Pro free for 30 days.No card.Start free →

AI agents and human approval: how to keep control of what they do

What changes when an AI can act instead of just answer, and how a small business can keep control: sort actions by risk, apply least privilege, enforce limits in the system rather than the prompt, keep a log, and use a setup checklist before switching an agent on.

By WorkBOS Content Desk6 min read
A human hand and a robotic hand meeting in a handshake in front of a computer screen (Photo by kiquebg on Pixabay)
Photo: kiquebg / Pixabay

An AI chatbot answers questions. An AI agent does things: it can draft and send an email, update a customer record, create an invoice or move a deal to the next stage. That difference is what makes agents useful, and it is also what makes them worth supervising.

This guide explains how a small business can give an agent real work while keeping a person in charge. It uses plain security guidance rather than hype, and it applies whichever tool you choose. It is general information, not professional advice.

What changes when an AI can act

When a chatbot gets something wrong, you read a bad answer. When an agent gets something wrong, a wrong thing has already happened: a message is sent, a record is changed, a payment is queued.

The OWASP guidance on "excessive agency" describes the problem for systems that can call tools or connect to other systems. It says harmful actions can follow when a model's output is unexpected, ambiguous or manipulated, with hallucination and prompt injection (malicious instructions hidden in content the AI reads) among the common triggers.

Three ways agents get too much power

OWASP names three root causes, and each has a small-business version:

  • Too much functionality. The agent can do more than the job needs, such as a tool meant to read files that can also delete them.
  • Too many permissions. The agent has access to more systems or data than necessary, such as write access where read access would do.
  • Too much autonomy. The agent takes high-impact actions without anyone confirming them.

The fix for all three is the same idea: give the agent the narrowest job, the narrowest access, and a person at the gate for anything that matters.

Sort actions by how much damage they can do

Not every action needs a human click. Approving everything defeats the purpose and trains people to click "yes" without reading. A simple three-level sort keeps the review effort where it counts.

| Level | Examples | Suggested control | |---|---|---| | Low: reversible, internal | Summarizing a thread, drafting text for you, suggesting a next step | Agent may do it; review the result when you read it | | Medium: visible or hard to undo | Sending a customer email, updating a contact or deal, posting publicly | A person approves each action before it happens | | High: money, legal or irreversible | Issuing invoices or refunds, deleting records, signing or sending contracts, changing permissions | Person approves, with a second check for large amounts |

OWASP's mitigation list includes this directly: require a human to approve high-impact actions, with the person confirming before the action runs. Its example is a user reviewing a drafted post before it is published.

Principles for keeping control

1. Least privilege

Give the agent only the access its task requires, and run it under a specific person's or role's permissions, not a catch-all administrator account. OWASP recommends this scoping and also recommends limiting the number of tools an agent has and avoiding broad, open-ended ones.

2. Enforce the rules in the system, not in the prompt

Telling an agent "never delete anything" in its instructions is a request, not a lock. OWASP advises enforcing authorization in the downstream system rather than relying on the AI to decide what is allowed. In practice that means the permission should be missing from the account the agent uses, not just discouraged in writing.

3. Keep a record

OWASP notes that logging, monitoring and rate limiting do not prevent excessive agency but can limit the damage when it happens. A log that shows what the agent proposed, who approved it and what changed lets you find and fix a mistake quickly.

4. Assign an owner

Someone in the business is responsible for each agent: what it may do, how it is checked and when it gets switched off. The NIST AI Risk Management Framework puts this under its Govern function, which NIST describes as setting policies, processes and accountability. See NIST's description of the framework core.

5. Plan for problems

NIST's Manage function covers post-deployment monitoring, incident response and recovery. For a small team that means knowing how to pause the agent, undo its last actions and tell affected customers.

A setup checklist before you switch an agent on

  1. Write one sentence describing the agent's job and what it must never do.
  2. List every system it can read and every system it can change. Remove what it does not need.
  3. Mark each action it can take as low, medium or high using the table above.
  4. Turn on approval for every medium and high action.
  5. Name the person who reviews approvals and a backup for when they are away.
  6. Run it for two weeks in a draft-only mode and read every output.
  7. Check the log weekly for the first month, then monthly.
  8. Decide in advance what makes you pause it, such as two wrong customer messages in a week.

Making approval work in real life

Approval only protects you if it is real review, not a reflex.

  • Show the action, not just the request. The reviewer should see the exact email, the exact record change or the exact invoice, not a summary of intent.
  • Keep the queue small. If you are approving fifty items a day, tighten the agent's scope or raise the autonomy of the truly low-risk steps.
  • Spot-check the approved ones. Look back at a sample of what went through to see whether review is catching problems.
  • Be careful with untrusted input. Agents that read inbound emails, web pages or uploaded files can be fed instructions by whoever wrote them. Keep such agents on read-and-draft only.
  • Review when scope grows. Each new tool or permission is a fresh risk decision.

Common mistakes

  • Giving an administrator login to an agent. It inherits every power that login has.
  • Relying on instructions as a safety control. Prompts guide behavior; permissions enforce it.
  • Approving in bulk without reading. That is automation with extra clicks.
  • No log. If you cannot see what happened, you cannot fix it or learn from it.
  • Launching on live customers first. Start in draft mode and widen slowly.
  • Forgetting the human in the loop has a bad day. Have a backup reviewer and a way to pause.

Where WorkBOS fits

In WorkBOS, AI agents are built to act only with human approval, so a person sees a proposed change before it is applied. Access is shaped by roles and permissions, and an audit log records activity so you can review what happened. Agents work across the same projects, CRM and invoicing data your team uses, which keeps the approvals in one place. Details are in the docs, plan options are on the pricing page, and the Academy has learning material. For the broader picture, see our guides on practical and safe AI uses and AI and customer data privacy.

Frequently asked questions

Q: What is the difference between an AI assistant and an AI agent? A: An assistant produces text or answers for you to use. An agent can take actions in your systems, such as sending messages or changing records, which is why it needs tighter limits.

Q: Should a person approve every agent action? A: Not necessarily. Reversible internal steps can run on their own. Anything customer-facing, hard to undo, or involving money or legal documents should have a person confirm it first.

Q: Can I rely on telling the agent not to do something? A: No. Instructions can be ignored or overridden by manipulated input. Remove the permission from the account the agent uses so it cannot do the action at all.

Q: What is prompt injection? A: It is when instructions are hidden in content an AI reads, such as an email or web page, to make it do something its owner did not intend. Limiting what the agent may do is the main defense.

Q: How do I know an agent is behaving? A: Keep a log of what it proposed and what was approved, review it on a schedule, and spot-check approved items for errors.

Sources are linked in the text above: OWASP GenAI Security Project, NIST AI RMF core.

Run it all from one place

Projects, clients, finance, people and AI employees in one system. Start free, no card needed.

Start free →