Unsupervised

Unsupervised is a Boulder, Colorado-based company that builds tools and systems for human supervision of AI agents, spanning what it calls agentic engineering and agentic data analysis. Its product line is split into two groups. Agent Control Systems include DeepWork (workflows agents can follow without constant supervision, offered as source-available software), Panopticon (open-source exception-based monitoring of unattended agent runs), Outfitter (open-source dotfiles management for coding agents, allowing profiles to be created, shared and switched) and 2119 (open source, enforcing that every requirement has a test). Agentic Data Analysis includes Finder, a free tool intended to give agents evidence before they act, and Finder for Teams, which runs against data-warehouse data to produce KPI-linked patterns ranked by estimated financial impact, with supporting evidence that an analyst reviews before action is taken.

The company began in 2017 by automating analytics for enterprises including AT&T and Disney, and describes Finder for Teams as the environment in which it learned to control agents in production. Its stated principles are that agent autonomy must be earned through explicit boundaries, checks and evidence in the harness; that humans retain judgment, direction-setting and risk approval rather than watching each step; and that runs should expose what changed, the basis for results, areas of uncertainty and items needing review. Earlier coverage described the platform as an AI built on unsupervised learning that scans structured, unstructured, incomplete, first-party and third-party data to discover and rank patterns relevant to a defined business question, with customers in CPG, retail, telecom and financial services.

Unsupervised publishes research and field notes alongside its software, including work on SemLang, described as a semantic language giving agents a governed view of company data, cost readouts comparing frontier and open-weight models under structured workflows, a published end-to-end run analyzing HHS Medicaid fee-for-service spending, and an AT&T case study.

Founding story

Unsupervised was founded in 2017 by Noah Horton (CEO) and Tyler Willis (co-founder, president and chief product officer). The two met while building Involver, a social media management software company co-founded by Horton that was acquired by Oracle and became the nucleus of Oracle Marketing Cloud; Horton subsequently led Oracle's public cloud efforts, while Willis worked in business and corporate development at a data company and later as a consumer CMO. Horton, a Boulder native with a math and computer science background, located the company in Boulder rather than Silicon Valley to reduce engineering attrition.

Business model

Unsupervised distributes a mix of open-source, source-available and free-to-use developer tools (Panopticon, Outfitter, 2119, DeepWork, Finder) alongside Finder for Teams, an enterprise product sold to organizations that connects to warehouse data and is presented through a demo-request sales motion. Earlier coverage describes enterprise deployments with Fortune 50 to Fortune 5000 customers.

Enterprise software sold to organizations for warehouse-connected analytics (Finder for Teams), with other tools distributed as open source, source available or free to use.

Traction

The company reports more than $1B in actionable insights found by customers since 2021, over $100M in estimated value from insights AT&T put into action, and $58M found for a Fortune 500 healthcare payer in six months. Benchmarks cited on its site include 65.7% fewer hallucinations than peer tools on natural-language data queries (DA-Bench) and, on AgentIF-OneDay, 36% fewer zero-score failures and 2.4x tighter run-to-run consistency when the same agent runs DeepWork workflows. As of 2021 coverage, it had 75 employees and worked with dozens of companies.

Latest developments

As of March 2026 the website presents the Agent Control Systems line (DeepWork, Panopticon, Outfitter, 2119) and Agentic Data Analysis line (Finder, Finder for Teams), a published full run analyzing HHS Medicaid fee-for-service spending, research on SemLang, readouts on model cost and speed under structured workflows, and an AT&T case study. The company is recruiting for work on agent reliability and control.