Main content

Case study

LifeOS: the studio's back office, run by AI agents

LifeOS is the AI agent operating system behind our own back office. Meetings file themselves, finance arrives as drafts, and a person approves only money, mail and names.

Client
ELASTO, own system
Sector
Back-office operations
Role
Agents, skills, schedules
Status
In daily use
  • ≈12h

    Of admin a month, now done by agents

    Estimate: measured runs, assumed minutes

  • 7min

    From call transcript to a filed note

    Median of 32 recorded meetings

  • 149of 151 days

    Logged since May, with no typing

    One log file per day

  • 3approvals

    A month, for company finance

    By design; first full cycle in October

FIG. 1The pattern behind every process
The LifeOS pattern: a trigger (a finished call, a new mail, the clock) meets a quick check with no AI; only new work starts an agent; code checks every figure; the result lands as a note, a draft or a reminder; a person approves money, mail and names.Six steps from left to right. The first five run on their own. A check with no AI runs first and stops when nothing is new. The agent does the judgement work, deterministic code does anything that must be exact, and the output lands where the work already lives. A person approves at one marked point only.TriggerCall, mail, clockQuick checkNo AIAgentClaude Code skillCode checksFigures, totalsNote or draftWhere work livesA person approvesMoney, mail, namesAutomaticOnly here

Processes

Every process, one pattern

Each took over admin that was done by hand, or not done at all.

  • Meeting notes

    The write-up after each call, when it happened at all, now files itself; a person only confirms speaker names.

    FIG. 2Call to filed note
    Meeting notes: a finished call transcript starts an agent that writes the summary and action items and files the note under its project; a person confirms the speaker names.Call endsTranscript savedAgent writesSummary, actionsFiled noteUnder its projectNames confirmedBy a person
    Notes filed this monthDemo data

    3notes

    Each one under its project

  • Company finance

    Statements, receipts and tax payments were typed by hand; now they arrive checked, and only a person pays or sends.

    FIG. 3Mail to payment file
    Company finance: a daily check with no AI finds new mail; an agent and deterministic parsers check every figure and prepare the payment file and a mail draft; only a person imports the payments in the bank and presses Send.New mailDaily check, no AIAgent, parsersFigures checkedPayment fileAnd a mail draftBank and SendA person only
    Finance documents checkedDemo data
    Passed
    23
    Held for a person
    3

    Last held: IBAN changed

  • Daily record and weekly review

    The day was rarely written down; now agents log each night, brief each morning and review each Sunday.

    FIG. 4Clock to daily record
    Daily record: on a schedule an agent reads the day's captures and the code repositories and writes the daily log, the morning brief and the weekly review; nothing is sent out, and a person corrects the record.The clockNightly, weeklyAgent readsNotes and codeDaily logBrief, reviewNothing sent outRead, corrected
  • Capture from anywhere

    Ideas used to scatter across apps; now a phone note or chat message lands in the right note, linked.

    FIG. 5Message to the right note
    Capture: a phone note, a chat message or a command from a code repository reaches an agent that files it into the right note with links, then commits and syncs; a privacy filter guards the chat side.Phone or chatOr a laptopAgent filesLinked, dedupedRight noteCommitted, syncedPrivacy filterOn the chat side

Services

The AI practice behind LifeOS

How LifeOS is built: agents, schedules and code checks

Quick checks first
Plain schedules with no AI decide when an agent runs, so a model is called only when there is new work.
Agents run skills
Claude Code runs custom skills for meeting notes, the daily record, capture and finance.
Exact work stays in code
Anything that must be exact, such as money and identity checks, runs as tested code.
No send path, no payment path
Mail stays a draft and payments stay a file. A person approves both, so the system cannot move money.
A second model on every pull request
A second AI model reviews every code change before a person merges it.

Stack

AI agents
Claude Code, Codex
Code
Python, TypeScript
Knowledge
Markdown notes, Obsidian

Next case

A system like this starts with a brief

Studio
Cluj-Napoca, Romania, EU