder.
A practical four-week pilot
Week 1: choose one repetitive request category, appoint a knowledge owner, build the evaluation set, and record baseline metrics.
Week 2: connect a test mailbox in read-only mode, configure classification and retrieval, and keep sending disabled. Verify access rules and the evidence behind each proposed answer.
Week 3: run in shadow mode. Agents work as usual while the system prepares drafts in parallel. Compare correctness, handling time, and the amount of editing required.
Week 4: expose suggestions in the operational interface for the safe category only, retaining human approval. Decide whether to expand based on acceptance rate and cost per useful result, not the number of processed messages.
Management takeaway
The RIS case demonstrates an appropriate first-step scope: one queue, a limited knowledge base, drafts rather than automatic sending, and measurable outcomes. A sensible next step for an SME is to test 300–500 real requests in shadow mode and determine how many acceptable answers the system actually saves. If the knowledge base is not ready, organising it should be the first project—not purchasing a larger model.
