What should my AI agent be allowed to do on its own, and what should it never do?
Short answer: Write one page with three columns: what the agent may do alone, what it must ask a person about first, and what it must never touch. An AI agent is software that uses an AI model to carry out tasks on its own, such as reading email or sending invoices, instead of only answering questions. Give it the narrowest access that lets it do its job, require a human yes before anything that sends, spends or deletes, and keep a log and a stop switch. That is the same advice OWASP, a nonprofit that publishes security standards, gives for AI apps.
You are about to let an AI agent loose on real invoices, customer messages or calendars, and you want a short, plain set of rules for what it may and may not do.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
Most guardrail advice is written for engineers. It talks about filters and policy engines. You do not need that to start. You need a rule sheet a human can read, and a setup that makes the rules real instead of hopeful. A good first use is AI agent to chase overdue invoices.
The reason this matters is simple. An agent can take actions, not just give answers. A wrong answer is an annoying message. A wrong action is a sent email, a paid bill or a deleted file. If you are new to the idea, start with what an AI agent is and how to build one, then come back here.
Why agents need rules at all
OWASP's list of risks for AI applications has an entry called Excessive Agency. It names three causes: the agent has too many functions, too much access, or too much freedom to act without anyone checking. Each one maps to a column on your rule sheet.
Anthropic, the company that makes Claude, writes in its guide to building agents that agents bring "higher costs, and the potential for compounding errors," and recommends testing in sandboxes (safe practice copies) with appropriate guardrails. A small error early in a chain of steps can grow by the end. That is why you set limits before the first real run, not after the first mistake.
The three-column rule sheet
Print this and fill it in for each agent you run. One agent, one sheet.
| Column | Meaning | Test for what belongs here |
|---|---|---|
| Allowed | The agent does it alone and tells you in a daily summary. | Easy to undo, no money moves, nothing leaves the company. |
| Ask first | The agent prepares it. A person approves before it happens. | Goes to a customer, spends money, or is hard to undo. |
| Never | The agent has no access and no instruction to do it. | Legal, financial or personal consequences, or you simply do not trust software with it. |
The "never" column works best when it is enforced by access, not by instruction. If the agent has no login to the bank, it cannot move money, however it is prompted. OWASP calls this least privilege: give the agent only the access its task needs. Its own example is an agent that only reads email being given a read-only connection, not one that can also send and delete.
Filled in: invoices
| Allowed | Ask first | Never |
|---|---|---|
| Read incoming invoices and fill in a draft record | Approve a bill for payment | Release a payment from the bank |
| Match an invoice to a purchase order | Pay a bill over a set amount, say $500 (your number) | Change a supplier's bank details |
| Flag a possible duplicate | Send a payment reminder to a customer | Delete or edit a posted accounting entry |
Changing a supplier's bank details is on the never list on purpose. A wrong bank detail sends money somewhere you cannot easily get it back, so a person should always confirm a change by phone. For the wider picture, see accounts payable automation.
Filled in: customer replies
| Allowed | Ask first | Never |
|---|---|---|
| Answer questions from your published opening hours, prices and policies | Offer a refund, discount or exception | Promise a result, a date you have not checked or a legal position |
| Collect details and book a callback | Reply to a complaint or a message that sounds angry | Give medical, legal or financial advice |
| Say it is an automated assistant when asked | Share any customer's personal data | Pretend to be a named human |
More on this use is in AI customer service agents. The key rule is an easy exit: any customer should be able to reach a person in one step.
Filled in: scheduling
| Allowed | Ask first | Never |
|---|---|---|
| Propose times from your free slots | Move or cancel a booked client meeting | Delete calendar events |
| Book a standard appointment type | Book anything outside working hours | See private events marked personal |
| Send confirmations and reminders | Overbook or double-book on purpose | Share your calendar details with outsiders |
The five limits that sit behind the sheet
The sheet is the policy. These five settings make it real.
- Access limits. Give the agent its own account, with the least access that works. Never share your own login. If something goes wrong you can switch off one account.
- Spending limits. If the agent uses a paid AI service, set a monthly cap in that service. If it can pay anything, give it a card with a low limit, or better, no card.
- Approval gates. The "ask first" column becomes a button a person must press. OWASP's advice is to require a person to approve high-impact actions such as sending, posting or deleting.
- A log. Every action goes into a record a person can read: what it did, to what, when, and why. OWASP notes that logging does not prevent bad actions but limits the damage and helps you notice them.
- A stop switch. One named person can turn the agent off in under a minute. Anthropic's guide also mentions stopping conditions such as a maximum number of steps, so a confused agent does not run all night.
Two kinds of guardrail, and why you need both
Some guardrails live outside the AI. Access limits, spending caps and approval buttons are examples. They work even if the AI is confused or tricked.
Others are filters on what the AI reads and writes. Amazon's Bedrock Guardrails, a cloud service for this, offers filters for harmful content, blocked topics, personal data and answers not grounded in your source material. These help, but they check text, not access. Do not rely on a filter where a locked door would do. OWASP makes the same point: enforce permissions in the systems the agent connects to, not by asking the AI to behave.
Common mistakes
- Using your own login. The agent then has all your powers.
- Rules written only in the prompt. A prompt is a request, not a lock.
- No testing on dummy data. Run the agent on a copy first. Anthropic recommends extensive sandbox testing.
- Approving everything on autopilot. If a person clicks yes without reading, the gate is decoration. Keep the approval list short enough to be read.
- Never reviewing the log. Read it daily for the first two weeks, then weekly.
How to start in one afternoon
- Pick one task, such as drafting replies to common emails.
- Write the three columns for it. Aim for five to eight lines in each.
- Create a separate account for the agent with only the access in the first two columns.
- Turn every "ask first" action into a draft a person must approve.
- Run it on dummy or copied data for a week and read the log.
- Go live with one person watching. Widen the "allowed" column only after weeks of clean logs.
NIST, the US standards agency, publishes a voluntary AI Risk Management Framework built around four functions it calls Govern, Map, Measure and Manage. You do not need the full document for a small business. The idea behind it is the same as this page: decide who is responsible, know what the system touches, check how it behaves, and have a way to respond when it goes wrong.
When to get outside help
If the agent touches payments, health or legal records, or customers in a regulated industry, have a developer set up the access limits and have a lawyer read the "never" column. If you are hiring for the build, how to choose an AI agent development company lists questions to ask, and RPA versus AI agents helps you decide whether you need an agent at all. Some jobs are better done by a fixed script, which cannot improvise. The idea of an AI employee is best treated the same way: like a new hire on probation.
How I'd approach this
I treat a new agent like a new hire in their first month. They get their own login, a short list of things they can do alone, and a boss who checks their work. They do not get the company card on day one.
My own rule is that anything which sends, spends or deletes needs a human yes at the start. I loosen that one item at a time, and only when the log shows weeks of good work. I also keep the never column short and firm. If you are not sure whether something belongs there, it does.
If you want a second pair of eyes on a rule sheet before an agent goes live, that is something I do as part of Blur Studio's business work.
Questions people ask
What are AI agent guardrails?
They are the limits that keep an AI agent safe: what it may do alone, what needs a person's approval, what it can never touch, plus access limits, spending caps, a log and a stop switch. See what an AI agent is for the basics.
Can I put guardrails in the prompt only?
Not safely. A prompt is a request the AI can misread or be tricked out of. Limits that live in access, such as a read-only login or a card with a low limit, work even when the AI gets it wrong. OWASP advises enforcing permissions in the connected systems instead of relying on the AI to decide.
What is the worst that can happen without guardrails?
An agent with too much access can send the wrong message to customers, pay the wrong bill, change records or delete files, and a small early mistake can repeat across many steps. Anthropic notes the potential for compounding errors in agent systems, which is why limits come first.
How much human approval is too much?
If a person approves dozens of items a day, they will stop reading. Keep approval for actions that send to a customer, spend money or are hard to undo, and let the agent act alone on easy-to-undo work with a daily summary.
Do I need a technical person to set this up?
You can write the rule sheet yourself. Creating a separate agent account with narrow access, and wiring approval buttons, is easier with a developer. If you want help choosing one, read how to choose an AI agent development company.
Is there an official framework I can follow?
NIST publishes a voluntary AI Risk Management Framework, and OWASP publishes a list of risks for AI applications that includes Excessive Agency. Both are written for larger teams, but the ideas map straight onto the one-page sheet in this article.
More in this topic: Problem solving: AI strategy and adoption (24 guides) →