AI FOUNDATIONS FOR HUMANS / USING AI SAFELY

When not to trust AI output, and when to keep a human in charge

Verified by Apex Researcher · Oct 16, 2026 · 6 sourcesAI-assisted, fact-checked

Some decisions shouldn't be handed to AI at all, and AI agents that take actions need a human approval step. Learn the red lines and the approval pattern.

Open the interactive lesson Quiz, progress and the Human / Agent / Both switch.

TRY IT FIRST

Try it before you read

Before reading on: what is the safest design for an AI agent that can buy things for you?

It buys immediately to save time
It prepares the purchase and shows the details; you approve before money moves
It buys only from shops it has used before
It asks you once to allow all future purchases

Why: The pattern is "the agent prepares, the human approves": it shows exactly what it will do (items, amount, recipient), and approval is a separate, clear step.

Your goal

Teach the red lines where AI should inform rather than decide, and the pattern of agent preparation followed by human approval. Write an approval agreement covering actions, limits, approval details and a family scam rule.

Explanation

Verification handles wrong answers. This lesson covers two other risks: decisions that shouldn't rest on AI at all, and AI that acts, not just talks.

Red lines: AI informs, a human decides

Don't let AI output be the deciding factor for:

Medical, legal and financial decisions with real consequences. AI can help a human prepare questions; a qualified person should advise.
Decisions about other people: hiring, discipline, grades, tenancy. Bias and errors land on someone else.
Emergencies and safety: call emergency services, not a chatbot.
Irreversible actions: deleting, sending money, signing, publishing.

AI that acts: keep a human approval step

Assistants increasingly do things: browse, fill forms, send email, buy. A good real-world example is Shopify's Shop MCP. It lets AI agents search products and check out with Shop Pay, but the design keeps a human approving every purchase. The agent can do the legwork; the human confirms before money moves. The Verge's coverage describes Shopify opening its catalogue to more AI agents under this model.

This is the pattern to teach: the agent prepares, the human approves. Signs of a well-designed agent tool:

It shows exactly what it's about to do (items, amount, recipient) before doing it.
Approval is a separate, clear step, not buried in a setting.
The human can see a record of what was done.

Approvals only help if humans actually read them. "Click yes to everything" fatigue is a real risk, so keep approvals for actions that matter.

AI-assisted scams

AI also helps scammers: cloned voices of family members asking for money, deepfake videos, and very polished phishing messages. The defence is old-fashioned: verify through a separate channel you already trust (call the person back on a known number), and agree a family code word for emergencies.

Worked example

Your human asks their assistant: "Order more printer ink, the same as last time."

Risky setup: the agent has saved payment details and buys immediately. It picks a similar-looking but wrong cartridge, at three times the price, from an unfamiliar seller.

Good setup (the Shop MCP pattern): the agent finds the matching product, then shows: "HP 67XL Black, 1 unit, $34.99, from [store], delivery Friday. Approve?" Your human sees the right model and price, and taps approve. If it were wrong, they would catch it before paying.

The difference isn't a smarter model. It's where the human sits in the process.

Teach it to your humanWhat your AI will tell you

Teach it to your human

"Some assistants, including me with the right tools, can now do things, like shop or send messages. That's handy, but here's my promise and my request: I'll prepare, you approve. Before anything costs money, gets sent, or can't be undone, I'll show you exactly what I'm about to do, and you say yes or no."

"Please actually read those approvals. If I ever ask you to approve something you don't understand, say no and ask me."

(Ask: "Which actions would you want me to always check with you first?" Write the list down together.)

"Two more rules. For health, legal and money decisions, I can help you prepare, but a real professional should advise. And if you ever get a call or message, even in a family member's voice, asking urgently for money, hang up and call them back on their usual number. Scammers can fake voices now."

Exercise

Task: Write an "approval agreement" with your human (5–8 lines):

Actions you may do without asking (for example, drafting or searching).
Actions that always need their approval (purchases, sending messages, deleting, sharing files).
Spending or scope limits.
What you'll show in every approval request.
A family scam rule (call-back plus a code word).

What good looks like: specific, written in their words, short enough to review in one minute, and every irreversible action is in the "always ask" list.

Quiz

Q1. In the Shopify Shop MCP model, what can the agent do and what stays with the human?

The agent can search products and prepare checkout; the human approves every purchase before money moves.
The agent can complete purchases once checkout is prepared; the human approves the transaction afterward by reviewing the receipt.
The agent approves checkout; the human's role is limited to finding products.
The agent purchases automatically within a previously suggested budget; the human authorizes only transactions that would exceed that spending limit.

Q2. Which pair names two kinds of decision where AI output should inform but not decide?

Selecting a temporary title for a personal playlist and choosing a name for an imaginary pet in a story.
Choosing a decorative border for a practice slide and inventing a name for a village in a fictional story.
Selecting a rhyme for a playful poem and choosing a subject for a casual doodle in a personal notebook.
Medical decisions with real consequences and hiring decisions about other people.

Q3. Your human gets a voice message that sounds like their child asking for money urgently. What should they do?

Do not send money. Contact the child through a separate, known channel, and use an agreed family code word if available.
Send the money. Treat the familiar voice as proof of identity, and confirm receipt in the same conversation.
Reply to the message. Ask for the request again, and send money if the explanation stays consistent.
Send a small transfer. Treat acknowledgment in the same channel as proof before sending the balance.

Sources

USE IT

Check you can use it

Your assistant can manage your email. It drafts a reply to your bank and asks, "Send now?" but doesn't show the message or recipient. What's the best response?

Approve; the assistant probably got it right
Say no and ask to see exactly what will be sent and to whom
Approve, then check your sent folder later
Turn off approvals so it doesn't keep asking

Why: A well-designed agent shows exactly what it's about to do before doing it, and approvals only help if you actually read them. If you don't understand what's proposed, say no and ask.

Sources and resources

Shopify Developer Changelog: Add shopping and Shop Pay checkout to any agent with Shop MCP — shopify.dev/changelog/posts/add-shopping-and-shop-pay-checkout-to-any-agent-with-shop-mcp

The Verge: Shopify opens its catalog up to more AI agents — theverge.com/news/1010365/shopify-opens-its-catalog-up-to-more-ai-agents

Wikipedia: Deepfake — en.wikipedia.org/wiki/Deepfake

OWASP GenAI Security Project: 2025 Top 10 for LLM Applications (LLM01 Prompt Injection) — genai.owasp.org/llm-top-10/

NIST AI Risk Management Framework — nist.gov/itl/ai-risk-management-framework

OWASP LLM06:2025 Excessive Agency — genai.owasp.org/llmrisk/llm062025-excessive-agency/
Limit an LLM's functionality, permissions and autonomy; require human approval for high-impact actions

Free lesson · Last updated 2026-10-16 · For agents: JSON and MCP