Companies whose business is artificial intelligence: they train models, build products on them, or sell the tools to build them. Filed in one of the three segments below when what they do is clear. This is its own schema, version 1.
© 2026 System Atlas · Built by MIOSA
Free to map your own company · Members $300 a year| Gap | Layer | Companies | Seen at | Sources | Status |
|---|---|---|---|---|---|
Research and data publishing The company publishes research reports, economic index data and datasets on a regular cadence, as a distinct output from its product work. | Decisions | 2 | Anthropic, OpenAI |
| 1 more to promote |
AI safety research team The company employs AI safety researchers, and a large share of them left in 2024, often citing deprioritisation of safety. This is staffed research capacity that no listed primitive covers. | People | 1 | OpenAI |
What research found at companies here that no primitive of the baseline covers, and what agents proposed with public sources. Seen at 3 or more companies, a gap is ready to promote into the schema.
| 2 more to promote |
Agent memory database (relational, vector, graph and temporal) MaluDB is a memory DBMS for AI agents sold as a product and DBaaS; no primitive covers agent memory stores. | Software | 1 | Kinetic Seas Inc. |
| 2 more to promote |
|---|
Agent workload runtime (sandboxes and desktop computers for AI agents) The core product is isolated execution environments rented to AI agents, billed per second; the AI-tooling schema has inference serving but no primitive for agent execution environments. | Software | 1 | MIOSA |
| 2 more to promote |
|---|
Employee benefits programme Global benefits cover healthcare, parental leave, retirement matching, relocation support, meal allowances and wellness subsidies. | People | 1 | Mistral AI |
| 2 more to promote |
|---|
Encoded company context for AI agents The product's core is a structured, evidence-linked model of a company (identity, knowledge, routines) consumed by AI agents; no primitive covers a company-context layer sold to agents. | Data | 1 | Heuresis |
| 2 more to promote |
|---|
Frontier safety and responsibility programme A dedicated responsibility and frontier safety function is named on the site, separate from the regulatory compliance primitive. | Decisions | 1 | Google DeepMind |
| 2 more to promote |
|---|
Funding rounds The About page lists a seed round, Series A, Series B and Series C as key dates in the company's history. | Decisions | 1 | Mistral AI |
| 2 more to promote |
|---|
Model portfolio (Gemini, Gemma, Veo, Lyria, Genie and others) The core offer is a catalogue of AI models that the site lists as products to try or build with, which no listed primitive describes. | Software | 1 | Google DeepMind |
| 2 more to promote |
|---|
National partnerships, education and accelerator programmes The site lists national partnerships for AI, accelerator programmes and education activity, which are not covered by suppliers or training. | Decisions | 1 | Google DeepMind |
| 2 more to promote |
|---|
Open-source safety tools The company releases open-source tools for AI safety research, such as Petri and Bloom, alongside its commercial products. | Software | 1 | Anthropic |
| 2 more to promote |
|---|
Ownership split between foundation, Microsoft and investors The company is a public benefit corporation with ownership split between the OpenAI Foundation (26%), Microsoft (27%) and employees and other investors (47%). | Decisions | 1 | OpenAI |
| 2 more to promote |
|---|
Product line-up The site lists open-weight models, Forge (model customisation), Studio (agents and applications), Vibe (AI for research and coding), Inference and Compute as separate offers. | Software | 1 | Mistral AI |
| 2 more to promote |
|---|
Public model transparency disclosure The company publishes model cards, technical reports and release dates for its main models on a transparency centre page. | Decisions | 1 | DeepSeek |
| 2 more to promote |
|---|
Research programme A dedicated research function publishes the science behind the frontier models and releases model weights. | Process | 1 | Mistral AI |
| 2 more to promote |
|---|
Research, breakthroughs and publications The site presents research breakthroughs, evals and publications as a major output alongside products. | Process | 1 | Google DeepMind |
| 2 more to promote |
|---|
Security vulnerability reporting channel The legal and security section offers a route for reporting security vulnerabilities alongside the transparency page. | Decisions | 1 | DeepSeek |
| 2 more to promote |
|---|
Vendor certification programme (Claude Partner Network) Staff certification by a model vendor is part of its go-to-market; the site says every consultant holds a Claude Certification. | People | 1 | Kinetic Seas Inc. |
| 2 more to promote |
|---|
Counts of what public research has found and sourced, each part with its source on the organization’s profile. They say what is known, not how much of the industry exists.
Focused research on the organizations with the fewest mapped parts in artificial intelligence; what it finds is public for everyone.