Lessons from Putting an AI Agent into Telecom Operations
What we learned putting an AI agent into telecom operations: approvals before any data change, the user's own permissions, counts computed in code, and every answer grounded in a tool call.

When we show an operator or a regulator what agentic AI can do with their data, the first question is almost never about accuracy. It is: where does our data go?
It is the right question. Call detail records, location data and fraud cases are some of the most sensitive data an operator holds, and in many countries they must stay in-country. A prompt to a language model is data leaving the building. That is why Prompto® supports on-premise language models, with the option to run fully air-gapped, so the model can sit next to the data inside the operator's own network.
Where the model runs is only the start, though. Building an AI agent for telecom operations taught us more about control than about models.
What the agent does
A chatbot answers a question. An agent works through a task: it decides which data it needs, calls the right tools, correlates the results and produces something a person can act on.
In Prompto, the AI Colleague answers questions about KPIs and dashboards in plain language, reviews high-scoring fraud alerts and sorts them into confirmed, suspected and insufficient evidence, runs scheduled workflows and drafts reports. It reaches the platform through tools exposed over MCP (Model Context Protocol), the same governed queries and functions a person would use.
The agent proposes, a person approves
Our rule is simple: any action that inserts or changes data needs a person's approval. When the agent wants to run a rule, save a dataset or mark a fraud alert as confirmed, it does a dry run and shows exactly what will change. Nothing happens until someone presses Approve.
Approve replays the stored call exactly as it was shown, so the model can't change its mind between the preview and the action. Only the person who asked can approve, the proposal expires after 30 minutes, and every decision is kept for audit. Deletes are not offered to the agent at all.
The agent acts as the person
It is tempting to give an AI layer one service account that can see everything. We don't. Every tool call carries the signed-in user's own credentials, and the platform's normal permission checks decide what is allowed. If an analyst can't see a dataset, neither can the agent working for them. A service account is used only for scheduled runs, where nobody is signed in.
Don't let the model count, or judge alone
Language models are poor at counting long lists, and our early versions proved it. Now every count is computed in code before the model sees the result, and the model may only use figures that came from a tool in the current answer.
Judgement needed the same treatment. When the fraud reviewer decided cases from its prompt alone, the same case could get different verdicts on different runs. We moved the core assessment into deterministic rules. The agent receives that verdict with the evidence, may only lower it, and must explain why.
Ground every number
An early version was asked whether it could forecast a revenue figure. It answered confidently without calling any tool, using an example number it had copied from a tool's description. We removed example figures from tool descriptions and added checks that catch answers not backed by a tool call. In revenue assurance, a plausible but wrong number is worse than no number.
Restrict tools, and pin versions
Each agent is a saved definition with its own instructions, model and list of allowed tools. A workflow pins the exact version of each agent it uses, so improving an agent for one workflow can't change how another behaves.
Trace everything, and test every change
Every model call, tool call, latency and error is traced, along with user feedback on each answer. A fixed set of test questions runs before any change to prompts, tools or models. That suite also guides model choice: when we upgraded, it showed a lower-cost model performing as well as a far more expensive one on our telecom scenarios.
Where to start
Don't start with a general-purpose assistant for everyone. Pick one workflow that takes up analyst time and has a clear output, such as reviewing yesterday's high-score fraud alerts or drafting a monthly reconciliation report. Keep every data change behind an approval, run it alongside the current process for a few weeks, and compare the results before widening the scope.
Our telecom AI page covers ML Studio and the AI Colleague in more detail, or get in touch to see it working on your own data.