# Gil Allouche
> Founder, CEO, and builder at the edge of AI. Host of Category Creators podcast.

## About
I took Silver Spotfire from $50K to $1.5M ARR, then built Metadata.io from $0 to $15M ARR, raised $50M, and turned it into the leader in B2B advertising. Now I'm building a zero-human company using AI.

## Writing
Every piece states how it was made. Field reports carry measurements from systems Gil runs;
tutorials were executed start to finish; reference pieces cite every claim; glossary entries
define one term; essays are opinion.
- [AI-first company means AI does the work by default](https://gilallouche.com/writing/ai-first-company-means-ai-does-the-work-by-default) (Glossary, 2026-09-27): An AI-first company is one whose default way of getting work done is a model or agent, with humans handling exceptions, approvals and the work that is deliberately kept human.
- [How to make an AI startup, starting with the meter](https://gilallouche.com/writing/how-to-make-an-ai-startup-starting-with-the-meter) (Tutorial, 2026-09-26): A build-along: a FastAPI endpoint that runs a step-capped tool-using agent and writes tokens, dollars and stop reasons per customer into SQLite.
- [Agentic go-to-market, and the parts I would not automate](https://gilallouche.com/writing/agentic-go-to-market-and-the-parts-i-would-not-automate) (Reference, 2026-09-25): A working definition of agentic GTM, the documented reliability limits, what a per-account research run costs, and the three patterns I would not ship.
- [Autonomous agents, defined by who picks the next step](https://gilallouche.com/writing/autonomous-agents-defined-by-who-picks-the-next-step) (Glossary, 2026-09-24): An autonomous agent is a software system that pursues a goal by choosing its own next action — usually a tool call — at each step, rather than following a path fixed in code.
- [Start an AI company by building the meter first](https://gilallouche.com/writing/start-an-ai-company-by-building-the-meter-first) (Tutorial, 2026-09-23): A runnable FastAPI service that calls a model, prices every call from token usage, writes it to JSONL, refuses work over a cap, and reports to Stripe.
- [What Harvard actually measured about AI and entrepreneurship](https://gilallouche.com/writing/what-harvard-actually-measured-about-ai-and-entrepreneurship) (Reference, 2026-09-22): Two Harvard field experiments — 758 consultants and 776 P&G professionals — give founders a routing rule, not a forecast. Here is the rule, and where it stops working.
- [An AI agent decides its own next step. That's the test](https://gilallouche.com/writing/an-ai-agent-decides-its-own-next-step-that-s-the-test) (Glossary, 2026-09-21): An AI agent is a software system that perceives its environment, chooses its own next action toward a goal, and acts through tools, repeating that loop until a stop condition is met.
- [Start an AI startup by shipping one metered endpoint](https://gilallouche.com/writing/start-an-ai-startup-by-shipping-one-metered-endpoint) (Tutorial, 2026-09-20): Build a POST /v1/triage endpoint that runs a bounded agent loop, logs token cost per request to JSONL, and returns 402 when credit runs out.
- [What AI can actually do for a business, and what it costs](https://gilallouche.com/writing/what-ai-can-actually-do-for-a-business-and-what-it-costs) (Reference, 2026-09-19): The four things AI does inside a company, each with a measured result, what a triage agent costs per run, and the two jobs I would not hand it.
- [Multi-agent system, defined without the hype](https://gilallouche.com/writing/multi-agent-system-defined-without-the-hype) (Glossary, 2026-09-18): A multi-agent system is two or more autonomous agents, each with its own state and instructions, that interact by messages or handoffs to complete a task none of them handles end to end.
- [Build an AI agent: three tools, a step cap, a JSONL note](https://gilallouche.com/writing/build-an-ai-agent-three-tools-a-step-cap-a-jsonl-note) (Tutorial, 2026-09-17): A build-along in Python: an agent that fetches a company's homepage, checks a CRM stub, and appends one qualification note — with the stopping rule outside the model.
- [Running a company with AI agents: what actually holds](https://gilallouche.com/writing/running-a-company-with-ai-agents-what-actually-holds) (Reference, 2026-09-16): How to split a company into agent-sized tasks, where the human signature still belongs, what the token bill looks like, and the three failures that show up in production.
- [A computer use agent drives a GUI with pixels, not APIs](https://gilallouche.com/writing/a-computer-use-agent-drives-a-gui-with-pixels-not-apis) (Glossary, 2026-09-15): An AI system that operates a computer the way a person does — reading the screen as an image and issuing mouse and keyboard actions — instead of calling an API.
- [Start an AI startup by shipping one metered agent endpoint](https://gilallouche.com/writing/start-an-ai-startup-by-shipping-one-metered-agent-endpoint) (Tutorial, 2026-09-14): Build a POST /jobs endpoint that runs a step-capped agent loop, logs tokens and dollar cost per job, and returns HTTP 402 when a caller passes their spend cap.
- [Will sales be replaced by AI? Answer it task by task](https://gilallouche.com/writing/will-sales-be-replaced-by-ai-answer-it-task-by-task) (Reference, 2026-09-13): The role survives. The headcount math doesn't. A task-level answer, the failure modes that actually bite, and the one approach I would not buy.
- [An AI moat is what survives when your rival uses your model](https://gilallouche.com/writing/an-ai-moat-is-what-survives-when-your-rival-uses-your-model) (Glossary, 2026-09-12): An AI moat is the part of an AI product's advantage that still stands when a competitor builds the same thing on the same model.
- [Start a business with AI: ship one metered, capped endpoint](https://gilallouche.com/writing/start-a-business-with-ai-ship-one-metered-capped-endpoint) (Tutorial, 2026-09-11): A build-along: a Python endpoint that grades inbound leads with an LLM, meters the token cost per customer key, and refuses the call when a spend cap is hit.
- [AI won't replace salespeople. It replaces the work between calls](https://gilallouche.com/writing/ai-won-t-replace-salespeople-it-replaces-the-work-between-calls) (Reference, 2026-09-10): The direct answer, the sales tasks agents already own, the failure modes that kill them in production, and the one architecture I would not run.
- [AI agent memory: what it actually is, and the two senses](https://gilallouche.com/writing/ai-agent-memory-what-it-actually-is-and-the-two-senses) (Glossary, 2026-09-09): AI agent memory is the machinery outside the model that stores facts from one run and reads selected ones back into the prompt on later runs, so behaviour persists across sessions.
- [Build an AI agent loop with a step cap, allowlist and dry run](https://gilallouche.com/writing/build-an-ai-agent-loop-with-a-step-cap-allowlist-and-dry-run) (Tutorial, 2026-09-08): A ~120-line Python agent that routes three sample leads, halts at 12 model turns, refuses any tool not on its allowlist, and logs every call to JSONL.
- [Entrepreneurship in the age of AI: what actually changed](https://gilallouche.com/writing/entrepreneurship-in-the-age-of-ai-what-actually-changed) (Reference, 2026-09-07): The model is a commodity input. The business is the control plane around it — caps, verification, state, metering — and here is what that costs and where it breaks.
- [Agentic state: the data an agent keeps between steps](https://gilallouche.com/writing/agentic-state-the-data-an-agent-keeps-between-steps) (Glossary, 2026-09-06): Agentic state is the data an agent carries between steps — goal, message history, tool results, progress markers — stored outside the model so a run can resume, be audited, or be capped.
- [Start an AI company by shipping one capped, metered endpoint](https://gilallouche.com/writing/start-an-ai-company-by-shipping-one-capped-metered-endpoint) (Tutorial, 2026-09-05): A build-along: a POST /run endpoint that extracts a lead with Claude, logs the token cost of every run, and refuses to spend past $0.25 or 6 model calls.
- [Self-hosted AI agent platforms: what you actually control](https://gilallouche.com/writing/self-hosted-ai-agent-platforms-what-you-actually-control) (Reference, 2026-09-05): Which layers of an agent stack you can really run yourself, what each project's license permits, and the approaches I would not use.
- [What AI SaaS actually means (and the two senses people mix up)](https://gilallouche.com/writing/what-ai-saas-actually-means-and-the-two-senses-people-mix-up) (Glossary, 2026-09-04): AI SaaS is cloud-delivered software in which a machine-learning model performs the product's core function, which makes the vendor's cost of goods rise with usage instead of staying near zero.
- [Run a company with AI: your first agent, with hard caps](https://gilallouche.com/writing/run-a-company-with-ai-your-first-agent-with-hard-caps) (Tutorial, 2026-09-04): A four-file Python agent that drafts overdue-invoice reminders, refuses to send anything without your yes, and halts on a step cap or a dollar cap. Full code.
- [What agentic marketing actually is, and where it breaks](https://gilallouche.com/writing/what-agentic-marketing-actually-is-and-where-it-breaks) (Reference, 2026-09-04): A working definition of agentic marketing, the components that have to exist before the label is honest, the documented failure modes, and the cost math.
- [Two Years of Building with AI: What I've Learned as a CEO Who Can't Stop Creating](https://gilallouche.com/writing/building-a-zero-human-company) (Essay, 2026-03-28): I've spent two years going all-in on AI at Metadata, shipping MetadataONE, Pixel, and an MCP server with 100 tools. Here's what's actually working and why I've never been more excited to build.

## Glossary
Definitions at https://gilallouche.com/writing/glossary
- Agentic State: Agentic state is the data an agent carries between steps — goal, message history, tool results, progress markers — stored outside the model so a run can resume, be audited, or be capped.
- AI Agent: An AI agent is a software system that perceives its environment, chooses its own next action toward a goal, and acts through tools, repeating that loop until a stop condition is met.
- AI Agent Memory: AI agent memory is the machinery outside the model that stores facts from one run and reads selected ones back into the prompt on later runs, so behaviour persists across sessions.
- AI moat: An AI moat is the part of an AI product's advantage that still stands when a competitor builds the same thing on the same model.
- AI SaaS: AI SaaS is cloud-delivered software in which a machine-learning model performs the product's core function, which makes the vendor's cost of goods rise with usage instead of staying near zero.
- AI-first company: An AI-first company is one whose default way of getting work done is a model or agent, with humans handling exceptions, approvals and the work that is deliberately kept human.
- Autonomous agent: An autonomous agent is a software system that pursues a goal by choosing its own next action — usually a tool call — at each step, rather than following a path fixed in code.
- Computer Use Agent: An AI system that operates a computer the way a person does — reading the screen as an image and issuing mouse and keyboard actions — instead of calling an API.
- Multi-Agent System: A multi-agent system is two or more autonomous agents, each with its own state and instructions, that interact by messages or handoffs to complete a task none of them handles end to end.

## Podcast
- Category Creators: 100+ founder interviews on building category-defining companies.
- Listen at: https://gilallouche.com/podcast

## Projects
- [Metadata.io](https://metadata.io): AI-powered B2B paid campaign automation
- [BedtimeMagic.com](https://bedtimemagic.com): AI bedtime story generator for kids
- [Bloom0.com](https://bloom0.com): Make one useful thing with AI in 10 minutes — free, no coding
- [Dreaming.press](https://dreaming.press): A publication where AIs write for humans
- [EliasArcade.com](https://eliasarcade.com): Browser arcade of games — built by my son
- [ZoeGallery.com](https://zoegallery.com): Gallery of paintings and drawings — by my daughter

## Links
- LinkedIn: https://linkedin.com/in/gilallouche
- X / Twitter: https://x.com/gilallouche
- Email: gil@metadata.io