AUGMANITAIAndreas Ehstand

Making decisions with AI

How does a plan become a promise?

In this fictional art sale, three AI agents pass on a seller’s note. What does the buyer actually read?

Three dark glass plates and a gold line of light that breaks between two plates.
AI-generated illustration

Twenty seconds

Catch the missing word

In this fictional example, three AI agents pass a sentence along. Watch what happens to the original intention.

Step 1 of 5

Read it once.

This sentence comes from an invented offer for a painting. Three AI agents will now pass it on. One summarises, one prepares a briefing, one writes the reply.

  1. The seller’s note · invented

    We plan to sell the painting in five years.

  2. Summary agent

    The painting will be sold in five years.

  3. Briefing agent

    Proceeds from the sale arrive in five years.

  4. Reply agent

    You can get your money back in five years.

From the talk

Zurich, 11 November 2026

About the Zurich talk

Plan or promise?A fictional art sale shows how an AI summary changes a tentative plan into a promise of repayment.Original English audio
Transcript

When there is an offer, we plan to sell the painting in five years. There could be an agent. You say, okay, that's my summary agent. And what he's doing, he says, you can get your money back in five years. Okay, that's just the incorrect summary, and the word plan isn't there.

Start with small stepsHow to begin working with AI in small steps.Original English audio
Transcript

It can be very easy, but you have to start in very little steps.

Find people who love the workA short observation on choosing who will do the work.Original English audio
Transcript

So I would say, find people who love this work.

A memory of thinkingAn idea for preserving the thinking behind a decision with local AI.Original English audio
Transcript

And now we have the chance with local AI to make a memory of thinking.

Three ideas I work with

What people decide at each handover

These are personal views, not rules. They come from a single-case study of my own work with AI systems that has been running since 2023: several hundred thousand turns by my own count. The examples are constructed.

I

The memory of thinking

Most records I come across keep what was decided and let the thinking go. I write decisions down with three things attached: the date, the reason, and the experience the reason came from.

To me this is stewardship in a plain form. The next generation receives the reasons along with the decisions.

Constructed example · a family office

The day of the decision

Without a record

Everyone in the room knows why. Nobody writes it down.

With date, reason and experience

14 March. We decline the co-investment. Reason: a ten-year lock-in, and the next generation asked for room to move. Experience: we were held in a similar structure once and could not leave when we needed to.

II

The human inside the swarm

One person now directs many agents. I do not picture that person above the swarm, signing off at the end, or outside it, reading reports.

I picture the human inside it, at the places where meaning can tip.

A personal view · not a measurement

Three agents. I can read every line they write. Oversight is reading.

III

The decision behind the decision

“He declined the stake” is what happened. It is not the decision.

What a successor, a board or an agent needs in order to act in the same spirit sits one or two questions further down.

Constructed example

  1. What happened

    He declined the stake.

On the record

Where this work has appeared

Andreas Ehstand is the founder of AUGMANITAI and an independent researcher and author. He works on how people and AI systems collaborate, how humans oversee AI agents, and how the knowledge behind decisions is handed on. His background is in competitive tennis coaching and performance analysis.

“I was especially pleased that with Andreas Ehstand we could welcome another proven AI expert to the BVMID.”

Dr. Thomas Graf, BVMID, in a public LinkedIn post, August 2026. Translated from German.

Explore the publications and sources
  • Prestel & PartnerFamily Office Forum Zurich 2026, The Dolder Grand. Partner and speaker, 11 November 2026.The talk on this site
  • Swiss Future InstituteMember of the AI Future Council.
  • BVMIDListed as Top Expert 2026 for IT and artificial intelligence in the expert directory of BVMID (Bundesvereinigung Mittelstand in Deutschland), an association of medium-sized German businesses.
  • IETF Internet-Draft“Terminology for Human Oversight Acts in Automated and Agentic Systems” (draft-ehstand-oversight-acts). An Internet-Draft is a working document, not a standard.The four terms explained
  • Private Wealth VoicesEssay “One Missing Piece of Legacy”, July 2026 edition (PDF).
  • CXO BoardroomArticle “The Approval Looks Right, But the Evidence Is Missing”.
  • Swiss Board ForumArticle on AI in board papers and what a board needs to be able to follow (in German).
  • The European ScientistArticle “An AI explanation is not evidence of human review”.
  • business-wissen.deArticles on handing over experience and on checking AI results (in German).

The full list with sources is on the profile page.

Working together

Three ways to begin

Keynote or talk

The audience tries the missing-word example for themselves, with examples built for the room, whether that is a board, a family office or the next generation. The talk is meant to leave them with a question they can put to their own decisions.

Working session with your team

We follow one of your own decisions on its way through people and agents and look at what arrives. The session looks for the places where reasons drop out, and sketches what a record with date, reason and experience could look like in your setting.

Research exchange

For people who work on human oversight, agent systems or the hand-over of decision knowledge. We compare terms, cases and open questions, and find out whether there is something worth writing together.

What this is not

I am not a lawyer, a tax adviser or an investment adviser, and I do not assess or certify specific systems. What I bring is a way of looking at decisions, drawn from my own research and practice. It rests on a single-case study and on constructed examples, so whether it fits your setting is something a first conversation has to show.

How it begins

A first message needs only the occasion, the audience and a rough time frame. Confidential documents are not needed for it. Format, scope and terms are agreed personally. Nothing is sold or booked through this website.

Start a conversation

A few lines about your setting and your question are enough to begin.