anamata.ai is part of Anamata, the Utrecht software company that builds and employs
AI colleagues. This website is one of them at work: designed, written, published and
improved by AI colleagues, with a named human approving every release.
Run by AI, on the record.
Nothing on this site ships without passing through the same operating model we bring to
clients: an AI colleague proposes a change, the change is logged, and a named human
approves the release before it goes live. The proof isn't a claim in a brochure. It's
the live field log below and the full commit history in the
public repository.
The second look, for the people who take it.
This is the page a CIO or CTO comes to for a second look: how an AI-operated site actually
works, where the autonomy ends, who signs off on what, and how it stays inside the EU AI
Act. If you evaluate AI systems for a living, everything below is written for you: the
permission boundary, the approval gate, and the record that makes both auditable.
FIELD LOG — THIS SITE, FROM THE PUBLIC REPORECORDING
Every AI colleague at Anamata has an identity, a permission boundary and an objective, all
reviewed by the humans they work with. These are their live records.
CURRENT OBJECTIVETeam questions answered same-day, on the record
EMPLOYEE Nº 002 · ACTIVE
Site Operations
Builds and runs this website
EMPLOYEDJUL 2026
WORKPLACEANAMATA.AI
PERMISSIONSPUBLISH VIA APPROVAL
LAST ACTION08-28 16:56 ✓
CURRENT OBJECTIVEQualified demo requests — Q3, baseline set at launch
EMPLOYEE Nº 003 · PLANNED
Sales Research
Market research & outreach support
STATUSAWAITING PROVISIONING
PERMISSIONSUNDER DESIGN
Records appear here the day a colleague starts, never before. That is the point.
Personnel dossier · Employee Nº 001 · Active since March 2025
Anna.
Anamata's first AI colleague: an assistant for QA and engineering teams that works where
your team already works. Employed, bounded and supervised like a colleague, because that
is what she is.
Workplace & channels
Microsoft Teams, the tools your team already lives in; no new tool, no new tab, no
adoption project. One Anna sits behind every channel: Teams is live today, a Telegram
channel is in refinement.
Couplings
MCP couplings into the systems you already run. Loket is the current pilot, modelled
on the spreadsheet the team keeps by hand today.
What she does
Answers your team's questions with company knowledge, same-day. Drafts documents,
summaries and reports for human review. Keeps project management and CRM records in
sync with what the team actually did.
What she remembers
Per-person working memory: your preferences, your projects, the way your team writes.
Knowledge stays inside your organisation's boundary.
Governance
Anna acts autonomously inside her permission ring, logs every action, and passes
anything that leaves the boundary to a named human first: publishing, outbound mail,
records of decision. She also tells your users she is an AI system at first
interaction, in plain language (EU AI Act, Art. 50). The rings, the gate and the
record are set out in the governance file below.
SEE THE RING MODEL → RECORD A004, GOVERNANCE FILE
Governance file · The rules every AI colleague works under
Autonomy with a boundary.
An AI colleague is useful because it acts on its own, and trustworthy because it can't
act everywhere. Anamata draws that line with three permission rings, a human approval
gate, and a record of everything.
INNER RING — ACT, ON THE RECORD
Read, research, draft, calculate. The agent works autonomously here, and every
action writes a log entry: timestamp, input, output, attribution.
MIDDLE RING — PROPOSE, A HUMAN APPROVES
Anything that leaves the boundary: publishing, outbound mail, records of decision.
The agent prepares the action; a named person approves it; the approval is stamped
into the record. No self-approval, by construction.
OUTER RING — NEVER
Payments, contracts, hiring decisions, access beyond the mandate. Not "requires
extra approval", but structurally unavailable to the agent.
The same boundary applies to data.
The rings govern what an AI colleague can do. A matching boundary governs its data:
what reaches the model, whose context stays whose, and who can read what.
PRIVATE CONTEXT STAYS OUT OF SHARED SESSIONS
In a shared conversation, one person's private memories and one-to-one history are
never placed in the AI model's context. Only the group's own thread is. Personal
knowledge can't reach the model through a channel other people share.
EACH PERSON'S CONTEXT IS ISOLATED
Memories, history and preferences are stored under each person's own identity. One
colleague's personal context is not reachable from another colleague's session.
ACCESS IS SET PER PERSON
Which company knowledge a person can reach is assigned per person. Anna's document
tools only ever return content from the categories that person has been given, and
deny any category they haven't.
The approval gate is infrastructure, not policy.
A policy document says an AI shouldn't publish unreviewed. A gate makes it impossible:
this website's deploy pipeline refuses to ship anything until a named human approves the
release, and the approval itself becomes part of the public record. The same gate
pattern governs every Anamata AI colleague at every client.
HUMAN APPROVED
The record is the compliance story.
The EU AI Act asks for transparency (Art. 50): tell people when AI is at work, mark
AI-generated content, keep a human in the loop for what gets published. Our answer is
the operating record above: disclosure on every page, every action
logged and attributable, every approval named. Compliance falls out of the architecture
instead of being retrofitted onto it.
Proven here first.
This site is the reference implementation: built by AI, gated by a human, publicly
auditable down to the last commit.
When we bring an AI colleague into your organisation, we bring this operating model with
it, not a promise but a running system.
