AI agents for business: decide what they can do without asking

AI agents for business got a lot more real last week. OpenAI and Meta both shipped agents that stay on all day, connect to the tools you already pay for, and do work while you are somewhere else. If you run a company, the useful question now is what you let one do without asking you first.

I want to walk through what shipped, and then how we answer that question in our own business, because we have been running AI inside our operations for a while and the answer turned out to be simpler than I expected.

What shipped last week

At its DevDay, OpenAI introduced dots, agents that keep working around the clock from a cloud computer. According to The Rundown's write-up, each dot can connect to more than 4,000 apps and reply in Slack, Teams or ChatGPT, and the first one comes with the Pro and Business Premium plans to start. OpenAI's own summary is in its DevDay recap.

The same week, Meta launched Muse for Small Business. It connects to tools a lot of small companies live in, including Slack, Shopify, Stripe, QuickBooks, Asana and HighLevel, along with your Facebook and Instagram business accounts. The line in Meta's announcement I would underline: "nothing publishes, sends, or spends without your approval."

So two of the biggest companies in the space are putting an agent next to your team, wired into your systems. For a founder, connecting an agent to the CRM or the inbox is turning into a setting you switch on.

The question that matters for AI agents for business

Once an agent can reach your tools, most of the thinking left for you is deciding which of its actions need a person.

Meta answered that in its launch copy: nothing that publishes, sends or spends happens without you. The Rundown's own setup guide for dots says the same thing in practical terms. Its advice for a first test is to leave access to your local computer off, set the agent to ask before taking action, and keep sending, publishing and purchases out of the task entirely. Then check what it produced against the facts, and before you put it on a schedule, give it an update that changes a deadline and see how it handles that.

I think that is exactly right, and it matches where we landed.

How we draw the line in our own business

We run BYLD on Byld Ops, the operating platform we built for ourselves and now set up for founders. Our proposals, invoices, projects and client records all live in it, and it has an AI assistant that works across all of it.

The assistant can search everything, draft replies and proposals, update records and create tasks when someone asks it to. Anything that leaves the company needs sign-off from a person. Sending a proposal, emailing an invoice or recording a payment waits in an approvals queue until someone on the team approves it. Agents we connect from outside, over MCP, get exactly the permissions of the person who issued their key, and nothing more.

That rule has held up because it is easy to remember and easy to check. An agent can prepare as much as it likes. The moment something would reach a client or move money, a person looks at it first. You can draw the line somewhere else for your business, but write it down, and make the agent's settings match what you wrote.

To be super transparent, the approvals queue is where the value shows up for us. Reading a drafted follow-up and pressing approve is faster than writing it, and the review is where you catch the draft that misread the situation.

Where agents earn their keep first

The best first jobs for an agent are the ones where a mistake costs you a minute, and where the work piles up when nobody owns it:

• Sorting inbound requests and routing them to the right person.
• Drafting follow-ups after a call, so the email is waiting for you instead of on your list.
• Keeping client records current from the messages and meetings that already happened.
• Preparing a morning brief of what is due, what is late and what needs a decision.

When we run operations for a company as its fractional COO, this is the kind of work we build agents for: sorting inbound requests, drafting follow-ups and preparing proposals, with anything going out waiting for approval.

We have built agents into products for founders too. BacklinksAI matches people with journalists' requests for expert sources and drafts responses, customizable to each user's writing style and expertise. Thredly is an AI agent that finds a company's customers' conversations across 8 platforms.

Before you turn one on

If you are about to connect OpenAI's dots, Meta's Muse or anything like them to your business, this is the order I would go in:

1. Start with read access. Let it summarize and draft before it can change anything.
2. Write down the actions that always need a person, then set the agent's rules to match. Sending, publishing and spending should be on that list on day one.
3. Test it on a change, not only on a clean first run. A good first answer tells you little about how it handles a moved deadline or a client who changed their mind.
4. Give it one owner. Someone on the team reads what it drafts each day and adjusts the instructions when it gets something wrong.
5. Keep a record of what it did. If you cannot see the history of an agent's actions, you cannot trust it with more.

If you want a second opinion on where agents fit in your operations, or you would rather have it built and running for you, our AI chief of staff page covers how we set up Byld Ops, and you can book a call below.