12 Mistakes Beginners Make With AI Agents, and the Quick Fix for Each One

Most people's first go with an AI agent ends one of two ways. Either it "did nothing useful", because the request was a one-line wish and the result was a one-line answer, or it did something alarming, like emailing a customer a draft that was never meant to go out. Both lead to the same conclusion, that agents are hype, and both are avoidable.

The pattern is not limited to beginners at home. Gartner predicted in June 2025 that more than 40% of agentic AI projects in companies would be cancelled by the end of 2027, blaming rising costs, unclear value and poor risk controls. Those are the same three mistakes, at a bigger scale. This guide lists the 12 mistakes beginners make with AI agents, what each one looks like, and the fix, which in every case takes under five minutes. It was checked on 8 October 2026.

How we chose the 12

1. The builders' own warnings. Anthropic and OpenAI both publish guidance on where agents fail, and the first eight mistakes are the ones they describe.
2. Gartner's three reasons. Cost, unclear value and risk are the reasons business projects fail, and mistakes 9 to 12 are their household versions.
3. What we see in real runs. Every example is a task we have watched go wrong and then fixed.
4. Fixable in minutes. A mistake that needs a course to correct did not make the list.

The 12 mistakes at a glance

Mistakes 1 to 4: how you ask

1. The one-line wish

"Write me a newsletter." The agent cannot know your trade, your customers or what happened this month, so it writes a newsletter for nobody. The fix is the sentence most people skip: who it is for, what it should achieve and what it must include. "Write a 200-word October newsletter for my plumbing customers in Leeds about getting boilers serviced before winter, mention the six slots I have free, friendly not salesy." Our guide on how to write prompts for AI agents has 30 examples built this way.

2. No definition of done

An agent keeps working until it believes the goal is met. If you have not said what done looks like, it decides for you, and either stops after a quick answer or wanders. End the request with it: "Done means a spreadsheet with one row per receipt, totals by month, and a list of any receipt you could not read."

3. Five jobs in one message

"Sort my receipts, chase the Patel invoice, find a cheaper phone contract and draft a job advert." It will do the first well, the second partly and lose the rest. Give one job per task, or write a numbered list and ask it to report after each item. On Jobbit you can queue messages while it works, so the next job is waiting when the first is done.

4. No example of your style

Agents write like the average of the internet unless you show them otherwise. Paste one email you were happy with and say "write like this". Save it in your instructions and you never need to say it again. Our guide to making AI writing sound human covers the rest.

Mistakes 5 to 8: how you check

5. Trusting a number you did not check

The model inside an agent predicts likely text, not guaranteed truth. OpenAI's own research paper, published in September 2025, explains that models are trained and tested in ways that reward a guess over an admission of not knowing. A figure in a quote, a tax threshold, a competitor's price: check any number you will act on, and ask the agent to show where it came from. Our guide on how AI agents work explains why this happens.

6. Never asking where a fact came from

The fix is three words in the request: "cite your sources". A fact with a link you can click is a fact you can check in ten seconds. A fact without one is a guess until proven otherwise. The best agents include sources without being asked; ask anyway.

7. Not checking the first scheduled run

Scheduled tasks are the quiet win of agents: a report that writes itself every Monday, a reminder that chases itself. They are also where a mistake repeats on a timetable. Read the first two runs in full. After that, skim.

8. Giving up after one try

The first result is a draft, not a verdict. Reply with what was wrong, as you would to a new assistant: "too long, wrong tone, you missed the deadline". The second result is usually right, and the correction can be saved as an instruction so the third task starts right.

Mistakes 9 to 12: safety and money

9. Pasting passwords or personal details

Never type a password, a card number or another person's details into a chat. Not because every service is careless, but because a chat is the wrong place for a secret: it is logged, it may be used for training, and you lose track of where it went. Agents that need a login to do a job should hold it in a secrets store that the agent uses without showing it. On Jobbit that is the Projects secrets feature. Our guide on whether AI is safe to use lists the seven things never to paste.

10. Letting it act on real accounts without approval

The agent that emailed the wrong customer was not malicious; it was told to "handle the invoices" and did. Keep approval on for anything that sends a message, spends money or deletes something, at least until you have seen it do the job right a few times. A good agent stops at those points by itself; make sure yours does before you trust it with the real inbox.

11. Ignoring credits until they run out

Agents are billed in credits, and a task that loops, retries a blocked website or reads fifty pages will spend more than a one-line question. The fix is to set the limit in the request: "spend no more than ten minutes on this" or "check three sites, not thirty". Watch the meter the first few times and you will learn what a task costs. Our guide to how much AI agents cost explains credits across the main products.

12. Using a chatbot for a job that needs tools

If you ask a plain chatbot to "book the table", it will tell you how to book a table. It has no hands. The job needs an agent with a browser, and the fix is partly choosing the right product and partly saying which tool to use: "use the browser to check availability on the restaurant's own site". Our guide to the difference between a chatbot and an AI agent shows when each is enough.

The one-page checklist

Keep this by the keyboard for the first month.

• Before you ask: who is it for, what does good look like, what must it not do, and what does done mean?
• One job per task. Queue the rest.
• Show it your style once, then save it as an instruction.
• Add "cite your sources" to anything with facts or numbers.
• Check every number you will act on.
• Approval on for money, messages and deletions.
• Secrets in the secrets store, never in chat.
• Set a budget in the request for anything open-ended.
• Read the first two runs of any schedule.
• Reply with what was wrong before you give up.

How an AI agent helps you avoid them

ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse all let you make and recover from these mistakes. An agent designed around supervision makes them harder to make in the first place.

• A plan before it runs. Jobbit's plan mode shows the steps and waits for your yes, which catches mistakes 1 to 3 before any credits are spent.
• Approval at the risky points. It stops before spending, sending or deleting, which is mistake 10 handled by default.
• A browser you can watch. The agent browser shows every page, and you can take the controls when a site needs a human.
• Instructions and secrets that persist. Projects hold your style, your rules and your logins so mistakes 4 and 9 are solved once.
• A meter you can see. The credit and context meters show what a task is costing while it runs.
• A person when the job needs one. If the task turns out to need a plumber or a designer, the agent briefs a vetted professional from Jobbit's human network, with payment held in escrow. The Free plan gives 100 credits a day with no card, and Go is £7 a month for one small app kept live on your own domain.

Your plan for this week

• Day 1: rewrite your last failed request using the four questions in the checklist and run it again.
• Day 2: paste one email you were proud of into your saved instructions as the style example.
• Day 3: move any password you have ever typed into a chat into a secrets store, and change it.
• Day 4: give the agent one task with a budget and watch the meter.
• Day 5: set up one scheduled task, and put "check run one and run two" in your diary. Then try the next task on Jobbit with the plan shown first.

Frequently asked questions

Why does my AI agent give bad results?

Usually because the request was a one-line wish with no audience, no definition of done and no example of your style. Rewrite it with those three things and the result changes completely.

How do I write a good prompt for an AI agent?

Say who it is for, what good looks like, what it must not do and what done means. Add "cite your sources" for anything with facts. One job per task.

Can I trust the numbers an AI agent gives me?

Not without a source. Models are trained in ways that reward guessing, so ask where each number came from and check any figure you will act on, especially money, tax and prices.

Is it safe to give an AI agent my passwords?

Not in chat. Use a secrets store that the agent can use without showing the password, such as Jobbit's Projects secrets, and never paste card numbers or other people's details anywhere.

How do I stop an AI agent spending money or sending emails by mistake?

Keep approval switched on for payments, messages and deletions. A well-designed agent stops at those points by default; check that yours does before trusting it with real accounts.

Why did the agent use so many credits?

Open-ended tasks loop, retry blocked websites and read many pages. Set a limit in the request, such as "check three sites" or "spend ten minutes", and watch the meter the first few times.

What should I try first with an AI agent?

One real task from your week, written with the checklist above. Start on Jobbit: it shows the plan first, asks before it acts, and gives 100 free credits a day with no card. Go is £7 a month if you want an app of your own kept live.