What Can Actually Run Itself in a Business With AI in 2026?

Let's say a lead fills out a form on a real estate website. And before the team opens the CRM, AI has read the answers, understood what they want, it qualified them, created or updated the record, routed this request or inquiry to the right person, decided and suggested the next step (and possibly executed the next step), and notified the team or certain members on the team. And if something falls outside the predefined rules, it stops, flags it and notifies about it.

And this whole thing can happen in under two minutes without anyone having to touch the keyboard.

And this is what I usually mean when I say certain parts of a business can run themselves. At this point, it's not about the whole company, it's not about every decision, and it's definitely not an AI with unlimited access to everything.

What I mean is a defined and outlined piece of work that can understand what is happening, decide what has to happen next, and take a certain action across the company's systems or tools. And it has to be able to keep going until it reaches an outcome or it hands it off to a person.

And that's what is already real in 2026.

We have moved past AI that only helps you work

Most companies first experienced AI as something that just helped a person do a task. It could be writing an email, summarizing a meeting, researching a company. It could also draft a proposal, for example.

All these things are useful, but the person still owns the process in this situation. The person starts it, moves the information around, makes the decisions, and pushes the work to the next step.

And agentic systems change that. What they can do: they can retrieve information, use your tools, make bounded decisions, which is important, take certain actions, see what happened, and keep going until the job is done or something just falls outside of their authority.

AI helping with the work and AI owning a piece of work are two very different operating models.

This is already happening inside real companies

And I like these three examples that demonstrate what's currently happening inside real companies.

There's this company, Danfoss. It uses AI agents to process B2B orders that arrive through email, and Google Cloud reports that more than 80% of transactional decisions are now being made by their internal AI. And at the same time, processing time is down by 50%.

The second company is Lemvigh-Müller. It uses AI agents to read supplier order confirmations. Their AI compares prices, quantities and delivery dates against purchase orders in their SAP, and it also processes their standard cases. And their system went from idea to production in about 10 weeks, and is designed to handle more than 100,000 confirmations a year.

I like another company as an example - Delivery Hero: they have an internal engineering agent that creates and merges more than 170 pull requests every day, and its success rate is 85% for tickets merged into production.

And here we're talking different industries and different workflows, but the pattern for all of them is very similar.

So what can actually run itself?

I usually don't start by asking which employee can or should AI replace. This is just too broad and not very useful.

I look for a piece of work that can drive an outcome and own an outcome inside predefined boundaries. And for me, a workflow becomes a serious candidate when six things are in place.

1. The outcome is clear

Because the system needs to know what done actually means. Let's say an order is entered and accepted. A support request is resolved. A lead is qualified, routed and logged. An invoice is matched, or moved into an exception queue.

And here, "help the business" is not an outcome. So the completion state has to be very tight, and that's what makes it easier to let AI own the work.

2. The AI can access enough trusted context

AI cannot efficiently run a workflow if certain information, information that's needed to make decisions, lives in someone's head. It could be an outdated spreadsheet, three inboxes, it could be an outdated policy document that nobody uses or trusts.

The system needs access to the right sources, such as SOPs - standard operating procedures, your CRM, ERP, etc. It could be inventory, customer history, current policies, information from your internal communication tools.

And the main point here: if the business does not know where its source of truth is, neither will the AI.

3. It has the tools to do something

Because providing an intelligent answer is not the same as completing work.

If the system understands that a customer wants to change an appointment but cannot access the calendar, it means the system still needs a person. If it has identified a qualified lead, but it cannot access the CRM or route that lead, the work stops.

And for work to run itself, AI needs access, controlled access, to the tools where the work actually happens.

4. Its authority is defined

And this is where I think a lot of conversations around AI agents go south. The main goal is not to maximize AI's autonomy. The goal is to make it as useful as possible, with the autonomy inside the rules you trust.

A customer service AI might be allowed to issue a €20 credit automatically, but a €2,000 refund has to go for approval. AI can for sure schedule a meeting, but in certain situations it may not sign an agreement.

So the question is not simply, can AI do this? The question is: what are the conditions under which we are comfortable to let it do something without asking?

5. We can verify what happened

This part matters more than repetition. Delivery Hero's engineering tickets, they're not all the same. And what makes their environment suitable for higher autonomy is that the result can be checked. It can be checked in different ways: through tests, security controls, some form of rollback.

And the business needs some equivalent: did the payment match? Did the CRM update? Did the order enter the ERP correctly? Did the tickets actually close?

And it's important to be clear here. If there is no reliable way to know whether the system succeeded, I become very conservative in these situations about letting it run unattended.

6. There is a clear exception path

Something eventually goes outside the normal path. Very, very often something can go outside the normal path. The customer could ask for something unusual, and suddenly two internal systems disagree, some key information is missing, the transaction is too large, and as a result a certain policy exception appears.

The important part is to design what should happen next. A good AI operating layer should be able to tell the person what happened, what it checked, what it identified, what actions it took. If certain information is missing, then what information is missing, and why it stopped.

So the human should deal with such a situation as an exception, and not the entire raw problem. And this part changes the economics even when AI cannot resolve 100% of the workflow.

What stays deterministic

Certain things have to stay deterministic, not everything inside an AI system should be decided by AI.

Very often a business already knows exactly what should happen every time, and I usually want deterministic logic to enforce that. It could be a refund ceiling, an approval hierarchy, it could be some form of authentication, inventory availability. It could be a compliance restriction.

AI is useful where language is messy, context matters, information needs to be interpreted, and the right path can at some point change. Deterministic systems are usually better where the same rule should be enforced every time.

And how I approach that - I use AI for understanding and coordination. And I use deterministic logic for certainty.

And some decisions should still stay human

With AI, some decisions should still stay human, and in this case Stripe is a very good example. Its compliance teams use AI agents to gather information and prepare parts of complex financial risk investigations. And Stripe reports a 26% reduction in median review handling time, while the final decision still remains with a person.

And that is definitely not a failure of autonomy. That is a very good system design.

Some decisions justify human ownership because of the severity of the consequences, such as large financial commitments, serious customer disputes. It could be employment decisions, safety, etc. The system can still do a huge amount of work around that judgment.

Where I start

I start with work that already has enough structure to define an outcome, but still requires people to spend time on things like reading, deciding, copying, routing, following up, or often switching between systems.

Then I look at volume, and I ask: how often does this happen? What gets delayed? What gets missed? What happens after hours? Where does somebody have to chase somebody else?

And then I define and design the system around how the work actually happens. Basic questions are: what AI needs to know, what data it needs, which tools it can use, which actions it owns, which rules stay deterministic, where its autonomy stops, what has to be flagged, and what makes the job actually complete.

My approach is, I don't try to make the entire business autonomous right away. I make one useful piece of work run properly, then I measure it, and expand and scale from there.

Questions business owners ask

1. How much of a workflow should AI be trusted with? As much as it can handle reliably inside clearly defined rules and authority. It can start small. The system just observes and recommends, then it drafts something, then it takes low-risk actions. And then more authority can be added as the workflow starts proving itself.
2. Does the workflow need to be repetitive? Well, high volume makes the business case much stronger, but repetition is not a hard requirement here. What matters more is whether the outcome can be defined and verified.
3. What happens when the AI gets something wrong? It's very important that that has to be designed before launch. Things like permissions, reversibility, logs, escalation, rollbacks are all part of the system, and it's not something added after the first mistake.

The business that runs itself is built one workflow at a time

So, the business that runs itself has to be built one workflow at a time. And the important development in 2026 is not that AI can suddenly start running an entire company. The key point here is that more pieces of real business work no longer need a person touching and checking every step.

In 2026, this is already possible: a request comes in, AI understands it, it pulls the context, chooses the right path, uses the systems, and the business rules define what it is allowed to do. And when the situation is outside those boundaries, a person gets to deal with an exception with the context already processed by the AI.

So now that is a very different business from one where every email, order, lead, document, request or decision has to be reviewed manually and passed manually from one person to another.