The word "agent" entered business vocabulary a little over a year ago and already means too many different things. For some it is the website chatbot under a new name. For others it is a system that handles orders on its own. This week Meta made it clear which side it is betting on, and the decision has practical consequences for anyone running a small company.

On 28 September 2026, Meta announced a new division dedicated to selling artificial intelligence to businesses, the Enterprise Platform. To lead it, the company hired Chirantan Desai, who stepped down as chief executive of MongoDB to become Meta's Chief Enterprise Platform Officer, reporting directly to Mark Zuckerberg. The initial focus is on four products: Muse, the Meta Business Agent, and the developer-facing Muse API and Muse Code. Zuckerberg described the bet as a way to use "our strengths that few other companies have: advanced models, leading agents, large-scale infrastructure, and years of working closely with many businesses."

Translated for anyone running a shop, a clinic or a workshop: Meta is going to push AI agents into WhatsApp, Instagram and Messenger, which are already the channels where your customers talk to you. The Meta Business Agent, in fact, launched globally across those three channels back in June.

The interesting question is not whether this is worth it. It is what exactly is worth handing over to a machine.

Three different things sharing one name

Most of the confusion disappears once you separate three levels.

A chatbot answers. It takes a question, looks up a response in a script or a set of documents, and returns text. It decides nothing and changes nothing in the business. It is useful for opening hours, addresses, prices and order status, and it is the lowest-risk form of automation there is.

A copilot suggests. It drafts the email, proposes the reply to the customer, prepares the quote, organises the information. The decision stays with a person, who approves, edits or discards. This is where most small companies find their best return, because it removes the cost of starting each task without taking anyone out of the loop.

An agent acts. It has access to systems and permission to execute: book, cancel, buy, send, reply on behalf of the business. It is the level with the most upside, and it is also the only one where a mistake becomes a done deal rather than a bad draft someone deletes.

Calling all three "agents" is what leads companies to buy risk when all they needed was convenience.

The case this week that shows where the line is

A few days before the Enterprise Platform announcement, a Canadian content creator, Matt J. Robb, let Muse manage one of his Facebook Marketplace listings. The assistant accepted an offer below asking price, arranged an in-person pickup, shared the pickup location and, when the buyer arrived, automatically replied "Yup, I'm here!" while he was not at home. He only found out late that night.

Meta's version is worth recording, because it changes the lesson. Muse said it had shared only the street-level pickup location, never the unit number or postal code, and that this was part of the auto-reply template the user approved when setting up Marketplace replies, while conceding that the user never specifically authorised handing out the address. David Singleton, of Meta Superintelligence Labs, responded publicly that in the reported cases the assistant was following direct instructions and correctly asking for permission. Robb disputes that reading.

Both versions can be true at once, and that is precisely the point. The system did what it had been configured to do. The user had not understood what he had configured. In a business, that gap between what was authorised on paper and what someone thought they had authorised is where expensive problems live.

The question that settles it

Before asking whether AI can do a task, there is a more useful question: what happens if it goes wrong?

If the answer is "the customer gets one piece of wrong information and someone fixes it in two minutes", automate away. If it is "a booking is lost", "a price goes out that we cannot honour", "something gets disclosed that should not have", or "a customer waits for someone who is not coming", then the task is not a candidate for full automation. It is a candidate for automation with a human approval step.

This hybrid model gets little attention because it makes a poor headline, but it is what works. The machine prepares, the person confirms with one click. You keep ninety per cent of the time saving and one hundred per cent of the control.

Where it genuinely pays off in a small company

A real example at Portuguese scale. A remodelling company in Braga gets a quote request on WhatsApp at 10:47 on a Friday night. Today that request sits unanswered until Monday morning, by which point the customer has asked three other companies for a price. A well-configured assistant replies within seconds, confirms the type of work, the area and availability for a site visit, and by nine on Monday there is a qualified lead on the table instead of an unread message. It closed nothing, and it is not supposed to. It bought back attention that was being lost.

A clinic taking fifteen calls a day about appointments has a different calculation. If the assistant books straight into the calendar, every error is a patient arriving at the wrong time or a slot lost. If it only captures the request into a list the receptionist confirms each morning, the same work gets done with the risk removed.

The difference between the two situations is not technological. It is the cost of the error.

Seven questions before automating a task

Any candidate for automation is worth running through this list.

How many times a week does this task happen? How long does it take each time? Does the right answer depend on information already written down somewhere, or on human judgement? What happens if the answer is wrong? Is there anyone who would catch the error before the customer does? Does the task require access to customer data or irreversible actions? If we removed the automation tomorrow, would the business go back to working as before?

Frequent, fast tasks based on written information with a cheap failure mode come first. Rare, slow tasks that require judgement and whose errors cost money or reputation come last, and some should never make the list at all.

The advantage of being small

There is an irony in this market. The large platforms are building generic agents to serve millions of businesses at once, which is exactly why they know nothing about yours. A twenty-person company can do the opposite: automate five specific tasks, with its own vocabulary, its own prices and its own rules, connected to the systems it already uses.

It is also worth saying what is not yet known about Meta's announcement. No pricing, general availability dates, or the administrative and contractual controls a business needs before letting an agent speak to customers in its name. Until then, anyone serious about automation does it with tools they already control.

The best automation is almost always invisible. Nobody praises the assistant that answered the 10:47pm enquiry. You only notice the business you did not lose.

How Digito helps

The work starts before the technology. We map the repetitive tasks in your operation, calculate how much time they consume each month, and separate the ones worth automating from the ones that are not, with the cost of failure always on the table. Only then do we build.

In practice that means three things: artificial intelligence assistants and automations connected to WhatsApp, your website or your inbox, designed around your vocabulary and your rules; custom software and integrations that connect those automations to the CRM, invoicing or calendar you already use, so information does not stay trapped in a conversation; and hybrid processes, where the system prepares and a person approves, for anything with a high cost of error.

If you want to work out which tasks in your business are genuinely worth automating, and which should stay with a person deciding, talk to us at digitosolutions.com.

Sources