What decisions can AI make on its own?

In 1 minute

The question "do we let it decide on its own?" is not answered with the technology, but with two questions about each decision: how much harm does it do if it goes wrong? and can it be undone? What is low-impact and reversible can run on its own, with a record. What moves money or affects a person, the AI proposes and a person approves. And there are decisions on which the AI only informs. International guidance agrees on something more: the person who approves has to be able to understand, stop and reverse, not just sign.

Imagine a mid-sized financial services company. It is fictitious, but it looks like many. It already uses AI to answer its customers' questions and to prepare internal reports. Now the operations area proposes the next step: that the AI, besides detecting duplicate payments, refunds them on its own.

There are two positions in the meeting. Operations says: "the AI is right nearly every time, and each manual refund costs us a day." Compliance says: "if it refunds money to someone who should not get it, who answers for it?" Both are right. And both are arguing about the AI, when the underlying decision is another: what kind of decision is refunding a payment.

This article is for whoever has to set how far AI goes in their operation. It makes one argument: each decision is classified, as a first filter, by its impact and by whether it can be undone; that classification, and not the technology, suggests whether the AI acts on its own, proposes or only informs.

Not every decision calls for the same thing

Besides answering, an AI can query systems, prepare transactions and, if authorized, execute them. That forces a decision, case by case, about how much autonomy it has.

The AI risk management framework of the United States' NIST says it plainly: human-AI configurations "can span from fully autonomous to fully manual". An AI can decide on its own, hand the decision to an expert or serve as an additional opinion for whoever decides. Some systems, such as those that improve video compression, may not require human oversight; others specifically require it [2]. That is: the right question is not "oversight, yes or no?", but "for which decisions?"

The two questions that order everything

1. How much harm does it do if it goes wrong?

Not in the abstract: in pesos, in customers, in people. A badly worded answer to a frequent question has a small cost. A payment refunded to the wrong account, a credit application rejected or personal data sent to someone who should not have it carry another weight.

As a practical rule, the impact is high when the decision:

• moves money —charges, pays, refunds, applies a discount—;
• changes a record that others use to decide —a balance, a file, an inventory—;
• affects a person's rights or opportunities —a loan, a job, a claim—;
• exposes personal or confidential data.

2. Can it be undone?

A wrongly applied label on an email is fixed with a click. A message sent to a customer can be hard to withdraw and to repair. A transfer can be reversed, but with cost, time and, sometimes, an upset customer. Reversibility is not yes or no: it is how much it costs and how long it takes to undo.

The matrix: three levels of autonomy

With those two questions, the matrix suggests a starting level for each decision:

Classification for illustration; the limits of each level are set by each company.

Impact and reversibility are the first filter, not the only one. Before giving the AI autonomy over a decision, add four questions:

• How likely is it to be wrong? Measured with real cases, not with a demo.
• How many people or transactions does it reach? A small error, repeated ten thousand times, stops being small.
• Is there a legal or contractual obligation over that decision?
• Can the error be detected and stopped before it escalates?

If any answer is uncertain, start with the AI proposing; the level goes up when the evidence justifies it. NIST frames it the same way: risks are prioritized by their impact, their likelihood and the resources available to address them [2].

Back to the duplicate payments. Detecting them is low-impact and reversible: the AI can flag them on its own. Preparing the refund too: it is left ready, not executed. Refunding the money moves money and is costly to undo: the AI proposes it with the evidence —the two charges, the dates, the customer— and a person approves it. The person receives the evidence already gathered and focuses the review on confirming or rejecting; before scaling the process, it is measured whether that way of working reduces time without lowering control. Compliance gets someone who answers.

Make the person's approval real

Putting a person in charge of approving does not work if they approve everything without looking. The European Union's AI Act, in its Article 14 on human oversight of high-risk systems, asks that whoever oversees be able, among other things, to remain aware of the tendency to over-rely on what the system proposes —automation bias—, to interpret correctly what the system produces, to decide not to use it or to disregard, override or reverse its output, and to interrupt it with a stop button or a similar procedure [1].

Outside legal language, that translates into four practical conditions for whoever approves:

• They see the evidence, not just the proposal: where the data came from and why the AI proposes what it proposes.
• They have time to review. If they must approve three hundred proposals in an hour, in practice they are not overseeing.
• They can reject at no cost to themselves. If rejecting creates extra work or a scolding, they will approve.
• They can stop everything: halt the AI in a process while what happened is reviewed.

Who answers

The last piece is responsibility. The OECD AI Principles ask that the organizations and people who develop, deploy or operate AI systems be accountable for their proper functioning, based on their roles and context, and that they ensure traceability of data, processes and decisions [3]. NIST, for its part, asks that human oversight processes be defined, assessed and documented [2].

In a company, that means putting in writing, at least: who owns each process where the AI acts; at what level of autonomy each decision sits; what evidence and what record remain of each action and each approval; who can stop the process; with which criteria it was tested before operating; and what happens when there is an error or when the process changes. If the decision affects a person, also how their case is reviewed and corrected.

And in Mexico?

An honest clarification: as of November 2026, Mexico has no general federal artificial intelligence law in force. There are initiatives in Congress, still not approved [4, 5]. The European regulation and the guidance from NIST and the OECD are cited here as a reference for good practice. The European regulation can reach a company outside the Union depending on the system, its role and where its output is used; if your operation touches Europe, confirm it with legal advice. And the lack of an AI law does not leave AI without rules: depending on the case, data protection, financial, labor, consumer and other rules from your sector may apply. Your legal area is the one that confirms which ones apply to your process.

When to leave it alone and when not to

It can act on its own when:

• The decision is low-impact and easy to undo: classifying, labeling, summarizing, prioritizing a work queue, notifying.
• There are hard limits that contain it —a maximum amount, hours, volume— and a record of each action.
• Someone reviews a sample regularly and can lower the level if something changes.

It is not advisable to leave it alone when:

• The decision moves money, changes a record others use or affects a person.
• There is no record of what it did and why, or nobody knows how to undo it.
• The process has not been measured yet with real cases: first it is tested proposing, and only with the evidence is it decided whether it goes up a level.

What we ruled out, and why

• "If it is right 98% of the time, let it decide on its own." The percentage says how often it fails, not how much each failure costs. A 2% error rate on labels is tolerable; on money refunds it may not be.
• "Let a person approve everything, just in case." It sounds prudent, but it produces automatic approvals and fatigue, and takes away the value the AI had. Oversight is concentrated where it matters.
• "The AI decides and then we audit." A review only afterwards can find the damage when it has already happened. For decisions that are costly to undo, preventive controls, monitoring and after-the-fact review are combined.
• "Let's wait for a law before deciding." Decisions are being made today. Classifying them does not depend on a law; it depends on knowing your own process.

Five questions for your next board meeting

1. Which decisions does AI make today in our company, and at what level of autonomy is each one?
2. Which ones move money, change a record or affect a person, and who approves them?
3. Does whoever approves see the evidence, have time to review and can they reject at no cost?
4. Can we stop the AI in a process today, and undo what it did?
5. Is there a record of each action and each approval, and who owns each process?

If you cannot answer the first, that is the first pending item: make the list.

Where to start this week

Make a table with each decision the AI takes part in —or where you plan to use it— and four columns:

• Impact: low or high, with the reason in one line.
• Reversibility: how much it costs and how long it takes to undo.
• Level: acts on its own, proposes or informs.
• Owner: who approves and who answers.

The rows where the current level does not match the one the matrix indicates are your work list.

What Hábil does

At Hábil we build custom AI that works with the data and rules of your systems.

If you need to set how far AI goes in each process, the proposal is in Hábil AI. Hábil AI serves your customers, looks up your information and takes steps in your processes on its own, with a person approving anything that moves money or data. You decide, by process and by tool, how far each agent goes: only look up and alert, propose for a person to approve, or automate within limits you set. An agent that acts needs permissions limited to the bare minimum, a record of every decision it made and a way to prevent the same action from happening twice. And if an incorrect answer or action is detected, it is stopped, the log is reviewed, the rule or the integration is corrected and it is tested again before resuming.

The conversation can start with your decision table, even if it is half done: what the AI decides today, what you would like it to decide and who answers. With that, we review together which level suits each one. Write to us on WhatsApp; to protect your information, do not share data about your customers or your systems through that channel: we will agree in the conversation on the channel for the case.

Request your first conversation, at no cost.

Consulted on 13 November 2026. Each reference carries its confidence: high (primary source read) or medium (secondary source).