Skip to content
Back to home

How it works

One controlled layer over the systems you already run.

Agents reach your tools through an orchestration layer that enforces the rules you sign off. It sits on top of what you have and replaces nothing.

Your people

  • Approve
  • Edit
  • Reject
  • Set the rules

Orchestrator

Runs in your environment

Orchestrator controls

Context: Gives each agent only the data its job needs, pulled live from your systems and never copied into a separate store.

Policy controls: Your rules, limits and approved wording are enforced in code, so an agent cannot step outside the mandate you signed.

Workflow routing: Sends each task to the right agent and each result to the right person or queue, with deadlines tracked.

Human approval: Anything that sends, pays, changes a limit or touches a customer waits for a named person. Four-eyes where you need it.

Audit trail: Every input, source, suggestion, approval and override is logged in your systems, ready for auditors and regulators.

Specialised agents

Read by default · write only through approved tools

Your existing systems

  • Email and chat
  • CRM
  • ERP and accounting
  • Helpdesk
  • Document storage
  • Spreadsheets
  • Payment APIs
  • Policies and rule books
  • Partner APIs
  • Case management
  • Data warehouse
  • Internal APIs
Your people approve, edit, reject and set the rules for an orchestrator that runs in your environment. It routes each task to the right agent; agents read your existing systems and write only through approved tools.

One task, start to finish

What happens when work arrives.

4 steps, the same for every agent. Nothing leaves without a person's say-so.

  1. Step 1

    Something comes in

    An email, a document, an alert, a form or a scheduled report triggers the agent.

  2. Step 2

    The agent prepares

    It reads what it needs, checks it against your rules and drafts the result, with every source linked.

  3. Step 3

    A person approves

    Your mandate decides what runs on its own and what waits for 1 or 2 approvers.

  4. Step 4

    Done and logged

    The action goes through an approved tool, and every step is recorded in your systems.

Mandate levels

You decide how far each agent may go.

Every agent starts at read-only. You unlock each level, one step at a time.

  1. L0 · Default

    Read

    Looks up, compares and prepares. Changes nothing.

  2. L1

    Act within limits

    Low-risk actions you list, such as tagging a case or opening a ticket, each with its own limits.

  3. L2

    One approver

    Sending a reply or updating a record waits for a named person.

  4. L3 · Off by default

    Four eyes

    Anything that touches money, limits or a client relationship needs 2 approvers.

Deployment

Your data stays where it is.

Pick the setup your rules require. Every option comes with the same 4 guarantees.

  • Most common

    Your cloud account

    Your AWS, Azure or Google Cloud, under your own contract.

  • Sovereignty

    Sovereign or local cloud

    A cloud provider owned in your country.

  • No third-party AI

    On-premises

    Open models on your own servers. Nothing leaves your network.

  • Best of both

    Hybrid

    Personal data stays local. Only masked text goes to the cloud.

  • Zero access by default
  • Runs inside your infrastructure
  • Switch models anytime
  • No vendor lock-in

See it mapped to one of your processes.