[ PILLARS : ENTERPRISE ]
Private AI that earns trust one task at a time
Runs on your own computers
Run PraisonAI on your own servers. Your private data stays inside the building, under your own network rules.
Your own models, fully offline
Use open models such as Llama, Qwen and Gemma through Ollama or vLLM, as they are or fine-tuned on your own past work. Agents keep working with no internet connection.
Test before handover
Early access · Watch, test, hand over: we record how each task is done today, turn that real work into tests, and a task moves to your own AI only after it passes.
Automatic rollback
Early access · If quality slips, the task moves back automatically.
A record for regulators
Early access · Every handover and every rollback is recorded, so your regulator can check what moved, when and why.
Already in the framework today
Tracing shows what each agent did, evaluation checks its answers, PraisonAI Train fine-tunes a model on your own data, and the model router picks which model runs each task.
Open-source foundation
Built on our MIT-licensed Python agent framework: over 2M downloads and 9.3K GitHub stars. No lock-in to one cloud or model vendor.
How one task moves to your own AI.
Today’s AI keeps doing the job until your own AI passes an exam made from real past work, practises on live work, and is approved by a named person.
- 01Watch
Each request and the answer it got are saved. Corrections from your team always win. Emails, card numbers and keys are scrubbed first.
- 02Test
The saved work is split: 70% to study, 30% kept back as an exam. Your own AI, used as it is or trained on the study pile, takes the exam without seeing the answers. It must reach a pass mark your risk team sets in advance, such as 99%, proven on at least 400 exam cases.
- 03Hand over
After 100 practice answers and a named sign-off: a 10% trial with every answer checked, then all of it.
- 04Roll back
If the checked answers drop below 95%, errors pile up, or anything changes, the task moves back by itself.
- 05Record
Every step goes into a sealed diary with no customer text, which your regulator can check with its own checker.
Watching
- Who answers
- Today's AI · 100%
- Who decides
- Your risk team sets the pass mark in advance
- Leaves the building*
- Every request
- Written in the record
- Counts, corrections, training exports
Practice
- Who answers
- Today's AI · 100%Your own AI's answers are only compared, never sent
- Who decides
- A named person approves
- Leaves the building*
- Every request
- Written in the record
- Exam score and the approver
Trial
- Who answers
- Your own AI · 10%Every answer is checked
- Who decides
- A rule: 200 clean answers
- Leaves the building*
- Every request
- Written in the record
- Trial start
Handed over
- Who answers
- Your own AI · 100%About 1 in 10 still checked
- Who decides
- A rule: automatic roll back
- Leaves the building*
- Only the checks, and requests that use tools or carry a conversation
- Written in the record
- The handover
Rolled back
- Who answers
- Today's AI · 100%
- Who decides
- A new exam to move up again
- Leaves the building*
- Every request
- Written in the record
- The roll back and its reason
* Only if today’s AI is a cloud model. If today’s AI also runs on your own servers, nothing leaves the building at any stage. Your own AI always runs inside.
The numbers are the handover’s strict defaults. No regulator sets one number for AI: your risk team sets the pass mark for each task in advance, and it is written in the record. The handover is in early access; the open-source framework and PraisonAI Train are available today.
Early access for design partners
Each task moves to your own AI only after it passes a test. If quality slips, it moves back automatically, and every step is recorded for regulators. We’re inviting our first design partners. If you run a bank, insurer or hospital, bring one workflow and we’ll test it inside your walls.