BPMN · DMN · CMMN on Flowable
One process, all of your companies.
Publish once, with the AI agents and the knowledge inside. Each company runs isolated in its own schema, with its own AI quota, and adjusts what is theirs. All inside your network.
Publish once, with the AI agents and the knowledge inside. Each company runs isolated in its own schema, with its own AI quota, and adjusts what is theirs. All inside your network.
For whoever operates more than one company
Headquarters, branches, a franchise or a group. The purchasing, approval or onboarding process is published once and applies to all of them — and the unit with its own rule overrides only what is theirs, without becoming a parallel version for someone to maintain.
A consultancy, a BPO or an accounting firm. Each client is an isolated tenant, with its own AI quota and its own knowledge — and the improvement you make to the process reaches all of them at once, without repeating configuration client by client.
Infrastructure
Process data, attachments and credentials stay in your database and your object storage, inside your cluster. The install is a chart on the Kubernetes you already run.
Robots never run on the platform machine: you register as many runners as you need, each one next to the system it integrates with. The platform itself grows in CPU and memory on your cluster, against your Postgres.
The executing agent is installed on a machine you choose and makes outbound connections only. It reaches the internal ERP with no inbound firewall rule and no tunnel.
Bring your own key, or point at any OpenAI-compatible endpoint: your own vLLM, the corporate gateway, an Ollama behind a proxy.
Model Context Protocol
The platform exposes an MCP server. Claude Code, or your IDE's assistant, creates processes, forms, agents and knowledge bases by talking to the platform — instead of you pasting JSON into a screen.
The assistant inherits the permission of whoever connected it, and nothing beyond it. A tool that writes does not answer to someone who could only read.
Authorisation sits on the process side, not the client's. A sensitive flow simply does not show up for the assistant.
Connect the leading market models or your own OpenAI-compatible endpoint: a vLLM in your room, the corporate gateway, an Ollama behind a proxy. Switching models is changing a setting, not rewriting a process.
The system counts real decisions from your own history and suggests values with the confidence score and the count behind it. It is counting, not inference: zero AI cost. And the suggestion goes to a person — the platform does not decide on its own.
Upload contracts, manuals and case history as PDF, TXT or Markdown. Agents retrieve the passages closest to that step's instruction, and the search is isolated per tenant at the database connection level.
The ERP password never enters the prompt, never stays in the conversation history and never shows up in the log. The platform swaps the secret in at the last instant, already on its way out to the network — the model never sees it.
AI consumption is attributed per process and per step, not as a total at the end of the month. Each company has its own quota: the copilot refuses when it runs out, and a running process never stops halfway — the overage is recorded.
Structured process, business rules versioned in a decision table, and unstructured case management on the same open-source engine. Every robot execution leaves a log sealed in storage, ready to download.
Designs the platform serves today
The agent reads the PDF attached to the task, cross-checks it against the knowledge bases linked to the process, and writes the opinion. The person still signs it: the next step is a human task with the text already filled in.
AI + human decisionApproval thresholds, risk criteria and sign-off policies live in a DMN decision table, which the administrator changes without shipping code and without a new version of the process.
DMN · no new deployThe platform counts what was decided in comparable situations and suggests values with the evidence in plain sight — "34 of 38 cases". The suggestion goes to a person, who accepts it or changes it.
Suggestion with evidenceA JAR of yours runs on the executing agent, inside your network, and talks to the platform over RPC. Every execution has its own log, sealed in object storage and available to download.
JAR · RPC · sealed logOperation and continuity
The whole install is a versioned chart, with the documentation alongside it. Your team brings the environment up from scratch without depending on us to operate it.
The code sits in escrow with a trigger defined in the contract. Once the trigger fires, it is yours — and the platform keeps running where it already is.
Every release ships with its dependency list and the vulnerability scan result, so your security team can check it without having to ask.
A critical vulnerability has a remediation window agreed in the contract, with a stated deadline — not left to goodwill.
Closed scope, written acceptance criteria and one of your processes live at the end. You talk directly to the person who wrote the code.
Book a proof of concept