AI agents
How do you connect 1C and ERP data to an AI agent?
Connect 1C and ERP data to an AI agent: read-only first, then proposals, approval and execution. Steps, model choice and pilot timeline.
In this article
Short answer: First, a connector is set up. It gives the AI agent access to your 1C, ERP or mail system, and only to what you allow. At the start, the agent only reads. It reads the 1C or ERP data you allow and prepares the result. Sending, writing and any other change runs only with your approval. The connection starts as a pilot with one request and one workspace.
Connector is the component that lets the AI agent enter your 1C, ERP or mail system, only as far as you allow. The connector is set up on your side.
Why doesn't the data in 1C give quick answers?
Many companies in Azerbaijan use 1C for accounting and operations. The data is in the system, but the answer to a question is often spread across several places. The question "Which customers have not paid this month, and who should talk to them?" has its answer spread across the receivables balance, the payment schedule and document headers.
To find the answer, several reports are opened, copied to Excel and put together by hand. The same work is repeated for every new question.
The AI agent is not there to replace this chain. It shortens it. It reads the data and gives the result as a written draft. The decision still stays with you.
How does the connection start?
The connection has four stages:
- AI readiness check (30 minutes, free). No access to your systems is needed. Together we identify which 1C configuration or ERP you have and which question takes the most time.
- AI readiness map (1 business day). 1–2 pages: where your data is, what can be connected, and what the first pilot needs.
- Pilot starts (within 7 days of agreement). One request, one workspace, read-only.
- Result (30–60 days). Based on the pilot result, the next step is your decision.
How does the agent read data and what does it do?
For example, the sales manager at Sample Distribution LLC asks: "Which payments are expected this week?" The process goes like this:
- Question. The question is written to the agent.
- Read. The agent reads the 1C tables you allowed: the receivables balance and the payment schedule.
- Prepare. The result is shown as a list: two customers this week, 15,950.00 AZN in total.
- Draft. If needed, the agent also drafts a payment reminder message.
- Approve and execute. Sending happens only after you click "Approve". Every step is written to the log.
The difference is in the source of the answer: the numbers come from your own database, not from general knowledge.
Propose, approve, apply: what is the rule?
The agent can do three things: read, calculate and draft. Writing and sending happen only after approval. This rule is the same for every system. A document correction, an order proposal and a reminder all go through the same path.
So at the start, the agent writes nothing to your system. It proposes first. You review the proposal, and then it is applied.
Model choice: cloud or your own server?
You choose which model runs. There are two options:
| Leading cloud model (OpenAI, Anthropic, DeepSeek and others) | Local model | |
|---|---|---|
| Where it runs | On the cloud model you choose | On your GPU server |
| What goes to the model | Snippets related to the request. Your whole database is not sent | The request does not leave your company network |
| When it fits | When your company rules allow a cloud model | When keeping data inside the company network is the priority |
In both options, the connector is set up on your side.
How is control over your data set up?
- The connector is set up on your side.
- The agent reads only the data you allow.
- With a cloud model, only snippets related to the request go to the model, not the whole database.
- With a local model, the request does not leave your company network.
- Every request, source and approval stays in the log.
What do you need to prepare?
- System. 1C configuration or ONYX ERP. If you have an older version, we check during the AI readiness check whether a connection is possible.
- One painful question. For example, receivables, stock or quotes. One question is chosen for the first pilot.
- Responsible person. The person who will read and approve the agent's draft.
- Model decision. Cloud or your own server.
We decide together during the AI readiness check which of these items are ready and which need more research.
AI readiness check. 30 minutes, free. No access to your systems is needed. A conversation is enough. Within 1 business day after the call, we send you an "AI readiness map": where your data is, what can be connected, and what the first pilot needs. The pilot can start within 7 days.
Frequently asked questions
I have an old 1C version. Can it connect?
We check whether a connection is possible during the AI readiness check. The check is free.
Can the agent change my data?
At the start, the agent only reads. Every write action runs only after your approval.
Which model is used?
You choose: leading cloud models such as OpenAI, Anthropic or DeepSeek, or a local model on your own GPU server.
We have no system and our data is in Excel. What can we do?
If you have no system, you can start with ONYX ERP. The ERP and the agent are built together, so your data is in a system from day one.
Which data does the agent get access to?
Only the data you allow. We decide together during the AI readiness check which tables will be opened.
How long does it take?
The pilot starts within 7 days and runs for 30–60 days. One request, one workspace.
For the internal logic of the connection, see How the AI agent works. If your process does not match the standard roles, see Custom agent. For where the model runs, read Where does your data stay? Cloud and local AI models. For how a pilot starts, read AI pilot in 7 days: what do you need to start?.