Before an AI tool clicks anything for your business, answer four questions: what can it see, what can it change, what needs a person's yes, and who owns it. Write the answers on one page for each tool, and you've decided what your AI is allowed to do before it decides for you.
That's the whole idea. The rest of this post walks through each question in plain English, with examples from the kind of work Orange County businesses do every day.
Give every AI tool a job description: what it can see, what it can change, and what needs your yes.
Why this matters now
Not long ago, AI mostly wrote things. It drafted an email, and you decided whether to send it. Now more AI tools can act. They can update a customer record, send a message, move a file, fill out a form on a website or click "Submit" on an order.
That's useful. It's also a different kind of tool. A tool that only writes drafts can't do much harm on its own. A tool that can click buttons needs the same thing a new employee needs: a clear idea of what the job is and where it stops.
Last week we walked through how to count the AI tools already in your business. This is the next step. Once you know what you have, decide what each one may do.
Question 1: What can it see?
Start with what the tool can read. Your inbox? Your customer list? Your accounting software? Shared folders with contracts and payroll?
Many AI tools ask for wide access when they're set up because it's easier. Someone clicks "Allow" and moves on. But a scheduling assistant doesn't need to read payroll, and a website chatbot doesn't need your full customer history.
Ask: What does this tool need to read to do its job, and what can we take away?
Question 2: What can it change?
Reading and changing are very different. An AI that reads your calendar to suggest times is low-stakes. An AI that can move or cancel appointments is not.
List what the tool can create, edit, send or delete. Pay close attention to anything that touches customers, money or records you'd have trouble rebuilding.
Ask: If this tool made a mistake here, could we undo it easily?
If the answer is no, that action belongs under Question 3.
Question 3: What needs a person's yes?
This is where you draw the line. Some actions are fine for AI to handle alone, like sorting incoming email or drafting a reply. Others should wait for a person to approve them.
Good candidates for "a person approves first":
- Sending anything to a customer for the first time
- Changing prices, quotes or invoices
- Anything involving payments or refunds
- Deleting records or files
- Anything you'd be uncomfortable explaining to a customer if it went wrong
A useful test: would you let a brand-new hire do this alone in their first week? If not, don't let the AI do it alone either. You can always loosen the rule later once you've watched the tool work.
Some AI makers now build this step into their tools. Anthropic's documentation for its Claude Agent SDK, a kit developers use to build AI agents, describes an approval step: when the agent wants to take an action that hasn't been cleared ahead of time, the request goes to an approval check that can allow it or deny it. Builders can also list actions that are always refused, no matter how loosely the rest is set. Ask whoever builds your AI tools whether they use a step like this.
Ask: Which of this tool's actions should wait for a person, and who is that person?
Question 4: Who owns it?
Every AI tool needs a name next to it. Not "IT" and not "the vendor." A real person in your business who checks on it, answers questions about it and decides when its job description changes.
Without an owner, tools drift. Access gets added and never removed. The person who set it up leaves, and nobody knows what it can do anymore.
Ask: If this tool does something odd on a Tuesday afternoon, who gets the call?
Put it on one page
You don't need software for this. For each AI tool, write four short lines:
- Can see: ...
- Can change: ...
- Needs a person's yes: ...
- Owner: ...
Keep it with your other business documents and look at it again when the tool gets a new feature, when you change vendors or when the owner changes roles. It takes a few minutes per tool, and it turns a vague worry into a clear decision.
This is a management decision, not a tech project. You already decide what each employee may do, approve and sign. AI tools deserve the same thinking.
How we handle it at StorenTech AI
When StorenTech AI builds an assistant for an Orange County business, we write that job description before anything gets built. We set clear limits on what each assistant can reach, keep those limits outside the AI itself, log what it does, and have a person approve anything risky.
At StorenTech AI, we put it simply: an AI tool should never be able to do more than the job you gave it.
Start with a map
If you'd like help deciding where AI pays in your business and what each tool should be allowed to do, start your AI Opportunity Map. We look at where your time and money actually go, then recommend the first AI job worth doing, with the limits written down before anything gets built.
FAQ
What should a business AI tool be allowed to do?
Only what its job requires. Decide four things for each tool: what it can see, what it can change, what needs a person's approval and who owns it. Write them down on one page.
Which AI actions should always need human approval?
Anything that's hard to undo or touches customers or money: first messages to a customer, changes to prices, quotes or invoices, payments and refunds, and deleting records.
Is deciding what AI can do a tech decision or a business decision?
It's a business decision. The owner or manager decides what the tool may do, just as they decide what an employee may approve. The technical setup follows that decision.
How does StorenTech AI decide what an AI assistant is allowed to do for an Orange County business?
StorenTech AI maps it during the AI Opportunity Map, before anything is built. Each assistant gets clear limits kept outside the AI itself, its actions are logged, and a person approves anything risky.
How often should I review what my AI tools can do?
Review a tool whenever it adds a feature, you change vendors or its owner changes roles, and at least a few times a year otherwise.