skillcraft.All articlesRU

FROM THE AI FRONTIER

Frontline news — November 18: decide who authorizes AI to act

A recent AI Engineer recording helps separate preparing an answer from permission to act. We use a customer email as a teaching example.

Skillcraft editorial team · Max character · 2026-11-18

Max holds a draft letter beside a latched gate, an envelope and the headline “Who authorizes sending?”

What is useful in today’s edition? Let AI prepare a reply, and define permission to send it separately. Today I look at a recent talk recording about the difference between helping a person and a program acting on its own. For a beginner, the key question is: who makes the decision that changes something for the customer?

Diogo Almeida: assistance and automation have different requirements Channel: AI Engineer. Original title: I love you so jevving much. Speaker: Diogo Almeida, TypeSafe AI. Recording published: November 16, 2026, at 00:00 Moscow time. Language: English. Duration on the official page: 23:19.

This is a conference talk recording, not news of a product launching today. In the 3:32–4:23 passage, Almeida distinguishes tasks where a person assesses an answer from tasks where a program needs to use a decision. His example contrasts a useful research report with automated order taking: correctness plays a different role in each.

At 4:37–4:44, the speaker adds that even a task that is easy to verify and can run on its own still has a person check the result afterwards. This is his position in a discussion of AI reliability. It does not prove that all automation is impossible or that a particular service can already make reliable decisions.

An example: reply without accidental promises The following is an editorial teaching example. We have not tested this process, and it is not a result reported by the team in the talk.

Suppose a customer asks to move a meeting and keep the original price. For the first step, give the assistant the customer’s message and the current terms. Ask it to draft a reply and identify questions the supplied information cannot answer. Leave the decision about the new date and price with the responsible employee.

You could use these instructions:

These are working instructions. If the assistant is connected to email, reinforce the sending restriction through available permission settings: allow draft preparation and have a person send the message. Check whether your tool provides these settings and what they are called. A sentence in a prompt does not establish that a technical restriction is in place.

Where to place the decision I suggest dividing the process into three parts.

What the assistant prepares A draft, a list of missing information, and references to the terms you supplied. The benefit is a starting draft. The drawback is that reading and checking remain necessary.

What a person confirms The specific date, amount, and exceptions. Review the promises made to the customer, as well as the wording. People can make mistakes too, so compare the reply with the original terms.

What happens after confirmation The approved text is sent. If you later automate sending too, define the permitted cases and when the process must stop because information is missing. That is a separate process configuration to check against your own tasks.

What this talk does not establish The talk’s technical sections are outside this edition’s scope. We examine one passage about assistance and automation, not every claim about models or TypeSafe AI technologies. The automatic transcript may contain errors. We have not tested the product and do not generalize the speaker’s claims to all AI services.

Sources

From the workshop