How it works · data · model · approvalHow an AI agent works.

The agent reads and prepares drafts. Execution happens only with your approval. Every step is written to the log.

Scroll: six steps
01 · Systems

Your systems stay where they are.

1C, ERP and mail are not moved anywhere.

YOUR NETWORKCLOUD · INTERNETquery-relevant part onlyanswerdraftexecuted after approval1CERPMailConnectoron your serverREAD-ONLYCLOUD MODELOpenAIyou chooseOpenAIAnthropicDeepSeekDRAFTPrepared by the agentAPPROVALYouedit · cancelApproveLOG09:14Query09:14Source09:14Draft09:15Approval: you09:15ExecutionYOUR NETWORKCLOUDquery-relevant part onlyanswerdraftexecuted after approval1CERPMailConnectoron your serverread-onlyCLOUD MODELOpenAIyou chooseDRAFTPrepared by the agentAPPROVALYouApproveLOG09:14Query09:14Source09:14Draft09:15Approval: you09:15Execution
Only the part needed for the query goes to the model.
  1. Your systems (1C, ERP, mail) stay where they are.
  2. The connector is set up on your server and only reads.
  3. You choose the model: a cloud model (OpenAI, Anthropic, DeepSeek) receives only the part needed for the query; a local model runs on your GPU server.
  4. The agent prepares a draft.
  5. You approve. Nothing is executed without approval.
  6. The query, source, draft, approver and time stay in the log.

The agent prepares. Execution is in your hands.

What it does
  • Reads
  • Calculates
  • Writes document drafts
  • Proposes actions

Read-only mode at the start

Not without approval
  • Writing1C · ERP
  • Sendingmail
  • Orderingsupplier
  • Document changesdelivery note · contract

Nothing is executed until “Approve” is pressed

Log

Who, what, when.

Every query, the source read, the draft prepared, the approver and the execution time are visible in the log.

QueryReceivables over 30 days
Source1C · Accounts receivable balances
DraftReminder letter
ApprovalFinance director
ExecutionLetter sent

We start with one query.

How we run this ourselves

Result: an AI readiness map for your company within 1 business day. No access to your systems required.

  • Scope1 work areaone agent, one request
  • ModeRead-onlyexecution with your approval
  • Duration30–60 dayspilot duration
  • Startwithin 7 days
  • PriceFixedagreed in advance

Questions.

We define the exact configuration together in the readiness check.

AI team: main page
What if the agent makes a mistake?

No draft is sent without approval. That is exactly what the approval gate is for. The log shows what happened.

Which systems does it connect to?

1C, ONYX ERP, databases and mail. The exact configuration is confirmed in the readiness check.

How does my data reach the model?

With a cloud model, only the parts relevant to the query are sent to the model, not your whole database. The local model runs on your GPU server and the query never leaves your company network.

Who chooses the model?

You do: leading cloud models such as OpenAI, Anthropic, DeepSeek, or a local model on your own GPU server. We decide the option together in the readiness check.

What can the agent do at the start?

It only reads, calculates and prepares drafts. Writing, sending, ordering and document changes are executed only with your approval.

Test it in your own system.

First, a free 30-minute AI readiness check. The pilot starts within 7 days.

First task · optional

Result: an AI readiness map for your company within 1 business day. No access to your systems required.

  • Read-only
  • Connector on your server
  • You choose the model
ONYX · AI agents · How it workssales@onyx.az · +994 12 310 20 40