Two weeks ago I wrote about thinking of AI as an alien intelligence instead of a person. This time I want to get into how you actually manage it.
Picture hiring a brilliant alien. It knows every language on Earth and has read every book ever written, but it has no common sense and no feel for how your office works. How would you manage it?
You wouldn’t give it a vague goal like “improve our marketing” and walk away. It would come back with 500 pages of perfect grammar that have nothing to do with what you needed.
That’s the trap a lot of leaders fall into with AI. We manage it like a human employee, relying on unspoken expectations and background we assume everyone shares. When the result is off we tend to blame the technology, though most of the time the problem is how we asked.
Managing AI well means swapping that unspoken supervision for clear boundaries. Spell out what it should work from and what it must not do, and check its output before it touches anything live. It’ll make you a clearer communicator with your people too.
Why AI reads differently from people
People read in order. You go word by word and sentence by sentence, building up the story in your head as you go.
A language model takes in everything you give it at once. It turns words into numbers that capture how ideas relate to each other, a bit like plotting every word on a giant map where related ideas sit close together. Technical folks call these vector embeddings.
To the AI, your document is that map. In practice that means two things:
- It measures closeness. When you give it your mission statement, it works out how closely its draft lines up with the words, rules and examples you provided. It has no sense of why the mission matters.
- It starts fresh every time. It has no memory of past conversations. Each time it writes, it weighs everything in front of it from scratch.
So long, chatty prompts work less well than structured ones. Clear labels, fixed formats and tightly scoped background get you more reliable results.
How to write instructions AI can follow
The biggest mistake I see from non-technical teams is writing a prompt like a search query or a quick text to a coworker. Here’s the framework I use to turn a vague request into a clear task:
- Give it a role. Instead of “Write a donor email,” try “You’re an experienced nonprofit development officer writing to a major donor who hasn’t given in 12 months.”
- Give it the background. It has no memory of your organization. Instead of “Use our usual tone,” attach your brand guidelines and last year’s successful gala appeal.
- Set the format. Instead of “Keep it short,” say “Under 200 words, in three paragraphs, with one clear ask.”
- Say what to avoid. Instead of “Make it sound professional,” say “Don’t use buzzwords like ‘synergy’ or ‘transformative,’ and don’t make it a hard sales pitch.”
Requests written this way cut down a lot on made-up details and generic copy. When the draft comes back, read it against the role, format and limits you set. If it missed one, tell it which one and ask again.
For leaders: keep AI in a safe workspace
If you hired an alien genius, you wouldn’t hand it the master key to the bank vault on day one. You’d give it a secure desk and copies of the files, and you’d check its work before it touched anything real.
Plenty of organizations still give AI tools direct, open access to their live CRM, financial records and public channels. The setup I’d look for is a walled garden:
- Read-only access to live data. The AI can read and analyze your data, but it can’t overwrite master records without permission.
- A sandbox for anything it builds. When it writes code or processes complex data, it works in a temporary, separate space. If something breaks, it breaks there.
- A person signs off. The AI drafts the report or prepares the CRM updates, then stops and waits for someone to approve.
I covered guardrails in more detail back in August. The short version is that trust comes from a setup where mistakes get caught before they reach anything live.
Questions to ask your team:
- Which AI tools can change our live systems right now?
- Do we test new AI work on copies first?
- Who approves AI-prepared changes before they go live or go out?
- When AI gives us a bad result, do we look at how we asked?
I’ve built that walled-garden setup into Cosly, the AI staff team I’m working on for small nonprofits.
