An AI operations assistant connected to the entire company's knowledge
We built it for ourselves. One assistant that knows our projects, our people and our numbers, takes a message or a voice note from anywhere, and keeps working through the night.
- Client
- SimpleAI, built for ourselves
- What we did
- An assistant that runs the company's daily operations
- In use since
- March 2026
The problem
Running a small company is mostly remembering. Who said what on which call, what was promised to whom, how the ads did yesterday, what is due on Thursday. The information exists, in notes and dashboards and inboxes, and somebody has to go and fetch it every time it is needed.
AI tools did not solve that for us. Each one could answer a question well, and none of them knew anything about the company. Every conversation began with the same explanation of who we are and what we were working on, and ended with us carrying the answer to wherever it was needed. We wanted an assistant that already knows the company, can be reached from a phone in the middle of a day, and does the routine work without being asked.
9 seconds
From a short text message to the answer. 4 October 2026.
Measured
17 seconds
From a short voice note to the answer, transcription included. 4 October 2026.
Measured
8 daily jobs
Run without being asked, listed below. 5 November 2026.
Counted
What others measured
The independent evidence on AI in office work points in one direction. Researchers at Harvard Business School, working with BCG, gave 758 consultants a set of realistic tasks. Those who used AI finished over 25 percent faster, and their work was rated over 40 percent higher in quality. In 2025 the UK Government ran a trial with 20,000 civil servants and found 26 minutes saved per person per day.
26 minutes a day
Saved per person, 20,000 civil servants.
UK Government Copilot trial, 2025self-reported
How to read these numbers
These studies measured people using a general AI assistant that knew nothing about their company. We found no independent study of an assistant that works from a company's own knowledge, so we show none. The figures under "The problem" are our own measurements, and they are the only ones on this page we put our name to.
Third-party results, not ours.
What we built
We built an assistant that is reached through Telegram, the way you would message a colleague. A message can be typed, spoken or photographed, in any language. It reads what it needs, plans the job, and answers in text or, when we ask, in a voice note. For larger jobs it hands parts of the work to other agents and brings their results back together.
What makes the assistant useful is what it is connected to. It reads the whole of the company's knowledge: the projects, the people, the decisions and the notes of every day, kept in one place that a person can read too. It works with Notion, Google Drive, the Meta ads account, Instagram, Luma and email. It runs on a private server, so it keeps working when every laptop is closed.
It also keeps the day in order. The notes of each day are filed where they belong, so that what was said in passing on Tuesday is part of what the company knows on Wednesday. The numbers from Instagram and the ads account are pulled in, and every morning starts with a brief, as text and as a voice note.
- MessageTelegram · text, voice, photo
- Transcriptionvoice notes
- AssistantClaude Code on a private server
- Company knowledgeprojects, people, notes
- Notion
- Google Drive
- Meta Ads
- Instagramanalytics
- Luma · email
- Sub-agentsfor larger jobs
- Savedto the company's knowledge
- Reviewedbefore anything goes out
What it does every day
Journal and memory
Notes, decisions and action items, filed after every call.
Briefs and digests
A morning brief, an evening overview, a weekly digest.
Ads tracking
Which ads win, where money leaks, what a customer costs.
Content analysis
Every Reel against our usual numbers; the best openings kept.
LinkedIn drafts
Posts in the founder's voice, ready for his edit.
Voice in and out
Talk to it, and hear the answer back.
Background tasks
Large jobs queued and worked through.
Reminders
Set in a sentence, sent on time.
- 01
We send a message, a voice note or a photo.
- 02
The assistant reads the company's knowledge and the connected tools, and plans the job.
- 03
It answers, saves the result, and keeps what is worth remembering.
- 04
The notes are filed, the numbers are pulled, and the morning starts with a brief.
Telegram, Claude Code, Notion, Google Drive, Meta Ads, Instagram, Luma, GitHub.
How it is kept safe
An assistant that can act needs limits. It answers to one person's Telegram account and nobody else's. Every run is sandboxed, so shell commands have no network access and cannot read secrets. Guard checks stand in front of anything that spends money and in front of the files that define how the system behaves. The ads reports are read-only. Nothing goes out to another person until one of us has read it.
Outcome
At this client
SimpleAI runs on the assistant every day, and has since March 2026. The notes from every call, the morning brief, the ads report, the reminders and the drafts all pass through it. A short message is answered in about nine seconds and a voice note in about seventeen. It is how a small team keeps the state of a company in order without a back office.
It is also where our client work comes from. When a company asks us for an assistant that knows its business, this is the pattern we build: first the company's knowledge in one place an AI can read, then the assistant on top of it, connected to the tools the team already uses. The system is private; we show it live on request.
- The operations assistant on Telegram
- Its connections to the company's tools
- The morning brief and the evening overview
- The eight daily jobs
9 seconds
Short message to answer.
Measured
17 seconds
Voice note to answer.
Measured
What it is worth elsewhere
26 minutes a day
Saved per person, 20,000 civil servants.
UK Government Copilot trial, 2025self-reported