Cases/AI agent development/European telecom
Telecom2026AI agent development
European telecomNDA

AI Customer Support Agent for a European Telecom

A European telecom operator decided to hand a share of customer support to AI. We built the agent on Microsoft Copilot Studio: a working proof of concept in 6 weeks, production-ready in 2 months.

European telecom
6 weeks
to a working prototype
2 months
to production readiness
80%
of support requests covered by the agent's topics
  • Telecom
  • AI agent development
  • 2026
The challenge

Their Business Challenge

The operator came to ProCoders with a support problem that grew with every new customer: the more people signed up, the more requests landed on the support team. In Europe, adding support staff is slow and expensive: employer taxes, strict labour law and long hiring cycles make every new hire a long-term cost.

Read more

The operator's priority was to make the business more profitable. Covering the extra work with new hires would make support costs grow at the same pace as the customer base.

The operator set two goals for the AI Customer Support Agent:

  1. 01

    Answer typical customer questions inside the customer portal, around the clock, using the company's own support knowledge.

  2. 02

    Meet strict security requirements: the agent is open to signed-in customers only, even a direct chat link leads to the login screen, and conversations are logged without personal data.

The hard part sits behind those two goals. In our experience, connecting an LLM to a knowledge base takes 10% of the effort. The other 90% goes into retrieval tuning, security, edge cases and rounds of testing that make an AI agent production-ready.

How we worked

Assembling the Team

The operator came to ProCoders on the strength of our track record with other companies in its home country. On the client side, the project was led by the operator's IT director.

ProCoders assembled a focused team of four:

  1. 01

    Solution Architect

  2. 02

    Full-Stack Developer

  3. 03

    QA Automation Engineer

  4. 04

    Senior Project Manager

Six weeks from kickoff to a working proof of concept

The team built the agent entirely on the native Microsoft stack, inside the operator's own Microsoft environment in Azure. That meant almost no custom backend code. The real expertise was in knowing how Copilot Studio infrastructure works, how retrieval-augmented generation (RAG) behaves on real support content, and how to test an AI agent until it is safe to put in front of customers.

Read more

After the proof of concept, we ran load testing, confirmed the agent performs under real usage patterns and prepared it for launch. From kickoff to production readiness, the project took 2 months.

Architecture of the AI support agent: customer portal, sign-in, Copilot Studio agent, support wiki index and logging
Everything runs on the native Microsoft stack inside the operator's own Azure environment.
Test run of one answer: message received, input check passed, support wiki search, source page read, answer sent with a link
One answer in the agent's test run: the input check, the wiki search, the source page and the reply.

An Agent That Is Ready for Production

ProCoders delivered the AI Customer Support Agent with the following components:

  1. 01

    Chat inside the customer portal

    The agent lives inside the operator's existing customer portal, so customers need no extra app. Only signed-in customers can open it: Microsoft Entra External ID checks every session, so even with a direct link an outsider lands on the login screen.

  2. 02

    Single sign-on passed to the agent

    The customer's portal session is passed to the agent through Microsoft Entra External ID. The sign-in stays inside Microsoft infrastructure, and customers never log in twice.

  3. 03

    Agent built in Copilot Studio

    We configured the conversation logic, topics and guardrails in Microsoft Copilot Studio. We compared several language models on the operator's real questions, and OpenAI GPT delivered the highest answer quality.

  4. 04

    Answers from the internal wiki

    The agent uses RAG over the operator's internal support wiki. Answers come only from approved company content, general model knowledge is switched off, and each answer points to the right page of the portal. When the wiki has no answer, the agent says so instead of guessing.

  5. 05

    Automatic and manual knowledge updates

    A scheduled sync copies the wiki into the agent's knowledge base every night, and the support team can start an update manually at any time. They keep editing the wiki as usual, with no code involved.

  6. 06

    Logging without personal data

    Usage, errors and answer behaviour are tracked in Azure Application Insights. Logging is configured so that customers' personal data stays out of the logs.

  7. 07

    Protection from prompt injection and hallucinations

    Attempts to override the agent's rules or extract its prompt are stopped before the message reaches the model, with a fixed reply. The agent does not claim access to contracts, bills or other customer records it cannot see.

  8. 08

    Testing at scale

    Rounds of functional and security testing covered prompt injection, hallucinations, requests for other customers' data, typos, slang and very short questions. Load testing ran 4,000 requests across 5 to 10 parallel sessions. The agent leaked neither its prompt nor its internal sources.

  9. 09

    Documentation and handover

    We delivered architecture and integration documentation and handed the agent over in full. The operator owns it, and its team can change the agent's content and behaviour on their own.

Together, these components turn a language model into a support channel the operator can trust with real customers.

Portal chat: a customer asks how to reset the password, the agent replies with three steps and a link to the portal page
A typical answer: steps from the support wiki and a link to the right portal page. Redrawn from test chats, client branding removed.
Portal chat: the agent says it has no information on a question outside the wiki, then gives a fixed reply to a prompt injection attempt
No answer in the wiki, no guess: the agent says so and points to customer service. Attempts to override its rules get a fixed reply.
The outcome

80% of Support Requests Covered by the Agent

The AI Customer Support Agent covers the topics behind 80% of the operator's support requests. A chat with the agent costs a fraction of a request handled by a support employee.

Read more

The operator first asked for the agent to read its wiki directly, so the first release passed the whole wiki to the model with every question, at about 10 cents per answer. We had recommended retrieval from the start. In the next iterations we moved the agent to a search index, and an answer now costs about 1 cent.

Customers get answers inside the portal at any hour, and the support team spends its time on the cases that need a person. The operator can keep growing the customer base without support costs growing at the same pace.

The ProCoders Difference

  1. 01

    A working proof of concept in 6 weeks and a production-ready agent in 2 months.

  2. 02

    Built on the native Microsoft stack inside the operator's own Microsoft environment in Azure, with no custom backend to maintain.

  3. 03

    About 1 cent per answer after the move to retrieval, down from about 10 cents in the first release.

  4. 04

    Security from day one: sign-in-only access, no personal data in logs and zero prompt leaks across 4,000 test requests.

  5. 05

    Full ownership on the client side: the operator's team updates the knowledge base and the agent without writing code.

Your story next

Tell us the outcome.
We will write your case study

Discovery in two weeks. ROI proposal after. Public case study optional, by your choice.