#security · Priya N.
94%
We ran Type II last year, I still have the evidence checklist and the auditor's punch list…
Intelligence that moves work forward.
State-of-the-art
company brain
7 sources · indexed in place · no migration
State-of-the-art
information redaction
fail-closed · audit stores counts, never values
State-of-the-art
computer use
sandboxed · every action audited
State-of-the-art
custom agentic workflows
fixed rails · frontier judgment only at marked steps
Architecture
Built for enterprise from day one.
AI only works in the enterprise when answers are secure, explainable, and permission-aware.
View A · the container
Your tools
Slack
Gmail
Notion
Jira
Drive
Docs
Outlook
Slack
Gmail
Notion
Jira
Drive
Docs
Outlook
Your infrastructure · self-hosted
Action Labs
Company Brain one living index
Every doc, message, ticket and screen, indexed where it already lives.
Orchestration deny by default
One fixed path per tool call. Personal data is masked here, before anything leaves.
Enablement where the work happens
Sandboxed agents that hold state and run for hours. Deterministic where the work is structured, frontier reasoning where it is not.
Self-hosted intelligence open weights, your GPUs
Your own open model, trained on your work, served in the container.
Cloud
Frontier models
Hyperscalers
View B · the dataflow
Your tools
Slack
Gmail
Notion
Jira
Drive
Outlook
Slack
Gmail
Notion
Jira
Drive
Outlook
read + write
scoped connectors
scoped connectors
Your infrastructure · self-hosted
1
Company Brain
One index across every tool your teams use.
2
Orchestration
Auth · policy · redact · audit. Deny by default.
3
Enablement
Sandboxed agents: computer use, custom workflows.
4
Self-hosted intelligence
Your own open-weights model on your GPUs, trained on your work.
audit log · every action · counts, never values
Cloud · optional
frontier reasoning
where it is needed
where it is needed
Benchmark
A model trained on your work beats the frontier at 1/222nd the cost.
We fine-tuned a 9B open model on one real task for about $500 of GPU time, until it beat every frontier model we tested at $0.50 per 1,000 items.
Benchmark · cost vs quality
zero-shot · 200 held-out episodes
Real benchmark · catalog integrity, 200 held-out episodes. Read the full study →
| Model | Quality (% of max achievable score) | Cost per 1,000 listings (USD) |
|---|---|---|
| Action Labs · Qwen3.5-9B + GRPO | 87.3 | 0.50 |
| Qwen3.5-9B base (untrained) | 64.2 | 0.50 |
| Claude Fable 5 | 75.9 | 111 |
| Gemini 3.1 Pro | 75.9 | 19 |
| GPT-5.6-sol | 71.4 | 29 |
| GPT-5.5 | 69.7 | 34 |
| GPT-5.5-pro | 70.3 | 172 |
TEAM
From frontier research to real-world execution.
Our team brings together frontier AI research, hands-on engineering, and decades of enterprise experience.
Fabian Hildesheim
Co-founder
AI research at Stanford HAI. Enterprise agent work at McKinsey QuantumBlack, with experience at AWS and Anyline.
Joël Hainzl
Co-founder
Built procurement agents at Tacto. AI investing at Fortino, a16z Scout experience, and a background at Kearney.
Justinas Zaliaduonis
AI Research & Engineering
ICML-published research at Stanford. Experience in knowledge retrieval at Biogenesis, engineering at carVertical, and model training and evaluation.
LinkedInWolfgang Nimführ
General Manager Europe
More than 40 years at IBM across consulting, business development, and sales leadership. Experience bringing AI, analytics, cloud, and industrial AR into enterprise organisations.
LinkedIn
GET IN TOUCH
Put intelligence behind your next breakthrough.
See Action Labs capture the work across your tools, act on it, and train a custom model on it, self-hosted from day one.
Start with a conversation about your workflow.