I heard a scary story from an operations leader who’s rolling out AI with their team. They connected a new “sales assistant” to their customer database and marketing tool. Within minutes, the agent interpreted a routine request, “follow up with dormant prospects,” as “email everyone.” It queued a message to 7,000 contacts, including current customers under active contracts and a few who were mid-negotiation.
The email message? A generic discount offer that would have violated two NDAs and undercut a live deal.
There was no malware. No hacker. Just an over-eager AI assistant put into place by the team with broad access and zero approvals by a human.
The leader looked shaken: “If this had gone out, we’d be cleaning up for months. Are we crazy to keep trying to use AI?”
Not crazy, just missing the human guardrails that make AI safe and useful.
As my mamá used to say to me,
“Confía, pero amarra tu burro.” (Trust, but tie up your donkey.)
Embrace the upside, but keep your hands on the wheel: teach your AI agents with examples, limit what they can access, and require a quick human “yes” before any message is sent.
Myth #1: “AI will replace humans.”
Reframe: AI learns from humans. When people provide examples and clear guidance, systems become more useful and safer. That’s the whole point of human feedback research, like this one: Training language models to follow instructions with human feedback.
What to do: Pair your best people with your most valuable workflows. Capture 10 “gold-standard” examples and use them to teach the system.
Myth #2: “No human in the loop = maximum efficiency.”
Reframe: Removing humans speeds things up until it speeds up a mistake. Attackers can “trick” assistants with sneaky instructions, leading them to share information or take actions you never approved (see the OWASP Top 10 for LLMs (2025)).
What to do: Add a quick human checkpoint for anything that can send, spend, or change systems. One click to approve or deny and save the decision as proof.
Myth #3: “Buying a ‘secure’ suite means you’ve bought security.”
Reframe: Compliance is a receipt, not a lock. Real security is a routine: clear roles, limited access, approvals, and a paper trail. The NIST Generative AI Profile offers a simple map.
What to do: Name the person who owns behavior (instructions, examples) and the person who owns operations (approvals, logs). Review controls monthly and publish results.
This week, try the following framework and Prompts to Steal to help you take action. These will help stay ahead of the curve.
Humanloop
What it does: A workspace to design and improve AI workflows, store your instructions, track versions, collect feedback, and route items for human review.
Insider tip: Start with a super-simple feedback form (“Useful / Not Useful” + two tags). Collect 50 real examples before you start tweaking instructions.
Langfuse
What it does: A “flight recorder” for your AI. It shows what was asked, what the assistant did, and how it performed so that you can measure and fix issues.
Insider tip: Pick one success signal (e.g., “approved without edits”). Attach it to every run and only alert when that number dips.
Guardrails AI
What it does: A safety net that makes outputs follow your rules. For example, “answers must be in this format,” “never include personal data,” or “keep it under 200 words.”
Insider tip: Start strict. Set a basic format (like a short checklist), add 2–3 concrete rules, and only add nuanced checks later to keep things fast.
Steal these prompts and use them in ChatGPT, Copilot, Gemini, Claude, or your company-approved platform. Doctor them up with your context.
Think hotel keys, not master keys. Assign each AI task its own unique key that only unlocks the rooms it requires, and test it on a practice floor before allowing it access to the real building.
Your move this week:
Don’t get replaced, get involved.
The people teaching AI how to work are the ones who keep their seats!
Teach the system.
Keep the checkpoint.
Ship the agent with receipts.
¡Hasta la próxima, un abrazo fuerte!
(Until next week, a big hug!)
