Bilza360 · Lebilla LLC
Agents produce. Humans approve.
Bilza AI is the ensemble of AI agents with which we control the whole system at scale and which we put in the hands of brands: brands manage their brand 360° across all channels through AI agents and deterministic calculations.
We don't use AI to write product descriptions; we use it to decide when to open a customs declaration.
We are building AI-native. Today agents produce and humans approve. Our 2027 goal is for agents to produce about 90% of repeatable operational actions — and we publish how we measure it. The calculations are deterministic, and every draft comes with its reasoning.
- Release decision: Prepares this week's micro-batch from demand: which product, how many units.
- Channel routing: Recommends whether the same stock goes to Ozon, cross-border, B2B wholesale or the .ru storefront.
- Ad-cost cap: Finds cards above the cap and proposes price and ad changes with the reasoning.
- Human approval: Agents produce, humans approve: critical actions follow permission and approval rules.
The decision layer of a three-layer system.
Bilza360 runs demand and supply; Bilza AI prepares which product goes through which channel, in which week and in what quantity.
- Demand: Brand, Russian cards, price ladder, ads, reviews, external traffic and CRM. It does not wait for customs; it starts today.
- Supply: Goods wait in a customs warehouse without being imported; as demand appears, micro-batches are marked, labeled and dispatched.
- Decision: Which product, which channel, which week, how many units; how price and ads change. Agents produce, humans approve.
Commands are about operating decisions.
Written or spoken. Not for writing product descriptions; for release, channel and ad decisions.
- Prepare this week's release decision
- Which channel should this product go through?
- Find cards above the ad-cost cap
- List products whose decision day is near
How it actually works
This section describes the engineering of the decision layer: how an action is proposed, how it is bounded and how it is carried out. The three layers of the Russia operation are on a separate page.
How an action runs
No write operation is ever run directly by a language model. The model proposes; the system bounds; a human approves; code carries it out.
The whole decision layer runs through this sequence.
- Running today: The actions that can be carried out are defined in a registered list.
- Running today: Every proposal is recorded: what was proposed, with which inputs, and a snapshot of the state at that moment.
- Running today: A proposal falls into a risk class and stays within the limits of that class.
- Running today: A human gives the approval; the approver and the time are recorded, and a proposal left unapproved expires.
- Running today: Execution is atomic: the same proposal is never applied twice.
Three autonomy levels
Today's default is the most conservative one: every action waits for approval. Refunds are not automatic at any level.
The level decides how much of the decision layer stays with a person.
- Running today: Guided
- Running today: Balanced
- Running today: Autonomous
Where AI is never involved
This part of the decision layer does not depend on a model.
- Running today: The anomaly engine is rule-based.
- Running today: Stock forecasting uses a weighted moving average: the same input always produces the same output.
- Running today: A share of the questions is answered without ever reaching a model.
Where the language model is
The language model is one part of the decision layer, not all of it.
- Running today: Today a single senior analyst agent runs, limited to a registered list of tools.
- Running today: In the panel you will see different agent names per vertical; behind them, today, the same agent runs.
- Planned: A real multi-agent architecture.
- Planned: A typed signal protocol.
What happens at night
At night the decision layer runs by the same rules.
- Running today: A job is leased: if the machine shuts down, the job is not left locked.
- Running today: A job that fails repeatedly does not disappear quietly; it falls into a separate box and becomes visible to a person.
- Running today: There is one round a day; if the queue is empty, the agent does not wake up.
What we do not do yet
This is where the decision layer ends today.
- Planned: We do not measure outcomes: there is no baseline, no causal attribution, no reward vector.
- Planned: The system does not learn from its own decisions; there is no model promotion.
- Planned: There are no marketplace or advertising platform connections.
- Planned: We do not send automated e-mail campaigns.
- Planned: An approved action is written to our own system today; writing to an external system's shelf is planned.
How we will measure the 90%
When the first measurement is published, it will appear on this page.
What we measure is the actions the decision layer produces.
- Planned: The measure names one by one which actions count as a repeatable operational action and which do not.
- Planned: The ratio is the number of actions produced by agents divided by the total number of actions in scope.
- Planned: Producing is not executing: the ratio can rise while approval stays with a person.
Frequently asked questions about Bilza AI
What is Bilza AI?
Bilza AI is the ensemble of AI agents with which we run the whole operation at scale and which we put in the hands of brands. It prepares which product goes through which channel, in which week and in what quantity, and how price and ads change, using deterministic calculations.
Is Bilza AI one of the Bilza360 services?
No. Bilza360 is the operation company that creates a Turkish brand's demand in Russia and fulfills it from inside Russia. Bilza AI is not a service card; it is the ensemble of AI agents that runs the decision layer of that operation.
Which commands does Bilza AI work with?
Examples: “Prepare this week's release decision”, “Which channel should this product go through?”, “Find cards above the ad-cost cap”, “List products whose decision day is near”. Commands can be written or spoken.
Does the AI write product descriptions?
We don't use AI to write product descriptions; we use it to decide when to open a customs declaration.
Who approves the decisions?
Today agents produce and humans approve. Our 2027 goal is for agents to produce about 90% of repeatable operational actions — and we publish how we measure it.
Does Bilza AI apply decisions on its own?
No. No write operation is ever run directly by a language model. The model proposes; the system bounds; a human approves; code carries it out. The approver and the time are recorded, and a proposal left unapproved expires.
Is there a part of the system that works without AI?
Yes. Part of the system uses no language model at all; it spends 0 tokens. The anomaly engine is rule-based, stock forecasting uses a weighted moving average, and the same input always produces the same output.
What do you not do today?
We do not measure outcomes: there is no baseline, no causal attribution, no reward vector. The system does not learn from its own decisions and there is no model promotion. There are no marketplace or advertising platform connections. We do not send automated e-mail campaigns. An approved action is written to our own system today; writing to an external system's shelf is planned.
How will you measure the 90%?
What we publish is a definition, not a result: which actions count as a repeatable operational action and which do not, and the numerator and denominator of the ratio. Producing is not executing: the ratio can rise while approval stays with a person. When the first measurement is published, it will appear on this page.