An AI workforce for mid-sized companies

Hire your first AI employee.

With a personnel file, a probation period and a boss who stays human. AI colleagues that work where your team works — and ask permission for anything risky.

Personnel file v4
Sabine K.
Invoice review · Finance
SupervisorM. Weber
StatusIn production
Monthly budget62 % · 240 €
Human zones3
Approval required
Sabine wants to release invoice RE-4471 for €8,420 for payment.
Approve Decline
Escalates to B. Krause in 3 h 12 min
Assessment centre
Safety 94
Honesty 89
Brand voice 92
Compliance 97
Process fidelity 86

An AI colleague is not a chatbot. It has a role, a supervisor, a budget and limits — and it follows your processes. Not configured. Hired.

21
building blocks for an AI workforce
5
assessment dimensions before day one
100 %
of actions in the audit trail
EU
operations in the EU — internal tools stay in your network
The hiring process

From job description to productive colleague.

No prompt engineering. You describe a role — the platform runs it through assessment, probation and sign-off before anything touches production.

Hiring and leading in detail
01
Write the personnel file

Role, persona, thinking tools, no-gos, model, skills, budget. Every change versioned and revertible.

02
Assessment centre

Adversarial testing across five dimensions. An independent judge model scores each one with reasoning.

03
Probation

Sandbox operation with tight feedback loops. Rehearse real processes safely — effective actions run as simulation.

04
Production — supervised

Approval rules, escalation deadlines, budget caps. Trust grows in steps, never all at once.

Human oversight

Control is an org structure.

Not a checkbox in a settings dialog. Every AI colleague inherits a responsible supervisor through the hierarchy — with deputies, deadlines and a safe default at the end of the chain.

Control and safety in detail
Approvals where the human already is
In the studio or as an interactive card in Microsoft Teams — mid-way through a delegated task and in overnight runs too. No new tool, no extra inbox.
Automatic escalation
If a request sits unanswered it moves up a level after 4, 8 or 24 hours, depending on risk class.
A safe default at the end
Critical and normal actions count as declined at the end of the chain. Only explicitly low-risk ones go through once the clock runs out — so small work does not hang on someone's holiday.
Planning mode

The plan comes first, then the work.

Larger jobs do not start with work but with a card: goal, waves, open questions — and explicitly the non-goals. You approve, return or reject.

Planning mode in detail
Waves, not a chain
What can run in parallel does. What needs another task's result waits — and the platform does not allow otherwise.
Non-goals are written down
The place where a quick read tells you whether the mandate was understood. Cheaper than any correction afterwards.
Version 2 says what changed
A revision must name and justify every change, otherwise the platform does not accept it. Nobody compares two long plans by hand.
Work plan · version 2
awaiting approval
Spring trade fair
Goal The booth is live on 12 March and eight confirmed customer meetings are in the calendar.
Wave 1
1
Agree on concept and key message
Result: approved concept
2
Collect booth space quotes
Result: three quotes compared
Wave 2
3
Ad copy based on the concept
Result: copy ready for sign-off
Non-goals
No new corporate design No commitments on discounts
Open question

Hall 9 like last year — or hall 4, closer to the entrance and 15 % pricier?

Hall 9 Hall 4
Approve Return with a comment Reject

A plan is always decided by a human — even when a team lead had it drafted.

In daily work

They work where your team works.

Nobody has to learn a new program. AI colleagues live in the internal chat, in Microsoft Teams and in the shared mailbox — reachable by @mention, replying in-thread.

Everyday work in detail
Chat, Teams, email

@mention in the channel, centrally managed mailboxes polled every minute, replies bundled and only when there is substance. Only someone explicitly assigned to a mailbox may send; meetings and calendar entries are part of it.

MW
@Sabine could you check yesterday's incoming invoices?
Yes. 14 invoices, 12 clean. Two deviate from the purchase order — I'll put them in front of you as an approval.
Helga operates the platform for you

The built-in assistant sets things up, explains and diagnoses — in plain language. Every change arrives first as an approval card with a reason you can actually read.

Why isn't Sabine responding?
Sabine has been suspended since 09:14: monthly budget fully used. I can raise the cap to €400 — shall I prepare that?
Resumable session · knows the full documentation
Proactive work

Schedules down to the minute, triggers from other systems via signed webhooks, handover notes from run to run. After three failures the platform stops the colleague itself.

Teams and delegation

A team lead breaks down the task, delegates to specialists, reviews and delivers consolidated. Four-eyes principle built in: whoever creates does not review.

Your processes, binding

Sick leave, incoming invoices, customer requests — described once in plain language, binding for every relevant colleague. Testable in a scenario beforehand.

Human zones

Territory that belongs to people.

You decide where AI simply does not belong — secured twice: as a binding rule at the employee's core and as a hard stop in the paths that lead outside. The Teams connector then refuses the DM reply, the mail system refuses external sending.

Human zones in detail
Direct messages External customer mail Public comments HR decisions
Working on software

And when the work is your source code.

The same colleague, a real workspace on a machine inside your network: fetch a branch, build, test, hand over. Your code stays with you.

For engineering teams
Connected outbound
A small service at your site dials into the platform. Nothing has to be opened in your firewall.
Build and test
Change files, run the build, check the tests, start a service and work against it.
Nothing is checked in
The result arrives as a package a human takes over. The service account holds no check-in permission at all.
Enterprise foundation

Built for companies that don't accept excuses.

Multi-tenant by design, not retrofitted. Data separation is enforced by the database itself — it does not depend on application code.

Foundation and operations in detail
Tenant isolation Row-level security at the database layer — separation between companies does not depend on application code.
Sign-in Enterprise SSO via OIDC/SAML, e.g. Microsoft Entra. Fine-grained roles down to a single approval permission.
Deployment Operated as SaaS, exclusively in EU data centres. No data leaving the EU. Tools that must not leave your network are reached by a small service at your site — connected outbound, with nothing opened in your firewall.
AI models Multi-provider out of the box. The model is chosen per role and swappable at any time — no vendor lock-in.
Skills Open MCP standard (Model Context Protocol): CRM, calendar, ticketing, your own internal services — including ones reachable only inside the company network.
Credentials Encrypted in the credential vault only, with a per-tenant key. Never in clear text — not even in approval cards.
Audit Tamper-evident, fully chained log: every call, every approval, every decision, every cent.
Knowledge Knowledge collections with semantic search during work, strictly separated per tenant.

Meet your next colleague.

30 minutes, one real use case from your company, no slides. We show hiring, assessment and approval live.

EU data centresEnterprise SSO from day oneNo model vendor lock-in