Piet Baudoin By Piet Baudoin, published on September 20, 2026, updated on September 26, 2026

How do I prevent AI from making mistakes and who checks it?

In brief

You prevent AI mistakes from reaching outside your business by not letting AI send or post anything without a person's approval, and by being able to review every step. The employee who currently does the work is responsible for checking. A mistake is corrected before it goes out, and that correction is remembered.

What does reviewing AI mean in practice?

Reviewing AI means that AI prepares the work and a person decides what happens to it. Each proposal includes the reason and source. Someone from your own business clicks Approve. You can look back at every step to see what happened and why. If that person improves a proposal, the correction becomes the rule for next time.

That approval must be a boundary in the program, rather than an agreement the AI can forget. Read why in the article on whether an AI assistant should send email on its own. Review then becomes the reason you can do more work, at a higher quality, with the same people.

How does the review work, step by step?

Review follows five steps, from incoming work to next time. Each example comes from administrative work.

Bombos reads incoming work and finds the information in your own software. A purchase invoice from email becomes a proposed entry with supplier, amount, VAT, project, and the reason for that project.
Purchase invoices someone retypes
Uncertainty becomes a question. If an invoice differs from the quote or receipt, Bombos first shows you the difference, with both lines next to each other, before anything is posted.
Invoices that do not match the quote or receipt
The employee who currently does the work clicks Approve. No payment reminder goes out without that approval. A customer who writes that they have already paid is referred to the coworker who checks it.
Reminders that go out late or never
Every action with an effect outside Bombos goes into a log. For a complaint, you can see what happened to the follow-up message after your approval.
Complaints that fade after the first reply
A correction becomes a rule. If you select a different project for a purchase invoice, the next proposal takes that into account.
Purchase invoices someone retypes

What disappears is the searching and retyping before every decision. Judgment stays with the coworker who knows the supplier, customer, and job.

Who is responsible if something still goes wrong?

The business using AI remains responsible for what it sends or posts, even if AI prepared it. Vendors say so themselves. Anthropic writes in the Claude Cowork safety guide: "You remain responsible for all actions taken by Claude performed on your behalf."

Microsoft writes in its shared responsibility model for agents that you always remain responsible for which actions require approval. It summarizes this as "Autonomy never reduces accountability." Microsoft presents this as an operational agreement, rather than a legal statement. A Canadian tribunal did rule on a case and held Air Canada to what its chatbot had told a customer.

The European AI Act does not currently have separate rules for agents. According to the Commission's FAQ, they fall under the existing definition of an AI system. Since August 2, 2026, a chatbot must say that it is AI. Rules for high-risk AI, the applications in Annex III such as assessing job applicants, apply from December 2, 2027.

Our assessment, rather than a legal prediction: the question then is whether you can show who approved what and why. Approval does not transfer anything to the vendor and does not release you from liability. It does produce that record. This is not legal advice. Put your own case to a lawyer or your insurer.

What does it cost and what does it deliver?

The pilot costs 2,000 euros for the first three months, with Bombos alongside you, then 1,000 euros per month. The approval boundary and log are always on. You do not need to arrange them separately.

What it delivers is fewer mistakes going out, because everything leaving the business first goes past a person. You see a reminder for a customer who already paid, or an invoice above the agreed amount, before the customer or supplier sees it. Bombos does the searching that currently makes that review so demanding. The same team does more work, at a higher quality, for less than half a day per week of the employee you would otherwise hire.

What does the person do?

The person decides. Bombos prepares. Before someone clicks Approve, they see the proposal with the reason and source: the invoice next to the quote, or the customer reply next to the payment status. If something is missing, Bombos asks you rather than guessing.

You can look back at every step to see what happened and why in a log you can access yourself. What happens to your information is explained on the privacy page.

Approval works in stages. At first, Bombos submits everything for review. For internal work, such as looking things up, checking, and preparing, you then decide how much Bombos can do ahead of time for each type of task, once it has gone well for a while. Customer messages and payments always wait for your approval. You can turn off each part separately. You cannot turn off that boundary yourself. Only Bombos can, at your explicit request and at your own risk.

Approval only helps if someone actually looks. The coworker who currently does the work therefore approves it. Bombos interviews them about what only lives in their head. Your team learns to work with AI along the way. They then teach Bombos the next task themselves, without needing technical skills.

When is this not a fit?

This is not a fit if there is no email stream being entered into software by hand, or if all the work already happens inside the software. In a very small business, most work still passes through the same hands. A separate approval step adds little.

Approval also does not solve everything. It does not help with a chatbot on your own website, because there is no approval in between. It does nothing if someone clicks Approve without looking. The other questions about AI in your business are in the overview.

Frequently asked questions

Can I turn off approval?

Not yourself. Only Bombos can move that boundary, at your explicit request and at your own risk. Outgoing messages and payments always wait for approval.

What happens to my information?

The Bombos platform runs on its own servers in Europe. Agreements about information Bombos processes for a pilot customer are recorded separately and personally. The rest is on the privacy page of bombos.ai.

Is the AI vendor liable if something goes wrong?

The vendors we reviewed place that responsibility on the business using AI. Anthropic writes this in the Claude Cowork safety guide. Approval does not change that, but shows who approved what. This is not legal advice.

What if I approve a wrong proposal?

Then you have sent or posted it, just as if a coworker had prepared it. The log shows what happened. Your correction becomes the rule for next time.

Do I need to check every proposal myself?

At first, yes. If a type of task goes well for a while, you decide whether internal work may continue without you seeing each step. Messages to customers or suppliers and payments still go past a person.

Does AI learn from my corrections?

Yes. A correction becomes the rule for next time, such as which project belongs to a supplier. Without corrections, Bombos does not learn your choices.

Leave your number and we'll call you back to talk through your situation.

A few lines about how things go at your company is enough, or just your number. You won't get a sales pitch but a conversation: what stands out to us, which work could get lighter first at your company and what your team can take on after that. It costs nothing. If we see something worthwhile together, we start with that first piece of work.