Running a company with AI agents

The problem

In January 2025 I founded AbilityArc to build durable-skills courses for IT students in colleges and career-tech programs. A bootstrapped company needs the work of a sales team, a marketing team, curriculum developers, platform administrators and a back office. It can’t afford to hire them. My question: could AI agents do much of that work safely, with me as the operator who sets the rules and approves what matters?

What I did

I’m not a software engineer. I designed these systems, set the guardrails, tested and approved the releases, and directed the work. Claude Code (and, for one product, a contract developer) wrote most of the code.

Designed an AI org chart. I built an operating model in Claude Code with 37 agent roles across sales, marketing, curriculum, web and platform, software and administration. It also has a six-persona board of advisors that argues a strategic question from different angles before a lead agent sums it up. About 50 scheduled routines run research, digests, inbox checks and weekly calendars.

Set the guardrails. Every agent works under rules I wrote:

• a human approves anything outbound or live
• each agent has narrow tool permissions
• shared state means two agents never contact the same person
• reviews happen in batches instead of interrupting me for every item

I also manage cost. When I found that two hourly routines drove 91% of automation cost, I cut scheduled runs by about 85%.

Built an AI research pipeline. A research agent works through a national queue of colleges, tech centers and districts. For each one it finds IT and CTE instructors and leaders, qualifies them and verifies emails. A separate step loads them into the CRM with dedupe rules. It has researched 4,941 institutions so far and is on pace to finish in about two months the work that would take one full-time researcher about a year (my estimate).

Automated the inbox, with approval. A real-time agent built on a Cloudflare Worker, Microsoft Graph and Claude reads prospect replies, drafts responses and logs CRM notes. It sends me a one-tap approval card, at about 0.1¢ per email.

Shipped an AI product. I directed the build of Reactor, an AI voice and chat simulator where students practice workplace conversations. It has rubric-based AI grading and passes grades back to the school’s learning platform. I authored its 32 live simulations. When the AI sometimes broke character, I traced it to a hidden prompt rule, removed it, and confirmed with automated tests across all 32 simulations: character breaks went from 6 to 0.

Onboarded people into the system. Two commission sales reps each got their own AI workspace, guarded CRM access and a selling guide. Each was up and running in about a week.

Results (as of October 2026)

• 37 agent roles and ~50 routines running a real business, with human approval on everything that matters.
• 4,941 institutions researched, on pace for about 2 months versus about a year by hand (estimate).
• Character breaks 6 → 0 across 32 live AI simulations.
• ~85% fewer scheduled runs after a cost analysis.
• Paying customers, including a Utah technical college that bought twice and uses the course program-wide.

Why it matters

For AI-enabled GTM and marketing operations: I don’t just use AI tools. I design AI systems with guardrails and run them in production. For fractional work: I can set up research, outreach and reporting agents for a team and leave them with rules they can trust.