Becoming an AI-Native Company

Updated September 22, 2026 · both talks are in Russian; this guide is the English map

An AI-native company is not a company with more copilots. It is a company whose context, rules and decisions live in one knowledge graph that agents read, write and act on - and whose people can say what they think in one sentence.

The short version

This page follows the rule it describes: one sentence first, then three supports, then details.

Watch the talks

The thesis

AI changes the economics of execution, but the model is only one part of the system. The durable advantage is the layer around it: connected data, explicit context, shared rules for writing knowledge down, evaluation, memory, governance - and a human who can judge the result.

The operating rules

The graph only works if people and agents write the same way. These are the rules we actually use at Improvado. The full text is open source.

Rule 1 · Minto: one sentence first

Barbara Minto, one of McKinsey’s first women consultants, wrote The Pyramid Principle. The idea is simple: any complex idea can be explained first in one sentence, then in one paragraph, then section by section. The reader can stop at any level and still have the point.

Why it is a company rule, not a style preference

How to apply it

1. Answer first. The first sentence is the conclusion, not the context.
2. Then 2-4 supports. Not seven. If you need seven, you have not found the structure yet.
3. Details only where they change the decision. Everything else is a link.
4. In a chat, Minto works at the sentence level: the first line must survive alone.
5. We ask the same of agents: every document starts with a one-sentence thesis.

▶ 03.04 at 16:54 · 22.07 at 47:17

Rule 2 · How to write a document

Agents write a lot of documents, so we wrote a document about how to write documents ↗. Every agent and every person follows it. It fits into seven principles:

The skeleton

Thesis       one sentence that previews every section
Overview     diagram: things (nouns) or process (verbs)
§1 … §5      3-7 numbered MECE sections
  ¶1           states the ordering principle of the section
  ¶2, ¶3 …     numbered paragraphs, one idea each
Sources      link + YYYY-MM-DD for every claim
LLM analysis synthesis is always labelled as synthesis

Which diagram? Ask BFO

“When I analyse client calls, ground truth inserts the quote and the minute of the call, so I can listen to that exact quote.”

Watch · 03.04 at 20:06 ↗

▶ 03.04 · 16:44-20:14 · The full rules and the agent skill that applies them: Knowledge Framework ↗ · knowledge-framework skill ↗.

Rule 3 · The graph is code for the business

Code describes how a program works. A knowledge graph describes how a business works. So it inherits the same engineering discipline:

1. DRYThe hardest and most important task: a metric is defined in one place, not in fifty.1:17:53 ↗
2. MECE at every levelLevel one is departments. Inside marketing: acquisition, conversion, experiments. And so on down.1:18:30 ↗
3. BFOTimeless things take part in events in time. Two taxonomies, not one tangle.1:19:02 ↗
4. KimballOne big table of all company events that every agent can query.1:20:13 ↗

Skills are evil: instead of hundreds of conflicting prompts, we keep one knowledge graph. Reusable knowledge belongs in the graph.

Watch · 22.07 at 1:01:53 ↗

Test yourself in five minutes

Before you send a document, a status update or a hiring homework, check it against the rules:

• Can I state my conclusion in one sentence - and is it the first sentence?
• Are there 2-4 supports, and does each one actually prove the conclusion?
• Do the sections overlap? Together, do they cover everything?
• Did I repeat a fact that already lives somewhere else?
• Does every claim have a source and a date?
• Is there a diagram - and does it show things or a process?

Transcript map · July 22 talk

Every timestamp opens the exact point on YouTube.

1. 00:00
   Daniel and ImprovadoThe company, the product, and what it means to dogfood an AI-native operating model.
2. 05:01
   Enterprise marketing OSImprovado's connector foundation and Secure MCP as one governed door for every agent.
3. 09:10
   Token metabolismHow agents continuously keep the company knowledge graph synchronized with reality.
4. 12:41
   From one event to coordinated actionA call becomes a cascade of actions; a campaign can cross 18 systems in 90 minutes.
5. 19:17
   Experiment and introspection loopsThe experiment registry, competitor analysis, company beliefs, and the daily Improvado Pulse.
6. 29:44
   The new economics of executionWhy model capability changes the bottleneck, while integrations become the system's eyes, hands and moat.
7. 43:27
   Redesigning the teamEngineers replacing conventional functional roles.
8. 47:17
   Minto and causalityWhy the pyramid principle is a company rule, and why people who cannot use it get persuaded by agents.
9. 1:01:08
   An agent is a harnessModels, hooks, tools, memory and governance - and why reusable knowledge belongs in the graph.
10. 1:17:53
    Ontology as operating infrastructureDRY, MECE at every level, Basic Formal Ontology and Kimball modelling applied to a company-wide event stream.
11. 1:31:08
    Live proof and durable moatsA real-data client demo, then the case for integrations, graph consistency and iteration speed.

Where each idea appears

Key concepts

Keep reading