AI agents for a mobile operator
- Scope
- 8 services
- reading four sources between them
Task: take the routine off the teams — eight jobs somebody did by hand every week, from watching the competition to writing up the meeting, each handed to an agent of its own.
Results
- 8
- agent services in use
- −70%
- less time on research and fact-checking
- 24/7
- monitoring with nobody watching
Nothing waits for the planning meeting
A competitor’s move arrives as an alert the same day, and a meeting is written up before the next one starts
One failure is one service
The eight run in separate containers, so an update or an outage in one leaves the other seven working
How it works
A daily crawl of their sites, alerts to Telegram
Somebody used to open a dozen sites once a week and try to remember what had moved since the last time. The agent walks them every day and sets each page against the copy it saw before, so what it reports is a change and not a page. It goes to Telegram the day it happens, which is the whole of the point: a competitor’s new tariff found on Friday and raised at Monday’s planning meeting is three days of a head start handed back.
Tasks into Jira, out of the shared base
A decision taken in a meeting becomes a task only if somebody writes it down, and the person who should is usually the one doing the talking. The agent takes what was agreed, writes the task into Jira with the context it needs, and ties it back to the record it came from, so the task carries its own reason with it. It reads the same base the reports and the legal review read — one place where what the company knows about a client lives, rather than three.
xlsx, pdf and docx out of the same base
The figures in a report are always the same figures; what differs is who is asking and in what shape. The agent assembles them out of the same base on request — a spreadsheet for the analyst, a PDF for the client, a document for the internal review — so nobody is exporting a table and reformatting it by hand for the third time this week. Because the source is one, two reports made a day apart cannot disagree with each other.
An independent LLM reviewer before signature
The risk in a contract is rarely in the clause nobody read; it is in the one read quickly by the person who wrote it. The agent reads the document before it is signed as a separate reader: it had no part in drafting it, and it checks against the company’s own rules rather than against what the draft was trying to say. It signs nothing and it is not a lawyer — it is the second pair of eyes there was never time for.
Written up from the recording of the meeting
Minutes written after a meeting are written from memory, and the memory belongs to whoever had the least to do during it. The agent works from the recording: who said what, what was decided, what was left open and with whom. The brief it produces is the one the presentation agent reads afterwards, so the deck and the minutes cannot turn out to be about two different meetings.
PPTX from the brief of that same meeting
Making a deck out of a meeting is an evening’s work, and most of that evening goes on retelling what has already been said once. The agent builds the PPTX from the brief the minutes agent wrote — that brief, not a retelling of it — so the deck says what the meeting decided and the figures in it come from the same base as the reports. What is left to a person is the part a person should be doing: the argument, and the order the slides make it in.
A live assistant inside the call itself
The help you need in a negotiation is needed during it, not in the write-up afterwards. The agent listens to the call as it runs and keeps a map of it in front of the person on the line: what has been agreed, what is still open, and what the other side said earlier that sits badly with what they are saying now. It prompts nobody and says nothing out loud — it is a surface to glance at, so the preparation does not have to be carried in somebody’s head.
Sub-agents over an LLM and a vector base
A research question is rarely one question: it is a dozen, and the answers have to be set against each other. The agent splits the request between sub-agents — as many as that request needs — each working its own part over the company’s own vector base and the model, and assembles what they bring back together with where each piece came from, so an answer can be checked rather than believed. Most of the seventy per cent off research and fact-checking is here.
Stack
LangChain Deep Agents · FastAPI · Docker · Telegram Bot API · Jira