2026 is the year AI stopped answering and started working

For three years we got used to asking artificial intelligence questions. Write this email. Summarise this document. Give me ten ideas for a campaign.

The AI answered, we copied the answer, and we carried on with the work.

That phase is changing. The generation of systems now reaching the public does not only want to tell you what to do. It wants to do it. And the past two weeks gave probably the clearest demonstration of that shift.

From chatbot to agent

On September 8, Meta launched Muse, which it presents as a personal AI agent.

The difference from a chatbot looks small until you see what it does. Muse opens a browser, fills in forms, sends emails, books services and makes purchases. It keeps working on tasks after the user closes the app, and comes back when it needs a decision or an authorisation.

On September 18, ten days later, Meta brought Muse to the Mac, where it interacts, under permissions granted item by item, with files, messages, calendar, notes and mail inside the system's own applications.

It is worth putting the right scale on this. Muse is available only in the United States, to users over 18, and requires a payment card even on the free tier. It climbed quickly up the US App Store charts, though it is worth noting that in the same week it sat around position 338 in Google Play's productivity category. The iOS figure flatters it.

We are not suggesting a Portuguese small business installs Muse today and transforms itself. What matters is the behaviour emerging: ordinary people experimenting with artificial intelligence that executes.

So what is an AI agent?

A chatbot answers. An agent executes. It is a simplification, but it is the fastest way to grasp the difference.

Imagine telling an AI system: I have a meeting tomorrow with this prospect, summarise the history of the relationship and prepare five points to discuss.

A traditional assistant helps write those five points if you supply all the information. An agent, with the right permissions, can consult the CRM, find the emails exchanged, review previous proposals, check the calendar, identify open items and leave the summary ready before the meeting.

The difference is moving from generating information to executing a process. That is where the change that matters to a business begins.

The leap is not a smarter AI

For a long time we judged these systems by the quality of their answers. Which writes better, which reasons better, which makes fewer mistakes.

Those differences still count. But for a company another question becomes more interesting: what can this AI actually do inside my business?

An extraordinarily capable AI, shut inside a chat window, still depends on a person to execute almost everything. An agent connected to the right systems takes part in the workflow itself. That difference can be worth far more than an improvement in answer quality.

Where this touches a real business

Picture a small company with five salespeople. Every day enquiries arrive by website, WhatsApp, email, Instagram and phone. Somebody has to log the contact, work out what the person wants, look up information, assign it, reply, schedule the follow-up, create the opportunity in the system and update its status.

None of those tasks is difficult. Together they consume hours every week.

The same happens in an online store. A customer asks whether a product will arrive before Friday. Answering properly requires knowing the product, the stock, the location, the carrier and the lead time. A chatbot built only on frequently asked questions cannot. An agent connected to the right systems can.

We are moving from rigid automations of the "if A happens, do B" kind to systems able to interpret what the person is trying to resolve and what actions get them there.

The website will have to change too

There is a less obvious consequence here.

Websites were built for decades for people: menus, buttons, forms, pages. But if someone tells an agent to find three companies capable of building an online store, compare the services and book a meeting with the most suitable one, the one browsing is no longer the person.

In that scenario a good-looking website is not enough. The information has to be organised so that people, search engines and AI systems can all understand who the company is, what it does, where it works, what the services and conditions are, and what the next step is.

This is where SEO, GEO and web development meet. We wrote recently about the first half of this story, appearing in the answers when someone asks an AI who does this locally.

The serious problem is trust

Giving an AI system access to a calendar is one thing. Giving it access to email is another. Authorising it to buy, change information or communicate on someone's behalf is considerably more serious.

The builders know this. Meta says Muse runs on a dedicated virtual machine with its own browser, and that it asks for explicit authorisation before sensitive actions, with those requests presented outside the conversation precisely so they cannot be manipulated by injected instructions.

That does not remove the risk. Reuters reported that internal testing identified reliability problems and incidents of unauthorised exposure or transmission of sensitive information. A Meta executive confirmed to the agency that a launch planned for April was delayed for additional security work.

For a company the lesson is direct: automating is not handing over indiscriminate access. A business agent should run with minimum permissions, logging of actions, human approval for critical decisions, data separation, and a clear way to stop execution. The greater the autonomy, the greater the control has to be.

So should we put agents everywhere?

No, and that would be the worst starting point.

The right question is not where can we put AI. It is this: where are we spending human time on repetitive tasks that follow a predictable process?

Many companies discover they lose dozens of hours a month copying information between systems, answering the same questions, preparing similar documents, qualifying contacts, hunting for information buried in emails and producing reports.

Five processes a small company can review now

  1. Contact qualification. Receive enquiries, interpret the need, collect what is missing and route to the right person.
  2. Sales preparation. Gather emails, history, documents and system information before a meeting.
  3. Customer service. Answer frequent questions with the company's real information and pass exceptions to a person.
  4. Internal operations. Collect data from several systems, update statuses and produce reports.
  5. Follow-up. Identify stalled processes, customers without a reply, or proposals with no follow-through.

None of these requires science fiction. They require integration, rules, organised data and control.

The real asset is still the company's own information

There is an irony in this revolution: the better the models get, the more the quality of the information we give them matters.

An AI does not know how your company works, what your minimum margin is, which clients take priority, when a discount can be granted, or when a complaint has to go to a person.

That knowledge exists, but it often lives in somebody's head, in a spreadsheet, in an old email or in a WhatsApp group. Which is why implementing agents rarely starts with AI. It starts with organising the company.

And that is where the real competitive advantage sits. Having access to artificial intelligence no longer distinguishes anyone, because millions of companies have the same tools. What distinguishes is being able to connect it to real processes, safely and usefully.

Frequently asked questions

What is an AI agent? A system able to execute tasks and use tools to reach a goal, rather than only generating a text answer.

What is the difference between a chatbot and an AI agent? A chatbot responds to requests. An agent can, depending on permissions and integrations, consult systems, fill forms, send information or update processes.

Can a small company use AI agents? Yes. The most useful cases sit in repetitive processes: contact qualification, customer service, sales preparation, reporting and administrative tasks.

Are AI agents safe? They introduce risks, particularly when they have access to data or the ability to execute actions. They should be implemented with minimum permissions, supervision, logging and human approval for sensitive actions.

Will they replace workers? There is no universal answer. In the short term, the most relevant use is automating the repetitive parts of processes, freeing people for what requires judgement and relationships.

How Digito helps

We do not start with the tool. We start with the process.

We analyse where your team's time is being consumed by repetitive tasks, identify which of them have real automation potential, and build agents and integrations adapted to how the business actually works, with the permissions, logging and approvals a company needs.

We also prepare the business's information to be legible to these systems, which is almost always the work missing before any automation makes sense.

And when the analysis shows a process needs tidying rather than automating, we say so.

If you want to know where artificial intelligence could genuinely save time in your company, let's talk about it.

Sources