OData in 1C — An Underrated Entry Point for AI
In April 2026, Anthropic launched Claude Managed Agents in public beta, OpenAI expanded MCP support in the Responses API, and monthly downloads of the MCP SDK surpassed 97 million. Forrester predicts that 30% of corporate SaaS vendors will release their own MCP servers this year. The world is systematically building an infrastructure where AI agents can read corporate data and take actions — not through ChatGPT in a browser, but directly from business processes.
This shift is passing by most Kazakhstani companies for one simple reason: they cannot connect an AI agent to their data because they don't have any data. More precisely, the data exists — it resides in 1C, which is installed almost everywhere — but it cannot be extracted in an automated way. Not because the technology is complex, but because the layer that turns 1C into a source for any AI tool is not built. This is OData. And in our view, from the practice of West Star Ltd in integrations with 1C, it is the most underrated lever of automation in Kazakhstan today.
Why 1C is a Bottleneck for AI in Kazakhstan
According to 1C on 1c.kz, more than a dozen localized configurations are supported in Kazakhstan: "Accounting for Kazakhstan", "Trade Management for Kazakhstan", "Payroll and HR", "Retail", "Document Management CORP", "Accounting for State Enterprises of Kazakhstan", "1C:Cashier". Releases are issued every two weeks — in the first week of May 2026 alone, updates were released for at least four Kazakhstani configurations. This infrastructure supports the overwhelming majority of LLPs, individual entrepreneurs, state enterprises, and large holdings. By our estimates, more than 80% of all business data in Kazakhstan that has commercial value physically resides in 1C databases.
What kind of data is this? Contractors with details and transaction history. Nomenclature with balances, prices, barcodes, characteristics. Documents — orders, sales, receipts, invoices. Cash flow movements through accounts and cash registers. Payrolls and HR records. Tax reports. Production analytics. This is the very "corporate memory" without which an AI agent is useless.
When a Kazakhstani company orders an AI project — a chatbot for customers, a sales manager assistant, an automatic invoice verification system, a stock forecast — the first question becomes: where will the model get the data from? And here a typical scene occurs. An accountant exports Excel from 1C once a week. A sales manager spends 40 minutes a day copying numbers from one system to another. Each AI solution requires manual export, and automation turns into a complicated version of the old process.
This is the wall. On the other side of the wall are all the possibilities of "AI agents on top of corporate data" that the Western press writes about. On this side — Excel and copy-paste.
What is OData in the Context of 1C
OData is an open data access protocol developed by Microsoft in 2007 and standardized by OASIS. Essentially, it's a REST API on top of any structured database: the request goes through HTTP, the response comes in JSON or XML, filters and sorting are specified in the URL. Nothing exotic — the same principles on which modern web services operate.
What is important: since the release of platform 8.3.5 (i.e., since 2014), the automatic REST service OData is built into the 1C platform itself. It is enabled with one command in the Configurator or through publication on a web server. It does not require configuration modification, does not require the development of special export modules, requires nothing but proper access rights setup. Any localized Kazakhstani configuration — from "Accounting" to "Enterprise Trade Management" — supports OData out of the box.
When the service is enabled, any database objects — directories, documents, registers, enumerations — become accessible via HTTP. A request like GET /odata/standard.odata/Catalog_Контрагенты?$filter=... returns JSON with contractors by the required filter. The request GET /odata/standard.odata/AccumulationRegister_ОстаткиТоваров/Balance(...) returns balances as of a date. Creating a document is a POST with a body in JSON. This is not a "crutch" or a "hack" — it is a standard mechanism that the firm "1C" itself recommends for integrations. And it remains massively unactivated in Kazakhstan.
What Changes When OData is Enabled
The picture after connecting the OData layer looks fundamentally different. An AI accountant in a Telegram bot does not ask the user to send photos of acts — it queries 1C itself to see which documents have arrived in the last month and what details are missing in them. The warehouse system in real-time answers the manager's question "do we have two hundred tons of A500C rebar with a diameter of 12 in stock". The stock forecast for the next month is calculated by an AI model that pulls the latest nomenclature movements every hour. The company's internal chatbot answers the CFO's question "how much did we pay this contractor for the quarter and how does it compare to the contract limit" — in five seconds, without the accountant's involvement.
This is not futurology or marketing. These are technical scenarios that work today — in our practice, with our clients, on the same 1C databases that are installed at thousands of Kazakhstani companies and are not used beyond their interface.
The emergence of MCP infrastructure in 2025–2026 amplifies this argument many times over. MCP — Model Context Protocol, an open standard from Anthropic — turns any data source into a tool that an AI agent can see and call. On top of the OData endpoint of 1C, a thin MCP server is written (literally a few hundred lines of code), and after that, Claude, ChatGPT, Gemini, or any agent in Microsoft Copilot Studio can read your 1C database just like they currently read Slack or Gmail. Without six-month integration projects. Without rewriting configurations. Without specialized connectors for each new model.
The key thought: OData is the basic layer without which MCP and AI agents, in principle, do not work on 1C data. Those who built the OData layer two years ago are now simply adding an MCP wrapper and gaining instant access to new technology. Those who did not will learn about MCP around the same time their competitors tell them about it — at an industry conference in 2027.
The Real Economy of the OData Layer
The most unobvious thing about this topic is how cheaply all this is done. Enabling OData on a working 1C database is a 1–2 working day job for a specialist who knows the platform. Publishing on a web server (IIS or Apache), setting up authorization, restricting access rights to sensitive sections, testing the main endpoints. In the RK market — 80,000–150,000 tenge for a one-time connection.
What you get in the end. A full-fledged REST API to your own accounting system. Access for any external applications — Telegram bots, mobile apps, BI systems, AI agents. The ability not only to read but also to write data back — that is, to automatically create documents from external sources. An analytical dashboard that updates in real-time without exports. Integration with Kaspi and Halyk marketplaces, warehouse services, banking APIs. The ability to connect ChatGPT/Claude via MCP within a month when needed.
Alternative paths are orders of magnitude more expensive. Custom development of a specialized web service in the 1C configuration — 800,000–2,000,000 tenge per project with a monthly term. External ETL systems with regular cloud export — from 300,000 tenge for implementation plus a monthly subscription. Custom connectors for specific tasks — each time from scratch, each time with support, each time with the risk that they will break when 1C is updated.
The comparison is not in favor of alternatives. OData is built into the platform, updated with it, supported by the firm "1C" as a standard mechanism. It is an infrastructural solution — set it up once, use it for years, layer any new scenarios on top of it.
Where This Approach Has Real Limitations
Ignoring weak spots means preparing for the same surprises that companies that rushed with MCP implementations in the West are now facing.
Security. Standard OData in 1C is an HTTP service with basic authorization. If you just publish it on the internet with default settings, in two weeks you will have brute-force attack attempts, and possibly — real access to the database with "Administrator" rights. The minimum required: HTTPS on the web server side, a separate 1C user with limited rights for each connection, white lists of IP addresses, control over which objects are exposed outside. This is not difficult but requires conscious work. Giving someone OData access as "try it, I'll send you a link" is a guaranteed way to leak accounting data to the public network.
Performance. The 1C OData service builds SQL queries automatically, and on large data volumes with unoptimized filters, it can slow down the database. A request "give all sales for two years" through OData without server caching can crash the production server. In real integrations, you either need to use incremental loading (read only what has changed since the last synchronization) or build an intermediate cache layer between OData and the AI agent.
Protocol limitations. OData works great for CRUD operations on standard objects but is poorly suited for complex business operations requiring the execution of procedures and algorithms within 1C. If you need to "process a document with posting to all registers and checking the accounting settings by 17 parameters" — this is not a task for OData, this is a task for a separate web service (HTTP service in the configuration). In practice, a hybrid is used: OData for reading and simple writing, HTTP services — for complex logic.
Localization features of 1C Kazakhstan. Kazakhstani configurations are regularly updated under changes in the Tax Code, reporting forms, regulator requirements. Within updates, the structure of details sometimes changes, mandatory fields are added, objects are renamed. If your OData client is tightly bound to the names of details and has no protection against changes — a 1C update can break the integration. The solution is a normal CI test on the data schema and a buffer layer that isolates external consumers from the internal structure of the database.
Dependence on data quality. An AI agent reading from 1C via OData inherits all the problems of the source data. If there are five contractors in the database with different spellings of the same company name, the bot will confidently answer at least five different questions about one client. If the balances do not match the actual ones — the model will give a beautiful forecast on top of nonsense. AI does not fix data. AI uses it. Therefore, before connecting an agent to 1C, it makes sense to conduct basic directory hygiene.
What to Do
The shift happening in 2026 is not about specific models like Claude, ChatGPT, or Gemini. It's about AI for the first time having a standardized way to read corporate systems. In a year or two, this will cease to be a feature — it will become the norm. The only question is whether Kazakhstani business will catch this wave with its data or remain the one who orders "innovative AI projects" but sends an Excel file manually with each model request.
What makes sense to do in the near future. Check the version of the 1C platform — if it is 8.3.5 or higher (and it is practically for everyone now), OData is available. Formulate one specific task for which external data access is needed: balances in a Telegram bot, a dashboard for management, a manager assistant, invoice verification. Enable OData in the test environment, write the simplest request, look at the response. This exercise is done in half a day and immediately gives an understanding of whether the database is ready for integrations.
Then — design the security layer. Do not publish raw OData on the internet. Set up HTTPS, a dedicated user, access rights restrictions, reverse proxy with logging. Fix in the company policy which data can be given to external consumers and which cannot.
And only then — connect AI agents, MCP wrappers, chatbots, predictive models. By this time, the basic infrastructure will be ready, and each subsequent solution on top of it will take not six months, but a week.
In our products — AI accountant, OData Hub, warehouse integrations — we are built exactly on this idea: 1C ceases to be "an accountant's program" and turns into a data source for any overlays. Previously, building such a layer required a serious budget and was justified only for large companies. Now, in May 2026, it costs less than an annual subscription for one user in an average SaaS. And this is the point from which it makes sense to start any conversation about AI automation in Kazakhstan.