AI agent development services: chatbots and agents inside your product

On our own CMS, an AI agent that tries to publish a page without a person's approval gets an HTTP 409 from the server, and nothing is published. That is the idea behind our AI agent development services: the agent can search, draft and propose, and the server decides what it may save, publish or change.

We put a chat or an agent inside a product that already has users: a catalogue, a B2B platform or an internal system. This work is part of our AI software development.

A project covers:

• a chat or an agent working on your product's data
• a separate AI service that your application calls
• tools the agent can use, each with a description and a schema
• rules enforced on the server
• tests that try to push the model off its task
• support after launch

The server side has its own page, MCP server development. Putting AI into an existing product is covered in AI integration services. If the data has to stay inside your own infrastructure, see private LLM deployment.

Building an AI chatbot begins with what the chat should know. The renovation platform is a system for managing renovations and fit-outs of properties: ordering products, a product search supported by AI, managing work and communication with workers. It is written in Python with Django and Django REST Framework, and we have worked on it since May 2022. Both the chat and the product-search agent run in production.

The chat answers the user. The agent next to it searches the texts and metadata of many products at once, across the offers of many suppliers. When a user asks about a price, the agent fetches the price information, and the chat's answer contains the product's source URL and the price from the platform. For questions about specific products the chat gives a link to the product.

The message passed to the AI carries the customer's context, so the conversation can move on to customer service.

The model handles the conversation, and Python code builds the search queries that have to come out the same every time. We moved those queries from the model to code for repeatability: the same question always produces the same query and the same answer. We made that decision after analysing the assistant's work on real conversations, in successive rounds. When a search finds nothing, the agent repeats it with changed parameters.

The chat and the agent run in a separate AI service. The main application calls it through a dedicated service in the backend, so the AI part can be switched off without touching the rest of the product. A flag set at project level in the application, ai_enabled, controls whether the AI is on. We describe this architecture in AI integration services.

The first version took shape in about four months:

• 15 December 2023: work on the product-search agent begins.
• 18 to 22 January 2024: the AI can be switched on and off for each project.
• By 2 February 2024: the agent searches across the offers of many suppliers.
• 2 to 5 February 2024: part of the search queries moves from the model to code.
• 27 February to 6 March 2024: the chat and the agent move to the separate AI service.
• February to 10 April 2024: further rounds of improvements to the chat.

For an agent that creates and saves things in your system, we put the rules on the server. An agent works through tools. The model chooses a tool by its name, description and typed input schema, and it never sees the code of the tool. On our server each tool calls an endpoint of the CMS API, so validation and permissions stay in one place.

Rules the server enforces

On our own MCP server the rules an agent must not break are response codes:

1. Creating a page without a brief returns HTTP 400 (request rejected). The brief is the description of a page, tied to a goal and a role in the sales funnel. A person approves it before an agent writes content.
2. Writing content before a person approves the brief returns HTTP 409 (conflict).
3. Publishing with audit errors returns HTTP 409. The audit is a check the server runs before publication. It flags facts outside the approved list, client names covered by an NDA and a missing verdict from an independent critic agent.

Every tool goes through the same gates. For live pages there is also a correction mode: it publishes a draft only when its audit finds nothing worse than the live version. The agent also learns the rules from the server. Two tools, list_page_types and list_blocks, return the available fields. The agent has to call them before it creates content, and that rule sits in the server's instructions, which the model receives automatically. When the server rejects data, the response points to the exact field, so the agent can correct the data itself.

Change control: revisions and rollback

A save creates a revision, and content goes live only after publication. Rolling back means restoring a revision. Deleting a page is a soft delete with an optional redirect, and deleting an image that a page uses returns HTTP 409 with the list of those pages.

Access management for enterprise use

Every request carries the token of one person or integration, which an administrator can switch off at once in the admin panel. Every request lands in a call log with the method, the path, the response code, the token used and the page it changed, so the log shows who asked an agent for what. Different roles have different token permissions.

Automation across 164 tools

The server has run in production since 9 July 2026, and on 20 October 2026 it had 164 tools. Our agents use them to automate work on pages and content, SEO, schema.org, media, menus, redirects, Search Console and GA4 analytics, and leads. Agents reach the server through the open Model Context Protocol. Anthropic created the standard and donated it to the Agentic AI Foundation under the Linux Foundation on 9 December 2025. According to Anthropic's announcement of that date, ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code support it, and Anthropic's assistant Claude works with it too.

We also run a public endpoint, https://ordergroup.co/mcp, open without a token since 8 October 2026. It has two tools: one lists our case studies and one submits a project enquiry. The endpoint is listed in the official MCP registry as co.ordergroup/public. How the server works in practice is in our article MCP vs API: what we learned running an MCP server in production.

Where a wrong output costs something, we keep a person in the loop.

Speech-to-text corrected by a specialist

Centrum Komunikacji is a web and mobile application run by the Polish Association of the Deaf, for people who are deaf, deafblind or blind. It offers audio and video calls, chat, history and call queueing, and it has run in production since 20 April 2026. Live speech-to-text is one of the features we added.

The association ruled out raw speech recognition, because automatic captions confuse names, addresses and the names of medicines. Instead the text is corrected during the conversation. Speech recognition writes a draft, and a transcription specialist corrects it in their own panel. The uncorrected text is never shown to the user. Only the corrected text appears, as a chat message. A session lasts up to 30 minutes. The recognition engine is the open-source Vosk with a Polish model, and it runs in the specialist's browser through WebAssembly, so the audio does not go to an outside API. In the interface the draft and the approved text have different colours. Personal data sits in encrypted fields, and deleting a user's data on request is logged.

AI summaries confirmed by an editor

Enerace.Online is an advisory platform for industrial buyers of energy and gas that we have built and developed since 2020 (Enerace case study). Market news items sit next to the price charts, and AI writes a summary of each item in English and in Polish at the moment the item is added, not on every read. The model also assigns the item a commodity type and a country or region, and rates the market direction (rise, fall or unchanged) with a short justification. The AI feature arrived without rewriting the platform, which has more than 20 functional areas in one codebase.

The rating is the model's assessment, and an editor confirms it. Each summary has a status: new, in progress, generated or confirmed, and only confirmed ones reach the reports and the news list in the app. The feature has been live since 2025.

We test a chat and an agent like the rest of the application, and add what AI needs on top. A round starts from real conversations: we read how the assistant behaved, decide what to change, and check the next iteration the same way.

PSONI, the Polish Association for Persons with Intellectual Disability, commissioned Generator ETR from us. It is a web and mobile app in which AI turns difficult text, such as an official letter, into easy-to-read language. Every model version is scored with a readability metric on a set of test samples. The metric, the Gunning Fog index, was designed for English, so on Polish text it works as a trend between versions. A reviewer trained in the easy-to-read standard then compares real outputs with their source texts, and the results feed the next iteration of the prompt and the model.

QA runs jailbreak tests that try to pull the model off its task, for example with a request for an apple pie recipe instead of a letter to simplify, and repeats them with every new iteration. QA reports set the source text, the result and its ratings side by side, so a reviewer can see what changed between two iterations. For legal and medical texts the app shows a notice that the simplified version is not a legal document and that a lawyer or a doctor can explain the details.

When you compare AI agent development companies, look at what you can check on the page: a chat, an agent and an editorial workflow that run in production, rules that sit in code on the server, and who owns the result.

A project runs in five steps: discovery, proof of concept, MVP, launch and support. It starts with a free basic discovery, one meeting. We look at your product and your data, including whether the data may leave your infrastructure, at the legal, security and company rules, and at the task together with its test. We use our own AI toolkit in discovery. A scope, an estimate and a defined success test come out of it, together with a basic proof of concept or a mockup that lets us estimate the cost. A wider discovery is a separate, larger scope, and sometimes the outcome is advice to skip the language model altogether.

After the discovery comes a kick-off, where we agree the methodology, set up YouTrack and the communication channels, and assign the roles. The first estimate covers the whole scope and begins with the simplest delivery route, with ready-made solutions where they fit. Then we build the AI service and the tools, test them in rounds, and launch.

The code is yours. Copyright passes to you with each payment, within the scope of use set in the contract. After launch we stay on for support and maintenance, because models, prompts and the data they work on keep changing.