Methodology

Once a day, early in the New York morning, a script reads every board and records what is open. A posting disappears when a role is filled or withdrawn, so anything not captured that day is gone for good. For that reason roles are classified and their text is read at capture time, and only the derived figures are kept. We do not store job descriptions.

Two safeguards protect the series. If fewer than four in five of the expected boards respond, the run records nothing, because a partial outage looks exactly like a hiring collapse. And a date is never recorded twice.

Each role is classified from its title into one of nine functions (engineering, research, product, design, go-to-market, operations, people, finance and legal, other) and one of seven levels, from entry to leadership. The rules were audited against real titles before first use: an early version filed “Senior Product Manager” under management and “Member of Technical Staff” under other. Titles are inconsistent across companies, and a share of roles will always land in the wrong bucket. We have not measured that error rate, which is a limitation.

A role counts as New York when its listed location names New York, NYC, or Brooklyn. Remote roles and roles at New York companies’ other offices are counted separately, because a company can have the largest board on the list and almost none of it local.

The AI skills glossary and the tools figures in the Hiring Index come from reading each job description for named tools (75 of them, from model providers to data platforms) and concepts (33, such as RAG, evals, and fine-tuning). On September 20, 2026, 90 terms appeared at least once across 1,567 postings.

• Companies first. The headline for any term is how many companies name it. Posting counts are secondary.
• The role’s own words. Most companies open every posting with the same paragraph about themselves. We remove any sentence a company repeats across a large share of its postings before reading. Without this, a voice-AI company’s finance roles count as voice-AI jobs. Tested on 1,131 postings, it left company counts almost unchanged and cut some posting counts by more than four-fifths.
• Ordinary words, strictly. Sales postings “fine-tune” messaging, finance postings set “guardrails,” and people get asked things “without prompting.” Such terms count only beside the technical language that gives them their meaning, acronyms are matched case-sensitively, and a term that cannot be separated from its everyday sense is left out.
• No self-counting. A vendor’s own job postings never count toward its own name.
• Named is not used. A posting names what an employer asks for. A nice-to-have counts the same as a requirement, and none of it is evidence of what runs in production.
• Pairings by lift. “Named alongside” lists rank terms by how much more often they appear together than apart, so that Python, which appears beside everything, does not crowd out the informative pairings. A pairing must be made by at least a fifth of the term’s companies.

An annual list of 100 AI companies headquartered in the New York area, in four tiers by size: $10 billion and up, $1 billion to $10 billion, a growth tier ($200 million valuation or $75 million raised), and an emerging tier. The top tier is ranked; the rest are alphabetical, because the public figures below that level do not support a ranking.

• Headquarters verified. Each company’s New York headquarters was checked individually. Metro-area headquarters are labeled with their city. Big-technology offices are listed separately as outposts and never in the 100.
• Figures dated. Every valuation or funding total shows when it was established. Where the latest mark is old, the date says so.
• Judgment disclosed. Companies where AI powers the product without being the product are footnoted.
• Verify or omit. Founding years, addresses, and rounds appear only where we found a source. Missing fields are left empty, not guessed.
• Annual, and never paid. It is an edition, compiled in August 2026 and revised yearly, not a live directory. No company can pay to be on it.

The episode guide covers all 37 episodes of the AI in NYC Show. Dates and running times come from the show’s public podcast feed, and video links from its public playlist. Guest names and titles are as given in the show’s own notes at the time of recording. The summaries are written by us from those notes and are not transcripts.

• Startup-weighted. Our coverage is venture-backed AI companies. The city’s banks, insurers, law firms, and hospitals hire heavily for AI and are not in it, so anything that matters most there, governance above all, is understated.
• Three posting systems. Companies that post elsewhere are missing from the hiring data.
• A few large boards. A handful of companies account for a large share of all postings. This is why companies are our headline unit, and why any posting-level figure should be read beside the company count.
• Short history. The series began in August 2026. We do not yet know what is seasonal.
• Postings are intentions. An open role is not a hire, a posted range is not an offer, and a named tool is not a deployed one.
• Pattern matching. Titles and descriptions are read by rules we wrote. They were audited on real postings, and they will still misread some.