The first agent is always the guardian of the brand.

Before a department writes, it learns to check. The first agent we build produces nothing. It guards how the brand sounds, what it claims and what it would never say. Everything else has to pass through it, and only then does a human see the draft. That sounds slow, and it is the opposite. A department that is allowed to do everything at once produces drafts nobody wants to approve. A department that knows the brand produces drafts that pass. The order decides. Behind the image of the gate sit two checkpoints with fixed questions. One checks the voice, the other checks the claims. Both work on the same brand knowledge that gets built in the first month.

What the guardian checks

Behind the image of the gate sit two checkpoints with fixed questions. One checks the voice, the other checks the claims. Both work on the same brand knowledge that gets built in the first month.

Every checkpoint knows exactly four answers: pass, revise, block, or hand it to a human. There is no vague middle. That makes the image of the gate concrete. A draft is either in front of it or behind it.

Disclosure belongs to the same gate for me. Where AI was involved, it says so openly, as agreed with the client. Hiding it comes out eventually, and then it damages exactly what the guardian is there to protect.

  • Does the text sound like your company or like generic AI text?
  • Does it contain wording you have ruled out for your brand?
  • Is every promise backed, and is the evidence attached?
  • Are numbers and client results placed in the right context?
  • Is it disclosed where AI was involved, as agreed with you?
  • Does the draft fit the rules of the platform it is meant for?

The order of the checks

From the outside the flow is plain. The agent puts a draft down. The checks run automatically, up to 21 automatic checks for a newsletter, 48 content checks and 26 platform checks for a blog post. Only then does the draft appear on the approval platform, and someone on your side clicks approve. That rule has its own name and its own note: draft before send.

  • The agent reads the kontext: what is going on in the company and what is coming up.
  • It works from the handbook of its department, written for this exact task.
  • It checks itself against its own brief.
  • Then come the two gates of the guardian: first the voice, then claims and evidence.
  • After that risk and legal.
  • At the end a human who approves, and a piece of feedback that stays in the system.

The trial by fire

On 21 July the content lead at DSB-ONE / MyGym reviewed the first newsletter drafts with the full brand standard: tone, local references, imagery, every claim. She gave feedback, a lot of it and precise. That is exactly what the system was built for.

A tool makes the same mistake twice. A department gets better with every check.

Since then the agents write in the language of the brand, with backed claims. Approval still happens with you. The guardian makes approval easy, it does not replace it.

In the running mandate it exists three times, once per brand: MYGYM, MYGYM PRIME and DSB-ONE, each with its own knowledge library and its own reviewer. And the gate is open to humans too. If you want, you can decide that your own handwritten posts run through it once before publication. What that looks like day to day is in the case-study.

The question underneath is keeping half the industry busy: who owns responsibility when an agent acts. KPMG writes about accountability for AI agents as well. My answer is a system of roles, rules and gates that takes hold again with every draft. Why that carries further than a good prompt is in prompting or systems.

  • 21 July 2026: The first newsletter drafts come back, with plenty of precise feedback on tone, local references and imagery.
  • 26 July 2026: Every piece of feedback has become a fixed rule in the system. The new versions pass every comparison.
  • 28 July to 4 August 2026: AI disclosure is agreed and becomes one of the checkpoints from then on.

What this means for you

When you start a department, the first question is not what it should write. The first question is how you recognise a good draft. Whoever can put that into rules has almost built the guardian of the brand already. That is what the strategy workshop is for. What an ki-department costs per month after that is in the note on the cost of a department.

This note keeps growing

2026-09-03: Grown: the six check questions, the four answers of a gate, the order of the flow, a timeline of the trial by fire, one guardian per brand, a source list.

2026-09-02: Loop drawn, section "What this means for you" added.

2026-08-24: Planted from the trial by fire.

ZukunftBilden GmbH · Salzburg · +43 681 81655313 · office@zukunftbilden.eu