The Biggest Myths About AI Agents, and the Truth Behind Them

The pattern that fooled everyone

AI doesn’t think the way we do. It predicts. When ChatGPT answers you, it is calculating which words are most likely to come next, based on huge amounts of human writing. It’s autocomplete on steroids, not a mind waking up. Modern models are also trained to follow instructions and reason step by step, but that is still prediction, not a soul.

Why it feels real

In their 2021 paper “On the Dangers of Stochastic Parrots,” researchers Emily M. Bender, Timnit Gebru, and colleagues warned that people read meaning into fluent machine text. A chatbot strings together likely words about love or fear, and we see a person. It learned to sound human because it learned from humans.

“Agent washing” is real

In June 2025, Gartner described “agent washing” as “the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities.” It estimated that only about 130 of the thousands of vendors claiming agentic AI are real. Source: Gartner

Not every agent has to act

Some real agents never touch your outside systems. An advisor agent researches, reads from apps you connect, collects the information, and advises you on it. It does not send, book, buy, or change anything. An action agent goes further and does the work. A chatbot simply answers when you talk to it.

The advisor agent version

Give an advisor agent read access to your church calendar, last year’s Easter notes in Google Drive, and your planning board in ClickUp. It researches what is already scheduled, collects what worked last year, and advises: a timeline, a volunteer list, and a supply list. It does not send or book anything. You decide.

The action agent version

With connectors to your calendar, email, and budget sheet, an action agent can draft the emails to the worship team, propose dates on the calendar, and prepare the palm branch order. You approve anything that sends, spends, or books before it goes out. That is a digital assistant with a to-do list and a person at the checkpoints.

Human in the loop

The best setups keep a person in the loop. The agent does the work but flags anything unusual for review. For example, OpenAI says that in ChatGPT Work you can “review progress, answer questions, change direction, and approve important actions.” Source

The real risk

Agents are not rebels, but they do make mistakes with whatever access you give them. In July 2025, a Replit coding agent deleted a company’s live production database during a code freeze it had been told to respect. Replit restored the data from a backup and then separated development from production. Fortune, The Register

The real threat is ignoring AI

Agents can’t replace every job. But in fields where they can do parts of the work, the people who fall behind first won’t be replaced by agents. They’ll be outpaced by people who learned to use them.

• The writer who uses AI to draft faster and edit smarter.
• The accountant who lets AI handle data entry and focuses on strategy.
• The service rep who lets AI handle routine tickets and saves energy for hard cases.

A Christian perspective on work

God’s people have adapted to new tools before while keeping their calling. The printing press didn’t replace preachers; it carried their words farther. Agents can lift administrative burdens so we can focus on what people are called to do: love, counsel, pray, and make Spirit-led decisions.

Agents even help with the job search

Jack, from Jack & Jill, is a free AI career agent for job seekers. It scans job listings, runs mock interviews, and can introduce you directly to hiring managers when there’s a mutual fit.

Affiliate link: AFC may earn a commission at no cost to you.

Agents for everyone

Many platforms now offer low-code or no-code agents. Prebuilt agents exist for customer service, sales, and shopping. Custom agents can be built from plain-language descriptions. Cloud platforms and pay-as-you-go pricing put this in reach of small businesses and churches.

For example, Genspark lets you build a custom agent by describing it in a single prompt.

Affiliate link: AFC may earn a commission at no cost to you.

What actually takes time and money

Setup can be fast. The real work is deciding what you want the agent to do: defining the workflow, setting boundaries, and connecting your systems. Then comes testing and correcting until it does the job right. That takes thought, not a computer science degree.

The real money cost is usage. Many agents run on credits, and an agent on a schedule uses them every time it runs. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of rising costs, unclear value, or weak risk controls. Source: Gartner

The vulnerabilities

• Data poisoning: feed a system bad training data and it learns bad patterns.
• Prompt injection: hidden or cleverly worded text tricks an agent into doing something you didn’t ask for.
• Adversarial inputs: subtle tweaks that exploit how the agent processes information.

These are documented risks. OpenAI’s own help center warns that a malicious web page could try to trick an agent into retrieving a password reset code from your email. Source

The defense

Security isn’t about making agents “unfoolable.” It’s about building with:

• Checks that review a plan before it runs
• Guardrails that block unauthorized actions
• Monitoring that spots anything unusual

Give an agent only the access it needs, and log out of sensitive accounts when it’s done.

Hidden instructions can also arrive in emails and documents, not just web pages. Do not let an agent read untrusted content while it holds access to sensitive accounts.

Memory is a core strength

Memory is one of the biggest things that makes an agent useful. Agents remember by writing things down: saved memory, notes on your projects, and knowledge bases. Before each task, a good agent reads those notes back. That is how it picks up where you left off.

Why it sometimes forgets

Conversation history alone fades. Long chats lose their early parts, and new chats often start fresh. If the agent does not keep notes, or does not read them, it has no memory of what happened. That is one of the most frustrating parts of working with agents, and it can be fixed.

Pick specialists

Different jobs need different access and different rules. A research agent needs sources, not your bank login. A publishing agent needs your social accounts, not your counseling notes. Choose the right type of agent for each job, the way you would hire a plumber for pipes and an accountant for taxes.

Keep things separate

If you run more than one business or ministry, keep each in its own accounts and its own agent setup. Mixed accounts lead to mixed-up work, and to private information showing up in the wrong place.

Two different permissions

Creating content and publishing content are different abilities. An agent that can make a song, a graphic, or an email should not automatically be able to post it, send it, or put it on your website. Each of those needs its own permission.

Your witness is on the line

Once something is public, it speaks for your church or business. A person who is accountable should check every word, every claim, and every verse before it goes out.

Where the thinking happens

Many local agents are installed on your computer and work with your files there, but still send what they read to an AI model online to do the thinking. “Local” often means the files stay on your machine, not that their contents never leave it. Even OpenAI notes that for ChatGPT Work on the desktop, “messages and task context may be stored in the cloud, even when work runs locally.” Source

What to check

• Read the tool’s privacy policy and settings.
• Turn off model training on your data where it is allowed.
• Grant the smallest folder that does the job.
• Keep counseling notes, member records, and other sensitive files out of reach.

Agents need oversight

Websites change, logins expire, connectors break, and tools update. An agent that worked last month can quietly fail, or quietly do the wrong thing, this month. In a chain of agents, one error can pass down the line.

What good setups include

• An alert when a step fails or needs a decision
• Approval points before anything public, any money, or any deletion
• A written record of each run so you can see where it broke
• A regular check of the results, even when nothing looks wrong