AI operating system for your company

What it is, concretely

Each agent is a role: analyst, support, release manager, tester, marketer, quality controller, scribe, guardian. A role has a written procedure it follows, a schedule it runs on, the tools it may call, the systems it may read, and the limits it may not cross. Between runs it keeps a memory of what it learned, so the same mistake is not made twice. A director role reads the others' reports and raises what needs a person.

The runtime is Claude Code or the OpenAI Codex and Agents stack, chosen per role, running on a machine or a cloud account you own. Reports go to the channel your people already read, Microsoft Teams, Slack or Telegram, and only when something changed, failed or needs a decision. Nobody reads a log of activity; people read exceptions and results.

The rules that keep it safe

• Every action that spends money, publishes, sends a message to a customer or changes production goes through an approval gate. The gate fails closed: no answer means no.
• A daily spend cap per agent and per company, enforced in code, not in a prompt.
• Iteration caps: a task gets three attempts and then stops with a report instead of retrying until it works.
• A kill switch that stops the whole fleet from one place, and a worktree per agent so two agents never edit the same files.
• Every run writes an audit record: what it read, what it proposed, what a person approved, what changed.
• The 80 percent rule: a role is automated only where 80 percent quality is enough. Customer money, legal positions and anything irreversible stay with a person.

What our own operating system does today

• Content and publishing for four brand pages: research, drafts, a cold-read gate with two independent models, legal checks and scheduled publishing, with a person approving anything about a named company.
• Product operations for a software product ran in September: a release manager and a tester six times a day, a customer advocate twice a day, an analyst every morning, a coach once a week. That fleet has been paused since 17 September while the product is reworked; the procedures and the gates are kept and restart with it.
• Marketing reporting, lead engines and social desks for a services business, reporting to Telegram only on deviations.
• Housekeeping of the system itself: memory hygiene, procedure updates proposed from observed work, a guardian that audits the fleet weekly.
• What it does not do: it does not talk to customers about money, it does not sign, it does not deploy to production without a person, and it does not run outside the limits above.

What we can show today, and what we cannot

• We can show our own operating system live in a 30-minute call: a role running on schedule, a gate refusing an action that is not approved, the spend cap stopping a run, the kill switch, and the audit record of a real run from the same morning.
• We can hand over the written procedures, the gate code and the audit format before any conversation, so your security team reads the controls rather than our description of them.
• We cannot yet point to a customer running this operating system; the only production installation is our own, since September 2026. The first customer build will be a four-week fixed-price pilot with agreed acceptance criteria and no obligation to continue, and this page will carry an anonymised case with numbers when it is done.

How we build yours

The build takes four weeks; the SAP and assistant pilots on this site take six, because they touch systems of record. Week one: we map the routines your people repeat and pick three roles where 80 percent quality is enough and the data is reachable. Weeks two and three: we set up the runtime in your environment, the identity, the channels, the gates and the caps, and write the first procedures with the people who do the work today. Week four: the three roles run beside the people, measured against the manual baseline, and your IT team reviews the audit records. Then you decide which roles to add, and we hand over the procedures, the memory and the operating manual; from there your people change procedures by editing text, not by calling us.

The whole thing is remote and end to end: the servers or cloud accounts, the identity, the model endpoints, the agents, the reporting and the operations. You provide access and the people who know the work.

What it costs, and why it is cheap

A fixed price for the four-week build, stated in the first conversation together with what it covers, and a monthly fee for operations if you want us to keep running it. The prices are a fraction of what a large consultancy charges for the same scope, for three reasons: we work remotely, the teams are small and senior, and our own operating system does the routine parts of our own work. Model usage is billed by your own cloud or provider account.

Who this is for

Companies with 50 to 5,000 employees in Europe or North America whose people spend their days on routines: reporting, content, support triage, QA, release checks, data hygiene. If you run SAP or Microsoft 365, the agents reach them through MCP servers with the same gates. It is not for a company that wants agents with no human in the loop, and it is not a chatbot.

Questions we get asked

Is this a product or a service?

A service with code you keep. The procedures, the memory, the gates and the configuration live in your repository; the runtime is Claude Code or OpenAI tooling under your own accounts. There is no licence from us and no lock-in to us.

Which model does it run on?

Anthropic Claude through Claude Code, or OpenAI through Codex and the Agents SDK, chosen per role. Both can run under your own cloud account; on Amazon Bedrock the Claude and OpenAI models have been available side by side with EU routing since June 2026.

What happens when an agent is wrong?

It stops. Every irreversible action waits for a person, a task gets three attempts before it reports failure, and the audit record shows what was proposed. The procedure is then corrected in text, and the memory keeps the lesson.

How much of our work can this take over?

The routine part where 80 percent quality is acceptable. In our own company that is content, reporting, testing, release checks and first-line triage. Decisions about money, customers and people stay with people, by design.

Do we need developers?

Not to operate it. Procedures are written text that your people edit. You need an IT contact for access, identity and the cloud account, and we provide the engineering for the build and the operations if you want them.

Where does our data go?

To the model endpoint under your own account, in the region you choose: EU or US. The agents run in your environment; we keep no copy of your data. The data-handling map per component is part of the assessment.

Which three routines would you hand over first?

A 30-minute conversation: we map your routines, name three roles for a four-week build and send a written assessment with a fixed price. Free, whether we work together or not.

Book a 30-minute consultation

Or write to us directly: info@iteractive.ai · +421 905 452 325