Corgent
Start
← All posts
AI for SMBsby Maximilian Koch5 min read

Why a good AI agent asks before it sends

AI agents now do real office work, but not all of it reliably on their own. Why human approval is not a brake, but the reason automation works for mid-sized companies at all.

illustration einer freigabe

A digital colleague has read a customer's request, looked up the order and drafted a reply. It is friendly, complete, and took two minutes instead of twenty. Should it just send it now?

The honest answer: not yet. Not because the reply is bad, but because sending is something you cannot take back. This is exactly where AI agents in small and mid-sized companies either become a reliable tool or a risk nobody wants to own.

Agents can do a lot today, but not everything on their own

AI agents have arrived in the German Mittelstand. According to the KI-Index Mittelstand 2026 by Salesforce and the Deutscher Mittelstands-Bund, 16.6 percent of the companies surveyed already use agents that take on tasks by themselves, up from 8.7 percent in 2024.

At the same time, research shows where the limits are. In Carnegie Mellon University's TheAgentCompany benchmark, agents had to solve 175 realistic office tasks, from research to coordinating with colleagues. The best agent tested completed 30.3 percent of them fully on its own (as of the version published in 2025). Models keep improving, but the pattern holds: agents reliably handle a large share of the groundwork, and get a smaller but decisive share wrong.

For a business, the question is not whether an agent makes mistakes. The question is whether a mistake leaves the building.

The real obstacle is trust

Ask companies why they hesitate with AI and you rarely hear that the technology is not good enough. In Bitkom's 2025 survey of 604 German companies with 20 or more employees, 53 percent name legal uncertainty as the biggest obstacle, and just as many a lack of know-how. 48 percent point to data protection requirements, 38 percent to results that are hard to trace.

All four share the same core: you do not know exactly what the AI does, and you cannot tell what you will be held responsible for. More autonomy does not solve that. It makes it worse.

The principle: the colleague drafts, a person approves

The answer is unspectacular, and that is why it works. Split the work in two:

  • Groundwork that can always be corrected: reading, summarising, researching, drafting, reconciling data. The agent does this by itself.
  • Actions that leave the company or cannot be undone: sending, paying, deleting, publishing. Here a person decides.

Action

Reversible?

Approval needed?

Read and summarise an email

yes

no

Draft a reply

yes

no

Send the reply to a customer

no

yes

Mark an invoice as checked

partly

yes

Trigger a payment

no

yes

Permanently delete a record

no

yes

This does not turn people into a bottleneck who redo everything. It makes them the final authority at exactly the points that matter. The twenty minutes of work are done; what remains are the ten seconds in which someone says "fine" or fixes a sentence.

What approval has to do with the AI Act

Since 2 August 2026, the transparency obligations in Article 50 of the EU AI Act apply. Anyone deploying an AI system that communicates directly with people must make clear that it is an AI, unless that is obvious anyway. For AI-generated text published to inform the public, Article 50 provides an exception where the text has been reviewed by a person and someone holds editorial responsibility.

So approval does not replace the transparency obligation. But it creates exactly what law and practice both ask for: a person who has checked, and a record of who approved what and when.

Approval has to take seconds, or it gets bypassed

The principle rarely fails on the idea, but on the execution. An approval that takes five clicks, three windows and a hunt for context will eventually be waved through without a look, or switched off altogether. Good approvals therefore share four traits:

  1. Everything needed at a glance. The draft and the evidence it rests on sit side by side. Nobody has to open the inbox or the CRM first.
  2. Edit instead of reject. If one sentence is off, fix it and approve. A "no" that restarts the whole task costs more than the work itself.
  3. Where people already are. A notification with a direct link brings the decision to the person allowed to make it.
  4. A trail that stays. Every approval is recorded with time and person. That helps with questions, audits, and learning where the agent still goes wrong.

This is how Corgent's digital colleagues work. They do the groundwork themselves and put every outward-facing action into an approvals inbox, with the draft, the evidence and the option to change something before approving. Every task is logged, and a failed task is not charged.

Five questions before you deploy an AI agent

Whether you build or buy, these questions separate a tool you can trust from one you have to babysit.

  1. Which actions can the agent take without approval, and can that be set per action?
  2. When approving, can I see what the proposal is based on?
  3. Can I edit a draft before approving, instead of only accepting or rejecting it?
  4. Is it recorded who approved what and when, and can I export that?
  5. What happens if nobody responds? Does the approval expire, or does the agent eventually act on its own?

If you can answer these with a clear conscience, there is no reason to fear AI agents. The software does the work; the decision stays with you.

Sources