Before an AI Agent Goes Live, Test What It Cannot Do
Before launching an AI agent, test prohibited actions, approval boundaries, retries and escalation. Verify both enforced controls and business quality.
Read article →Writing
Thinking on AI implementation, product strategy, technical leadership, and the operational realities that most consultants skip over.
Before launching an AI agent, test prohibited actions, approval boundaries, retries and escalation. Verify both enforced controls and business quality.
Read article →AI decision records must distinguish proposals, discussions, approvals and completed actions, retaining their evidence, conditions and effective dates.
Read article →A first AI workflow should leave reusable tests, approval boundaries and operating history. The second implementation shows what the company retained.
Read article →AI tool renewals need evidence of useful work, acceptable quality and the full cost of support. Usage alone does not justify continued spending.
Read article →Highly capable people spend enormous time re-verifying work a well-built system should already guarantee. Here's what changes when it doesn't.
Read: Trust work vs. judgment work →AI can produce a polished strategy document in minutes. It cannot do the observation, judgment, and tradeoffs that make a strategy real.
Read: The document was never the strategy →Founder-dependent work feels efficient long after it's gotten expensive. Scaling means turning that judgment into a system the team can run without translation.
Read: When the founder becomes the OS →A fractional CTO turns business ambiguity into technical clarity and execution rhythm. AI makes that judgment more valuable, not less.
Read: What a fractional CTO actually does →Every SaaS product that doesn't quite fit gets patched with spreadsheets, manual approvals, and tribal knowledge. AI is changing the economics of closing that gap yourself.
Read: The build-vs-rent question →SMBs face different constraints than enterprises, but they adopt the same playbook. The better approach starts with understanding the actual work, not the tools.
Read: What SMBs get wrong about AI →Token costs are a legitimate question. But they are not the right first question. The real unit of analysis is the workflow — what it costs today and what it would cost if it were better.
Read: Token-level vs workflow-level cost →Autonomy is not the hard problem. The hard problem is the structure around an agent — context, evaluation, guardrails, feedback loops — that lets it operate reliably in a real workflow.
Read: Agents need a harness →Speed is not the same thing as wisdom. A mediocre strategy with AI is still a mediocre strategy. It just gets executed faster.
Read: AI makes mediocrity faster →The mistake many companies make is starting with the tool. The better starting point is the workflow — where time, money, and attention are actually being consumed.
Read: Start with workflow pain →AI changes the leverage equation. The person who creates the most value is not always the deepest specialist — it is the person who knows where to aim the tools in the first place.
Read: The generalist advantage →Sometimes the best first step is a conversation — not a proposal.