Measurements, not opinions
What this blog is for: numbers I measured myself about what AI costs inside a company, how far off-the-shelf subscriptions go, and where agents break.
Most writing about AI in business falls into two piles. Vendor material, which tells you what a product does. And agency content marketing, which invents a price range so the article ranks. Both are easy to produce and neither survives contact with a real deployment.
This blog is the third thing: numbers I measured, with tools I pay for, inside processes I ran.
What you will find here
Measured. What a Claude or Codex subscription actually covers in daily work: where the limits hit, how much of a task the agent finishes on its own, and when NotebookLM over your own documents beats both. Not “saves hours” but minutes per task, with the plan, the model and the task stated.
Taken apart. A claim from a release note or a report, checked against a primary source. When the number does not survive the check, that is the post.
Deployed. What happens when you put agents in front of a team: the policy, the rollout, the metrics that turned out to be useless, and the thing that broke in week three.
The rules I hold myself to
- A number needs a primary or research source. Agency content marketing is not a source of fact, at most an indicator of what people are asking for.
- Every source carries its publication date. “In 2026” without a date is not a fact.
- Measurements state their conditions: tool, plan, model, the task itself. A benchmark you cannot reproduce is an anecdote.
- When I am wrong, the correction goes in the post, not in a comment.
Two languages, on purpose
Posts appear in English and Ukrainian. The Ukrainian version is not a translation. It is written for readers who work in a different market, with different budgets and different constraints, so the examples change.
If you work on this and disagree with something here, the disagreement is the point. That is what the discussion thread under each post is for.