AI agents in SMBs: how much access do you give your digital assistant?

Reading time: 6 minutes · For owners of SMBs in Ausserschwyz, the Linthgebiet and around Lake Zurich who are still writing quotes in the evening

It is 9 pm, the children are asleep, and you are still working on three quotes, a supplier order and the question of why last week’s customer never called back. This is what marketing and sales look like in many SMBs: nobody really does it, everybody does it on the side. This is exactly the work AI agents now want to take over. The only question is how much access you give them.

This week, journalist Joanna Stern spoke with Mark Zuckerberg about Muse, Meta’s new AI agent. I watched the interview for the first time on camera and commented on it:

Why this now affects SMBs too

Muse is not a chatbot that answers questions. It works on its own virtual computer at Meta, accesses emails, accounts and appointments, and completes tasks on its own, even overnight. Meta is also bringing it to its own glasses. Zuckerberg’s examples are deliberately everyday: finding and cancelling unnecessary subscriptions, getting a permit, sorting out the recipe and the ingredients at the weekend.

Tools like this have been around for professionals for a while, such as Gemini Spark from Google or Claude Cowork from Anthropic. What’s new is that they are becoming simple enough to reach a business without an IT department. And with that, the question of trust lands on your desk, not just on the desks of large corporations.

Joanna Stern herself did not connect her private Gmail. She doesn’t fully trust Meta and worries that the agent will send emails in her name that she would never have written that way. That isn’t paranoia. It is the right question. Meta answers with architecture: according to Zuckerberg, Muse runs in a confidential virtual machine that not even Meta can see into, and a second agent monitors everything that goes in and out and asks before sensitive actions.

What to do now

1. Start with admin, not with customers

The biggest relief comes where no sensitive data is involved: reports, orders, appointment suggestions, first drafts of quotes. In my own work, agents have taken over exactly these tasks over the last twelve months. The time I gained didn’t go into even more screen work. It went back to people. What this looks like in the back office is something I described in the post on AI as a lever in the back office.

2. Set sign-offs before anything goes out

Meta is building a guard into Muse that checks back on logins, payments and sensitive messages. Do the same on a small scale: an agent may draft, sort and prepare. Anything that goes to customers, suppliers or banks is signed off by a person. That takes a few seconds and prevents the one email you can’t take back.

3. Don’t automate a bad process

An agent doesn’t make a chaotic follow-up process better, only faster. If nobody records today who enquired and when to follow up, the agent will widen exactly that gap. First you need clean lead and customer nurturing, then automation.

4. Answer the data question honestly

My personal stance is pragmatic: if you don’t want to share something, don’t use the tool. For a business, that stance alone is not enough, because it is about your customers’ data. The revised Data Protection Act has applied since 1 September 2023. You need to know which personal data goes where, and inform your customers about it. Keep a list of which agent has access to what. That is all you need to start with.

5. One person who keeps the overview

The real risk is not the tool. It is that nobody is responsible. If everyone in the business connects their own assistant, in the end nobody knows what is happening in whose name. This is exactly where marketing without your own marketing department comes in: one point of contact who leads marketing and sales with you and decides which tasks an agent may take over and which not. If you need a data protection or technical specialist, they join the table in the same week.

Where the limits are

For medical practices, fiduciary offices and anyone working with health or financial data: no test with real customer data until it is legally clear where the data is processed. And where the personal relationship is the actual product, such as the regular customer who has been calling for twenty years, the contact does not belong in the hands of an agent. There, the agent should free up time so you can have the conversation yourself.

Zuckerberg is well aware of the limits of today’s tools. Not everything works the way it should. But the pace is enormous, and for a business in Lachen the same maths applies as for a corporation: hand off the admin and you win time for your customers. Hand off customer contact without thinking it through and you lose trust. How Meta itself experienced this with the shopping agent is covered in the post on Amazon and AI shopping agents.

Frequently asked questions

What is Meta’s Muse?

Muse is Meta’s personal AI agent. It runs on its own virtual computer at Meta, can access emails, accounts and appointments, and completes tasks on its own instead of just answering questions. Meta is also bringing it to its own glasses.

Can an SMB let an AI agent access customer data?

In principle yes, but the revised Data Protection Act applies here too: you need to know which personal data goes where, and inform your customers transparently. With particularly sensitive data, for example in practices or fiduciary services, it is worth getting advice before the first test.

Where is the best place to start?

With recurring admin that involves no sensitive data: reports, orders, appointment suggestions, draft quotes. That saves you time straight away at little risk. Only hand over customer contact itself once the process behind it is clean.

Do I need an agency for this?

No. You need someone who understands the process and sets the sign-offs correctly. That can be someone in-house or a person on a mandate who leads marketing and sales with you.