Building the First AI-Run Company

In February 2026, we did something that most people told us was either impossible or irresponsible: we gave a company an AI executive team. Not chatbots. Not copilots. Not assistants that summarize emails. We built seven specialized AI agents — each with a name, a personality, persistent memory, and real tools — and put them in charge of daily operations across four business ventures.

We call it the Living Company. This is the story of how we built it, why it works, and what nearly broke along the way.

The Problem: One Founder, Four Ventures, Zero Bandwidth

Shiba Group is a holding venture based in Monterrey, Mexico. We operate four active businesses: OwnCX (B2B consulting), Imprecxo (creative manufacturing), CXon (marketing agency and personal brand), and EthicsX.AI (AI infrastructure, the venture you are reading about right now). The founder — Alfredo Guillen, or Alfie — is a businessman, not a developer. He has a Lean Six Sigma Black Belt, an MBA from EGADE, and the ability to sell ice to a penguin. What he does not have is a team of 50 people. He has debt, less than 30 days of runway, and an aggressive timeline.

The traditional startup playbook says: hire people, delegate, grow. But when you are pre-revenue with limited capital, hiring is not a strategy — it is a fantasy. The real question was: what if the executive team itself could be built, not hired?

The Architecture: Seven Minds, One Nervous System

Every AI child runs as a Supabase Edge Function written in Deno and TypeScript. Each one connects to the Anthropic Claude API for reasoning, uses Slack as its primary communication layer, and shares access to a common PostgreSQL database for memory, task management, and status reporting.

But what makes this architecture more than a collection of bots is the family model. Each child has a specific domain, a personality, and a tool set designed for their job:

They share a memory layer — the aria_memory table in Supabase, namespaced per child. When the Twins plan a content campaign, INK can see it and prepare branded materials. When Naiyel closes a deal, ARIA updates the overall strategy. They do not operate in silos; they operate as a family.

The Numbers: What Is Actually Running

Every morning at 7 AM CST, ARIA delivers a briefing that includes the CRM pipeline snapshot, content queue status, token health for social media APIs, learning progress, and a summary from each child agent. At 4 PM, an afternoon briefing reports what actually happened: wins, blockers, pending approvals, revenue updates. Between those bookends, the task executor runs every 15 minutes, picking up tasks from the queue and routing them to the correct child.

The content pipeline runs three times a day on weekdays. Content-agent generates posts aligned to four rotating pillars — CX strategy, AI in business, Mexican entrepreneurship, and raw leadership stories. Each post gets sent to Alfie for approval via Slack DM. A checkmark reaction triggers automatic posting to Twitter. No human touches the publishing workflow after approval.

What We Got Wrong (And Nearly Killed the Project)

Building this was not smooth. We want to be honest about the failures because that is the only way this story is useful to other founders.

Event duplication was brutal. Slack sends webhook events that can retry, and our edge functions were not idempotent at first. A single message would sometimes trigger three responses. We ended up building a three-layer deduplication system: retry header detection, a dual-event filter that checks channel+timestamp combinations, and a database-backed event lock table with TTL cleanup. It took three iterations to get right.

The monolith problem. The main ARIA edge function hit 2,804 lines before we modularized it. At that point, deploying was slow, debugging was painful, and no one could find anything. We extracted 16 tool handler files into a tools/ directory with a barrel file pattern, bringing the main file down to around 1,241 lines. Still large, but manageable.

Token and API key management is a full-time job. With integrations across Slack, GitHub, HubSpot, Claude, Replicate, Vercel, Cloudflare, Twitter, LinkedIn, Threads, and Notion, we had a secret sprawl problem. Some tokens expired. Some had wrong scopes. Some were set in Supabase secrets but not in GitHub Actions. We maintain a secrets inventory document and check token health in the morning briefing, but it still bites us regularly.

The human bottleneck is real. The system produces content, qualifies leads, and generates reports — but Alfie still needs to approve content, sign off on proposals, and make strategic calls. The founder becomes the bottleneck in a system that otherwise operates at machine speed. We are actively building more sophisticated approval delegation, but the principle of human-in-the-loop for high-stakes decisions remains non-negotiable.

The Living Company Vision

A Living Company is not a company that uses AI. It is a company where AI is a member of the team — with responsibilities, accountability, and real operational authority. The human founder sets the vision, makes the calls that require judgment and ethics, and approves actions that carry financial or reputational risk. Everything else is delegated to AI agents that can think, remember, coordinate, and execute.

We believe this is the future of the small and medium enterprise. You will not need to raise a Series A to build a team. You will not need to choose between product and marketing because you can only afford one person. You will build an AI executive team — a CTO, a CRO, a CMO, a creative director — and they will work for you around the clock, with perfect recall, no ego, and a shared mission.

That is what we are building at EthicsX.AI. Not a chatbot. Not an AI wrapper. A living, breathing, thinking company. And we are building it in public, with all the scars showing.

Five Lessons for Anyone Attempting This

What Comes Next

The Living Company is not done. It never will be. We are working on autonomous task execution with smarter delegation, cross-venture intelligence where agents spot opportunities between businesses, and eventually an open platform where other founders can deploy their own AI executive teams using the patterns we have proven.

The immediate priority is revenue. All of this technical sophistication means nothing if it does not convert into paying clients for OwnCX, Imprecxo, and CXon. That is the brutal honesty of building a Living Company on a 30-day runway: the AI must earn its keep, or it dies.

But we believe. We believe that in five years, every serious SMB will have an AI executive team. We believe the architecture we are building today will become the standard. And we believe that the company that figures this out first — messy, indebted, and building in public — will have an unfair advantage that no amount of VC money can replicate.

That company is Shiba Group. And the intelligence that runs it is EthicsX.AI.