LifeOS
In daily useCustom AI agents that run one real process
We build custom AI agents for back-office work: meeting notes, the monthly finance pack, invoice intake. Each agent runs the routine end to end; money and mail still wait for a person.
From our own back office
Admin that runs itself
Of admin a month, now done by agents
Meeting notes filed minutes after the call
From call to filed meeting note
Company finance down to a few approvals
Approvals a month, for company finance
How an AI agent runs one process
Plain code decides when the agent runs and checks its work. A person signs off once.
What a custom AI agent includes
- A working agent for one process, in the client's environment
- Access where the team already works: email, Slack, Teams or a web screen
- Tests built from the client's real cases, run before launch and on every change
- An action log the client can read, plus a runbook and a handover
- Scope
- One process per agent
- Models
- Chosen per step, Claude by default
- Cost
- Build and monthly run, quoted separately
- Ownership
- What the client owns
Discover, build, deploy
- Discover
- The process is mapped with the people who run it, and the success metric is written down first.
- Build
- Progress shows on a staging copy, tested against real past cases.
- Deploy
- The agent goes live in the client's environment, then gets monitored, tuned and handed over.
Rules first: which step goes to code, a model or a person
Some steps are cheaper and safer as plain rules. The first meeting sorts every step.
| Step | Decided by | Why |
|---|---|---|
| The supplier's IBAN checksum is valid | Rule | Arithmetic gives the same answer every time. |
| The amount in words matches the figures | Rule | Checked by code, never by a model. |
| Which mail attachment is a company receipt | Model | No two suppliers send the same layout. |
| A supplier writes that last month's invoice was wrong | Model | Free text, with the meaning in the wording. |
| Paying a supplier whose IBAN just changed | Person | Irreversible, and a known fraud pattern. |
AI agents already at work
Mozar
In productionAn AI agent that edits a live catalogue, after a person confirms
AI agent services: four ways to start
When the agent is unsure, a person decides
Payments, customer emails and deletions wait for a yes. Every step is logged.
- Filed
- 1,184
- Held
- 25
Last held: Total does not match
Six guardrails, and where each one runs
- Money and mail wait for a personLifeOS finance
- The agent writes the bank's payment file and a mail draft. It has no way to pay or to send: a person imports the file and presses Send. LifeOS case study
- Admin changes wait for a yesMozar admin
- An agent can change mission types and prices only after a person sees the change and confirms it. Mozar case study
- Exact checks stay in codeLifeOS finance
- Plain code does every exact check, from bank details to totals. The model only sorts receipts.
- Names are never guessedLifeOS meeting notes
- A speaker the agent cannot confirm is flagged for a person to check, never filled in.
- Access starts read-onlyMozar admin
- The agent can look before it can change anything, and every change is shown to a person first.
- Every change is on recordMozar
- Every change lands in an audit trail that nobody can rewrite.
Prompt injection heads the OWASP Top 10 for LLM applications, and excessive agency, meaning permissions wider than the task, is on the same list. How the agents we build are secured: cybersecurity and NIS2.
Where the data goes
- Model input
- Only the text a step needs. Personal data becomes placeholders wherever the process allows it.
- Training
- Anthropic does not train models on API inputs and outputs without explicit permission. Any other provider gets the same check before use.
- Region
- The first-party Claude API processes requests in the US or globally. When data must stay in the EU, the model runs in an EU region on Amazon Bedrock or Google Vertex AI. Last checked 28 September 2026.
- Contract and access
- A GDPR data processing agreement (Article 28) is signed before real data reaches the agent, and every account and API key it needs is approved up front.
- Billing
- Model usage can run on the client's own provider account, so the bill goes straight to the client.
Custom AI agents: common questions
Difference from ChatGPT with a custom prompt
A custom agent runs one defined process with code around the model: integrations with the client's systems, rules that check every value, an approval step and a log. ChatGPT with a prompt answers one person at a time and leaves no record in the client's systems.
In LifeOS, most of the engineering sits outside the model call: the schedule, the checks and the approval points.
AI agent vs chatbot
A chatbot answers questions. An agent takes actions: it reads a document, updates a record or drafts a payment, within permissions set in advance. Many projects need both: a RAG chatbot for company documents answers, and an agent acts on what it reads.
When the agent is wrong or unsure
An agent that is unsure hands the case to a person, with the reason and the source. Irreversible steps (payments, emails to customers, deletions) always wait for approval, and every decision lands in the log.
Wrong answers found after launch become test cases, so the same mistake fails before the next release.
Data location, model training and EU hosting
Data goes only to the systems listed before the build, and the model receives only the text a step needs. Anthropic does not train models on API data without explicit permission. When data must stay in the EU, the model runs in an EU region on Amazon Bedrock or Google Vertex AI. The full list is under where the data goes.
Cost to build and cost to run
The build is quoted after the process is mapped, because the integrations drive most of the effort. The run cost is estimated per month from the expected volume and the model chosen for each step, and it appears as a separate line.
The AI workshop produces both estimates for one process.
Time to production
It depends on how many systems the agent touches and how clean the inputs are. The first release covers one process and one metric, so it reaches production before a second process starts. The workshop brief gives the estimate for a specific process.
AI models used, and switching later
Claude is the default, and each step gets the smallest model that passes its tests. The model call sits behind one interface, so switching provider means a configuration change and a rerun of the tests.
EU AI Act and the agents built here
Most back-office agents fall outside the Act's high-risk categories, which cover areas such as hiring, credit scoring and critical infrastructure. The transparency duties in Article 50 still apply when people interact with an AI system. Each project's classification is checked during scoping and written into the brief.
One process on the table: the AI workshop
Where an agent pays off, and what it costs to build and run.
- Studio
- Cluj-Napoca, Romania, EU