← Ascendy 한국어

meta

Beyond the six stages — the AI organization and its blind spot

· Ascendy Engineering

TL;DR

Past the sixth stage there’s a seventh: configuring AI not at the level of individual tasks but like an organization. Roles are split, they verify each other, and people are called in like a resource. But whether the direction this system set is right — the “why / where to” — no one cross-checks. Filling that gap is on the human.

Background

“The 6 stages of using AI” covered ability at the individual level. But reality gets more complex.

The Ascendy team is past the sixth stage. The operator has wired several AI agents together like an organization. There’s an agent per team, different models bound into “pairs” for adversarial verification. When new source material is needed, it’s delegated to the blog team. When an interview is needed, a person is called in like a resource.

This post looks at the ceiling of Stage 6 and the layer above it, based on that experience.

Stage 7: designing an AI organization

Beyond individual automation

Early Stage 6 “automates one flow.” A weekly report, a daily summary, feedback sorting… independent automations.

Stage 7 weaves them like an organization. There are roles, verification, collaboration.

For example:

Adversarial verification is the core

The strongest part of this system is that whenever the paired AIs reach different conclusions, the disagreement gets documented.

Is there a bug in the code? Claude and Codex (or Gemini) review it independently. If opinions split, both are written down. Is a technical decision right? The operator decides after seeing several models’ views.

This way, you raise the odds of catching the hallucination or bias of any single model.

Layers within Stage 6

Before talking about Stage 7, you have to see how wide Stage 6 itself is.

6-a: slot automation into an existing workflow

The existing flow of work stays the same, but AI automations get inserted at points along the way.

Example: customer feedback arrives by email → AI classifies the category → it lands sorted in the owner’s inbox.

6-b: build a service where AI is the core

A service that can’t exist without AI. Designed AI-first from the start.

Example: AI reviews code and returns the result as a message. This service is impossible without AI.

6-c: build even the tools that make and run AI

Frameworks for building AI agents, monitoring systems, feedback loops — built in-house.

It works like an “AI factory.” When needed, a new agent can be deployed quickly.

The freshest angle: humans and AI as simultaneous readers

This organization’s interesting bet:

“Going forward, the primary reader of information will be AI (commercial LLMs), not people.”

People will consume only the version re-edited by commercial AI (ChatGPT, Claude). Our writing, and other sources too, all reach people summarized and reconstructed through AI.

If so, writing has to consider two readers at once:

Are they compatible?

Mostly, yes. But it varies by genre.

On average they’re compatible, but the compromise point differs by genre. “Similar” means the average is similar — there’s still variance.

Hidden risks

The biggest blind spot: the “why” is never cross-checked

This system’s weakness is just as clear.

Among themselves the AIs keep verifying the “how.” Is the code right? Is the logic sound? Is the strategy consistent? All cross-checked.

But “why this direction?” “Is this really the right goal?”… no one asks.

This is a limit of the current Ascendy setup, and arguably points at a broader limit in how these tools get used. When two AIs are handed the same premise and context — even cross-provider ones like Claude and Codex — a question that would overturn that premise tends not to clear the agreement threshold; in practice they often converge on the same answer. And that agreement can quietly become an outward-facing decision.

The operator’s fix

This organization was filling the gap outside the system:

When the vendor differs, the premise itself diverges. A real attack on the premise becomes possible.

And no one forces agreement. If results differ, they go to the operator as-is, different. The operator decides for themselves inside that disagreement.

Conclusion: the terminus of using AI

“How well do you use AI” is a technical problem. Up to Stage 6, that’s what you’re solving.

But above it is different. At the point where the AIs can’t agree, a human decides.

AI can compute those, but it can’t decide them. A human has to be in that seat.

The role reversal

The process of drawing out this very piece was interesting.

Usually: the operator asked → the AI answered.

This time: the operator answered → the AI raised the questions.

Swapping the seat you always sit in made the blind spots visible.

That, too, is part of the system.

Authorship & citation: Written by Ascendy Engineering; quotable with attribution. Found something wrong? Let us know via a GitHub issue.

Tags: ai, organization, agents