# Bootstrap prompt

Paste the prompt below into an agent that can reach the Metix AI Platform: one with the [metix-skills](https://github.com/MetixAI-Official/metix-skills) installed, the MCP server connected, or plain REST access with `METIX_KEY` set. Followed end to end, it costs about 2,400 to 2,800 API Credits: about 2,000 for the counts, about 240 for the searches and records it reads to choose and check (the largest companies, a slice of their staff, the small teams with big rounds), and 150 to 450 for the audits (the definition slices, the theme slices, the growth check). It stops and asks before 3,000. Every published company number is a count; the profile numbers are counts too, and records are read only to choose and check.

```text
Answer one question with the Metix AI Platform: ChatGPT launched on 2022-11-30; how many companies that describe themselves as AI companies were founded in the years after it, where are they, what do they say they build, how big have they grown, how are they funded, and where did their staff work before? Work only through the public Platform (REST at https://mira-api.metix.ai, the MCP server, or the metix-skills) with the key in METIX_KEY, and never print the key.

1. Read before querying. Call GET /contract (free) and build every condition from querySpecByEntity.company and querySpecByEntity.profile; read GET /docs/api/companies and GET /docs/api/query-spec (free). keywords, name, industry, and headquarters.city are free text: match and eq both mean every word present, in any order, ignoring case, and in means any of several such values. size takes nine closed values. A total of 100,000 or more comes back as the string "100000+". A query holds at most 64 conditions. A count with size 1 costs 1 API Credit, a search 1 API Credit per 25 IDs, and a detail read 1 API Credit per 5 records. Check the balance with GET /auth/key/status (free) at the start and the end, and stop and ask before the run passes 3,000 API Credits.

2. Population. An AI company is one whose type is not Nonprofit, Educational, or Government Agency, and either whose name has the word AI and whose keywords match any of AI, artificial intelligence, generative AI, machine learning, large language models, LLM, deep learning, computer vision, natural language processing; or whose keywords match any of deep learning, computer vision, natural language processing, large language models, generative AI. Cohorts by founded_year (a year only): 2019 to 2022 (before; the launch fell on 2022-11-30, so 2022 counts as before), 2023 to 2025 (the class), 2026 so far. Write the definition and the cohorts to files.

3. Audit the definition before any measure. Search ranks by match quality, so never judge a definition from the top of a search. Read whole slices: every match in one founding year with linkedin_followers in a narrow window, 15 to 50 companies, one slice per year from 2016 to 2025 (about 200 companies; _source id, name, industry, keywords, type, founded_year). Classify each by hand as AI is the product, AI is part of the core product in another field, a generalist firm listing AI among many services, or not AI or not a company; the first two are AI companies. Report the share, by founding era. Then read the matches of the nine loose terms that the definition leaves out, in the same kind of slices, and report how many are AI companies. Expect about 78% and about a third: tags are written freely, generalist firms list AI terms, and a name alone keeps newer companies more than older ones. Say so rather than tuning the rule to the reads.

4. Founding wave. For each founding year 2010 to 2026, count AI companies and all companies. All companies is banded in most years: split it into size bands and no size, split a banded band by headcount (0, 1, 2, 3 to 4, 5 to 10, 11 or more, none), and add the parts. Publish the AI share of all companies founded that year; the share survives the lag that thins every recent year. Beside it, count the same share for each branch and for wider and narrower definitions: the name branch alone; the technical branch alone; the technical branch with only deep learning, computer vision, and natural language processing, which no naming fashion or post-2022 word can move; the probe terms; the loose terms. State only what holds across them. Show the lag: count AI companies per founding year by updated_at month and by whether a headcount is stated.

5. Where. For the class and for 2019 to 2022: headquarters.country for about 60 candidate countries, the top 15, the rest, and no country, adding up to the cohort. Metros as a country, a list of cities in one in condition, and a state that is either named or missing (San Francisco has no state on 165 of 891 class companies); Singapore is the whole country. Publish each share both of the cohort and of the companies that give a city (or a country), and repeat the Bay Area, the city of San Francisco, and London under the probe and loose terms. Count all companies headquartered in China per year to show how thin that coverage is.

6. What they build. For each founding year 2016 to 2026, the share of AI companies whose keywords match each of 15 themes: generative AI; LLM; AI agents or agentic; computer vision; NLP; machine learning; robotics (not robotic process automation); data analytics; SaaS; health; fintech; cybersecurity; developer tools; voice and speech; AI infrastructure (GPU, inference, MLOps). Read one small slice per theme and drop any term that matches through words from different tags (financial technology, information security, AI infrastructure) or marks service firms (devops). Count the same themes inside the loose population too (any loose term, same type exclusion), so the name branch entering without tags cannot drive a theme's fall.

7. Size, growth, funding. Size bands per founding year. headcount_growth_yoy_pct in six bands and not stated, for headcount 11 or more only, class against 2019 to 2022; read 30 records first to check the values are percentages. Share of each cohort with last_funding.date and with last_funding.amount (the latest round only). Latest rounds by quarter 2023Q1 to 2026Q3. Amount bands only for United States headquarters: amounts elsewhere are in local currency (a Korean team of eight shows 23,000,000,000). Then read every United States AI company founded 2022 or later with headcount 50 or less and a latest round of 50,000,000 or more, and drop rounds dated before the founding year, billion-dollar rounds for such teams, and companies where AI is one tag among many. Count the growth bands, funding coverage, US amount bands, and size bands for every company of the same types as well, with no AI filter, so an AI figure can be told from a young-company figure. End the quarter chart at 2026Q1: records were refreshed in April and May 2026.

8. Staff. Read the AI companies of the class with headcount 60 or more and take the largest 150; the same for 2016 to 2019 with headcount 200 or more. Count current profiles per name (has_experience with experience.company.name eq the name and experience.is_current eq true); drop names that return more than 1.5 times the headcount, and divisions, labs, and communities that are not companies. Classify every remaining company by hand with the scheme of step 3 and keep only AI companies; mark data-labeling marketplaces, whose visible staff are mostly contractors, as their own group, and report every staff figure with and without them. Read one whole slice of the kept companies' staff (total_experience_months in a narrow window) with experience.company.id and drop names where most profiles carry another company's id. Then count people currently at any kept company (one in condition) with an earlier, non-internship entry at Google, DeepMind, Meta, OpenAI, Anthropic, Microsoft, Amazon, Apple, NVIDIA, Stripe, Uber, Airbnb, Salesforce, Palantir, Databricks, and Scale AI (not for the 2016 to 2019 baseline, which includes Scale AI itself), and at any big tech or any frontier lab; and those whose current_title matches founder.

9. Outputs and charts. Aggregates with "unit" (companies, or profiles for the staff), the snapshot date, and the query files. Profiles: 1 to 9 as "<10", shares only on a base of 30, and no rollup published that differs from its shown members by 1 to 9. Charts: the AI share by founding year with the partial years marked; top countries and metros for both cohorts; the theme shares as small multiples across years; growth bands; rounds by quarter; the small-team list as a table; earlier employers against the baseline. Titles state findings in neutral words.

10. Limits. A precise core, not a census: the definition keeps about four in ten AI companies and about one in five of its matches is not one. How large the rise after the launch looks depends on how companies describe themselves: by names and new vocabulary it roughly triples, by the older technical tags it is about a fifth, and counts cannot tell more AI companies from more companies saying so. founded_year is a year, and records were refreshed in April and May 2026, so 2025 is incomplete and 2026 covers a few months. last_funding is one round, amounts are in local currency outside the United States, and headcount is what the company page reports. Staff counts are visible profiles, joined by name, and never a headcount.
```

## Adapt it

| To change | Edit | For example |
| --- | --- | --- |
| What counts as an AI company | Step 2, then audit it again in step 3 | Add machine learning to the technical terms, and expect lower precision |
| The launch you date from | The cohorts in step 2 | Companies founded after GPT-4 (2023-03-14) would need a month, which founded_year does not have |
| The country | headquarters.country in steps 4 to 8 | Only United Kingdom companies, with London and Cambridge as metros |
| The employers | The list in step 8 | Add xAI, Mistral AI, or ByteDance |
| The API Credit ceiling | Steps 1 and 8 | Skip the staff join (step 8) to save about 530 API Credits |

## What to ask before running it

When someone brings a looser version of this question, settle these first. Each one changes the query or the cost:

1. What makes a company an AI company: its name, its tags, or both, and how much noise is acceptable?
2. Which founding years count as after the launch, given that the Platform stores a year only?
3. Which countries, and should amounts be compared across currencies at all?
4. Should the staff join run, and against which baseline cohort?