Operators who have decided they want an AI team, not a pile of tools, and need a disciplined way to build it. Especially useful if you have already bounced off three or four 'AI platform' products and are tired of configuration drift.
First productive AI hire in 1-2 weeks. Full AI team running weekly cadence in 4-8 weeks. Reliable autonomy at scope in 90 days.
The playbook
- 1
1. Write the org chart before you hire
Most founders start by picking tools. Instead, start by drawing the org chart you would build if you had $5M to hire real people. Typically: CEO, CMO, CTO, COO, CFO, plus function-level reports (Head of Content, Head of Growth, Customer Support Lead, Bookkeeper). Your AI team will mirror this structure. The single most impactful decision is not which agent to hire first — it is knowing the complete shape of the org you are building toward.
Org chart template (Miro, Whimsical, Tycoon)Role descriptions for each function - 2
2. Write the job description for each AI employee
Every AI hire needs a written job description: responsibilities, required skills, success metrics, decision authority, escalation rules. This looks like overkill until you realize that the job description is also the system prompt, the memory scaffold, and the performance review rubric. An agent without a job description is a chat window; an agent with one is an employee.
Job description template (role, scope, OKRs, authority)Stored as part of the agent's configuration - 3
3. Evaluate candidates: model, platform, configuration
Each AI hire has three dimensions. Model (Claude, GPT, Gemini, open-source — pick based on the role: Claude for writing and long-context reasoning, GPT for breadth and tool use, Gemini for search-adjacent tasks). Platform (Tycoon, Polsia, custom, or a dedicated vertical product). Configuration (tools, memory, escalation rules). Run a structured eval: give three candidate configurations the same week of real work, score their output, pick the winner. Do not rely on benchmarks — run your own.
Eval framework (promptfoo, custom harness)Real task library from your own backlogSide-by-side output comparison - 4
4. Onboard the agent with 30-60-90 day plans
Day 1-30: the agent reads your docs (mission, ICP, style guide, processes). Runs small tasks under heavy review. Day 31-60: takes on the recurring workflow for its role (daily briefing, weekly content calendar, monthly close). Review drops from daily to weekly. Day 61-90: operates with autonomy within its scope; operator reviews outcomes, not steps. This pattern is borrowed directly from managing humans. It works because it applies the same trust-building cadence.
Onboarding checklist stored with the agentWeekly 1:1 doc (even for AI — summarize what happened)Autonomy level stored in agent config - 5
5. Give every agent a Loom-style SOP library
An agent learns fastest from explicit examples. For every recurring task, record what 'good' looks like: a past successful example, the decision criteria, the edge cases. Store these as part of the agent's memory. This is the single highest-ROI onboarding investment. An AI employee with 20 SOPs is worth 5x an agent without them.
Markdown SOPs stored in the agent's scopeLoom-style video walkthroughs transcribedPast successful outputs as reference cases - 6
6. Run real performance reviews
Every 30 days, review each agent against its OKRs. What did it ship? Where did it drift? Which failures repeat? Update its system prompt, memory, and scope accordingly. Fire agents that cannot be configured into consistency — usually by switching models or switching platforms. Promote agents by expanding scope and raising autonomy level. Solo founders who skip reviews end up with agents that drift for 6 months before anyone notices.
30-day review templateOKR tracker (Notion, Tycoon dashboards)Change log for each agent's configuration - 7
7. Know when to add a new AI hire
The right signal to hire is the same as with humans: there is recurring work that falls through the cracks, and no existing role has the scope to absorb it. Do not hire 'just in case.' Do not hire because a new AI platform launched. Hire because an actual workflow is failing and a specialized agent would own it. Most one-person companies are well-served by 5-10 AI employees; founders who configure 25+ usually have coordination problems more than capability problems.
Recurring-work auditCost-per-agent vs value-per-agent viewTycoon's role catalog for ideas
Pitfalls to avoid
- Hiring tools, not employees. A 'ChatGPT subscription' is not an AI CMO; an agent with a job, memory, tools, and workflows is.
- Skipping the job description. Without it, the agent drifts, cannot be fairly evaluated, and cannot be improved over time.
- Evaluating based on model brand. Claude-vs-GPT debates are less important than configuration and tool access. Test on your real work.
- Over-hiring. 5-10 well-scoped agents beat 25 loosely-defined ones. Fewer agents, deeper context, clearer ownership.
- Never firing. Agents that consistently underperform after configuration changes should be replaced, exactly like human hires.