Pailot
How is a company AI agent different from a chat tab?
| In the company, Pailot | A chat on one computer | |
|---|---|---|
| Where it runs | On-premise, on hardware the company controls | On one laptop or in a vendor cloud |
| What it does | One process step: files, a draft, a CRM card | A reply in the same window |
| How you publish it | Web chat first, then API or A2A | Open the chat and type |
| Log | Every run is in a shared journal | Usually a private thread |
What do people usually ask for?
People ask for an AI agent and mean a cloud chat. Here the agent sits on the canvas next to conditions, RAG (answers from your files), speech and channels. Each run is logged.
Pailot is built by ООО «Профитайзер» (Russia, since 2022). The public site is https://pailot.ru.
How do you create the agent?
Open the copilot, describe the job. It places blocks and links. You fix a wrong step by hand and run it yourself. There is no self-serve agent store.
- Describe the job in plain language.
- The copilot places blocks and links.
- You fix a wrong step by hand and start the run yourself.
- Domestic channels in the box include Bitrix24, amoCRM, MAX, VK and Telegram.
Where does an on-premise agent run?
PailotGPT can run on the customer equipment so process steps do not have to call an external API. Document search and speech can stay in the same contour.
Where can the agent answer?
The fastest path is web chat: a link, a person types in the browser. Your own system uses an HTTP API. One process calling another uses A2A. Channels are listed below.
- Web chat: a link, a person types.
- API: your system starts the process and reads the result.
- A2A: one process calls another.
What is in the box
Core of the delivery, not a connector list. Integrations sit under this.
- Copilot
- Drafts the process from a brief, checks links and publishes. At install time you choose a cloud copilot with redaction or a copilot that stays on your servers.
- Language model
- An LLM runs the agent step: reads text, decides, writes the reply. In the contour that model sits on the customer’s hardware.
- Document search
- RAG answers from your files and shows the source. The index stays in the contour.
- Speech and diarization
- Speech-to-text transcribes audio and video. Diarization splits speakers. A meeting can become a protocol or a reply in the same thread.
- Document parsing
- PDFs, spreadsheets, audio and video are parsed on-site. The file does not have to leave for a third-party parser.
- Browser step
- The process opens a live page and works like a clerk when no connector exists.
Domestic connectors in the box: Битрикс24, amoCRM, MAX, VK, Telegram. The rest come from the catalog (162 blocks and connections) when they already live in the company contour.
Where an agent can answer
| Channel | What it does |
|---|---|
| Web chat (fastest) | The fastest way out: the process becomes a chat page. A person types, the agent answers. Use it before wiring other systems. |
| API | Your system starts the process with a request and reads the result. |
| A2A | Agent-to-Agent: publish the process so other agents can call it over the A2A protocol. |
| MCP | Model Context Protocol: publish the process as a tool another assistant can call. |
Questions
- Is this a ChatGPT wrapper?
- No. The agent is a block on a canvas with a run log. External models are optional and can be redacted before the request leaves the contour.
- Can we self-host?
- Yes. That is the default delivery story. Hardware and scope are scoped after an access request.
- How do we get access?
- Submit «Получить доступ» on the site. Price is on request.
- Is there a self-serve agent store?
- No. Access is by request. On a demo we assemble one narrow process.
- What is the fastest way to try an agent?
- Web chat: a link, a person types, the agent answers. API and A2A come later.