Samaya AI

The research bench is the deepest in our coverage. Members of Samaya's founding team co-authored some of the most-cited work on how language models retrieve and use evidence, and it shows in passage-level citation and agent design. The Agent Control Plane, launched in February 2026, runs scheduled and event-triggered research agents over long horizons. Samaya is also one of very few vendors to publish an evaluation: its FrontierFinance benchmark is self-built, and should be read as such, but its methodology is public.

Public enterprise evidence has not kept pace with the deployment. The company states it serves more than 10,000 professionals at one of the world's largest banks, and Morgan Stanley has described a partnership across its Institutional Securities Group. Against that, we did not locate the SOC 2 report type, a public no-training commitment, a subprocessor list or private-cloud deployment options, and Morgan Stanley remains the only named client. A tier-1 bank deployment implies these questions have been answered privately; buyers should ask to see the same answers.

We lower the rating to Established, at 6.82, with an Improving outlook. The change from 7.5 reflects recalibration to the full sub-criteria standard applied to every company covered since October, rather than any change at the company; our review of public sources found no material company news since September 2026. Controls we did not locate publicly — a no-training commitment, private deployment, information-barrier documentation — are now scored as absent, and company-reported scale no longer lifts commercial and workflow scores. The largest moves are in Workflow Fit (7.10 to 6.33), Commercial Value (6.20 to 5.33) and Enterprise Readiness (6.50 to 5.29). Data & Content, at 7.92, remains among the strongest in our coverage, and the shipping cadence and content partnerships support an Improving outlook.

Strengths

• Research pedigree. The founding team came from Google Brain, the Stanford NLP group, Meta FAIR and AWS AI, and has published on retrieval and long-context reasoning in venues including TACL, EMNLP and Nature Machine Intelligence.
• Broker research and entitled content. Samaya connects to a firm's approved brokers through BlueMatrix, to LSEG broker research and to Morgan Stanley Research, alongside Third Bridge, Visible Alpha, FactSet, S&P Capital IQ, PitchBook and Daloopa.
• Tier-1 bank deployment. Morgan Stanley's Global Director of Research described a partnership "across all divisions of our Institutional Securities Group", covering research, sales and trading, and banking.
• Agentic depth. Agents run one-off, on a schedule, or on triggers such as an earnings release, an FOMC decision or a new quarterly filing, and deliver finished reports to an inbox.
• Published evaluation. FrontierFinance, posted to arXiv in August 2026, comprises 220 analyst queries scored against 11,543 rubric items, with a public leaderboard that reports cost per query alongside accuracy.
• Distribution through assistants. A Query API and an MCP service bring Samaya's sources and five finance skills (earnings recap, model build, model update, variance table and fact checker) into Claude and ChatGPT.
• Capital and strategic backing. A $43.5m Series A led by NEA in 2025 was followed in February 2026 by strategic investment from NVentures and Databricks Ventures.

Cautions

• Security and deployment documentation. SOC 2 documentation is available on request through a Vanta-hosted trust centre, but the report type is not stated publicly and we did not locate an ISO 27001 certification.
• Data governance. We did not locate a public commitment that client data is not used for model training; the company refers this to its contractual terms.
• Client concentration. Morgan Stanley is the only client named publicly, and its endorsement appears in Samaya's own announcement rather than in a Morgan Stanley release.
• Desktop delivery. We did not locate native Excel, PowerPoint, Teams or Slack integrations.
• Total cost of ownership. Pricing is not published and every access path is priced by proposal.
• Team scale. Company job listings put headcount at approximately 32.
• Self-reported performance. The "98% versus 53%" financial-reasoning figure cited by the company has no published methodology, and the earnings-prediction study published in October 2026 has not been peer reviewed.