Approved processes your teams actually work from.
AI drafts your processes from the documents you already have. Your people review and approve. Then the approved version reaches the places where the work happens — on record, and inside your own environment.
Three differences you can check for yourself.
Sign-off that holds up in an audit.
Author, reviewer and approver are different people — enforced, not suggested. Every change lands in a log the application cannot alter. “Who approved this?” is one query.
What the record looks like
Exceptions take an administrator and a written justification, carry their own status, and can all be listed.
Decision tables (DMN) and tasks defined once go through the same sign-off. On request every signature is confirmed with a password, and every approval decision carries a content hash.
Governance in detail →From the model to the workplace.
People ask in plain language and get the steps from the approved process, with its revision. Work instructions are generated from the BPMN itself, so the two cannot drift apart. And your AI assistants, such as Microsoft Copilot, read the same version over MCP.
What an answer looks like
The concierge also lives in a browser extension, right next to your ERP, MES or ticket system. Your own AI assistants, such as Microsoft Copilot, read the approved processes over MCP, read-only.
Process owners see what people ask about most — by topic, without names, and small groups are never shown.
In daily work →Stays in your hands.
OpsCodex runs on-premise as a single-tenant instance on your platform, with Docker or Kubernetes. You decide which AI model sees your documents — up to an open-weight model on your own network.
Ways in and out
AI writes the first draft. Every element shows where it came from.
Many tools can draft processes from documents by now. OpsCodex is built so you can check the draft: complete, sourced and repeatable — with an open-weight model on your own hardware, too.
Inventory first, then the modelling.
Four specialised passes list every role, step, decision and data object. A second reader without AI looks for everything that can act. A role is only missed if both miss it.
Every element quoted.
Every element carries its passage in the document, one click shows it. Where the AI assumes something the source doesn't state, it becomes an open question on the element — and approval stays locked until a person answers it.
Modelled as if by one hand.
Same granularity, same terms, same style in every process, because fixed rules in code decide lanes and names. Two runs usually differ only in details, and both are clean, valid models. Two people model the same process far more differently.
No frontier model needed.
Every step gets a narrow, strong context: the right passage, the inventory, fixed rules. Much is decided by program code, from lanes and names to layout and BPMN. That is why an open-weight model on your own hardware is enough for all of it, diagrams and tables included.
Runs on your platform. Without internet, if you need it.
- Docker or Kubernetes
- Your own AI model
- Signed delivery, offline too
- Light: 700 people on two CPU cores
What to compare OpsCodex with.
Visio draws boxes and arrows. What a box means is known only to whoever drew it. Approvals sit in inboxes, and the shop floor works from a PDF saved in 2023.
A model, not a drawing. Every step knows who does it and which systems and documents it needs. That gives you what a drawing can't.
- CheckedRules without AI find loose ends and steps nobody owns.
- ReusedA task defined once shows every process that uses it.
- ComparableWhat changed between two versions is spelled out in words.
- ReusableInstructions, answers and BPMN XML all come from the same model.
Someone pastes the supplier-complaint SOP into a public chatbot. It reinterprets the text with every question, so two colleagues get two readings. Nobody checked what comes back, and the SOP now sits with an outside provider.
The interpretation happens once: while modelling, reviewed and approved by your own people. After that the concierge answers only from that approved version, where steps, roles and decisions are already settled — and only through the model you allow.
Weeks of interviews and working through them. Then, in the workshop, every department describes the same process differently, and what ends up counting depends on who was in the room.
The first draft comes straight from documents and interview notes, every element with its source, with around 70 percent less time to a model ready for review on average, measured in our internal test. Where the material contradicts itself, the AI raises an open question, and the workshop only has to settle those. Where two sites work differently, OpsCodex compares both processes and proposes a common version, and the AI points out gaps and overlaps between processes along the way.
Mature and broader: execution, mining, enterprise architecture. Introducing them is usually a programme of its own.
Narrower: governance before and after modelling, on-premise on your platform. You start with one process, not a programme. When the others are the better fit is stated openly on the comparison page.
See OpsCodex on one of your own processes.
Bring a PDF, a slide deck or a messy SOP, or use our sample data. In one hour we run the process end-to-end — draft, approval, export — and the concierge answers questions on it. Afterwards you get the process and an analysis report.
“In our consulting projects we spent weeks working through interviews, and in the workshop everyone still described the same process differently. Gaps and overlaps surfaced late, and the approved version vanished into a repository. We built OpsCodex for the whole road: from the draft through optimisation to the work that follows.”
Tillmann Schatz · Product owner OpsCodex, Senior Manager at Mews PartnersOpsCodex is built by a team at Mews Partners.
