AI implementations
A chatbot that responds from your knowledge (RAG), and does not make things up. Classification of emails and leads. Reading invoices and documents. We build on your data, host it wherever you want, and show a working prototype before you pay the second installment.
If you recognize at least two, it makes sense to talk about implementing AI.
01
The salesperson doesn't know if this product fits this machine, so he calls the technologsst. Onboarding a new person yeses months because the knowledge is not in one place. The RAG assistant responds with your documents in seconds, with a link to the source, not with fabrications.
02
The contact@ mailbox is a bag: inquiry, complaint, invoice, spam. Someone reads it and sends it around. AI classification identifies and directs the report to the right person or to a machine immediately, without waiting for the morning inbox check.
03
Invoices, orders, specifications, contracts. Someone opens the PDF and copies the data into the system. AI reads the document, extracts fields (Tax Identification Number, amounts, items), and gives it to a human for validation only when it is not certain. Less rewriting, fewer typos.
04
A ready-made chatbot does not know your price list, procedures or customer history, so it gives untruths in a confident tone. RAG connects the model to your knowledge base: it responds only from what you have, and when it doesn't know, it says it doesn't know.
A working prototype on your data before we write a line of production code. Payment divided into four installments.
Week 1
What should AI do, on what data, where is this data (PDF, database, Customer Relationship Management, emails). We define one specific use case for the MVP and leave the rest aside. We check GDPR requirements and whether data can go beyond your server. Deposit: 30%.
Weeks 2–3
Selection of the model (cloud or own server), RAG architecture, method of measuring the quality of the response. Clickable prototype on a Sample of your real documents. You see him respond before you decide on the full build.
End of Week 3
You judge the quality of answers on real questions. Without approval, we do not start building the MVP. Upon approval: another 30%.
Weeks 4–7
The complete process: data preparation, indexing, hallucination control, panel or integration with your system. Tests on your data with accuracy measurement. A working version is shown every week. After the MVP release: another 30%.
Week 8
Migration to production, quality and cost monitoring (how many queries, how much they cost, where the model is wrong). Documentation in Polish and handing over to the team. After implementation: last 10%.
The new salesperson asks the assistant instead of interrupting the technologsst. The customer gets a response in the chat instead of waiting for an e-mail. The company's knowledge is no longer trapped in the heads of three people who are on vacation.
RAG doesn't make things up. Each answer identifies the document it comes from, so you can verify it. When a model has no basis in your data, it says it doesn't know, instead of probably providing an untruth.
Invoices, orders and e-mails are read carmatically, data is extracted into the system, a human intervenes only in uncertain cases. With the Same staff, you will process a larger volume.
You see how many inquiries and how much they cost. Model on your server when data cannot go outside the company. Without dependence on one supplier, when prices or policies change, we switch to another model.
Cursor and Claude in development, AI in content pipeline, lead classification and quality control. We don't sell something that we don't use every day.
JSON Crew · own processes
AI-first
We write code in Cursor and Claude, generate and control the quality of content, classify and evaluate leads, and conduct research. We know where AI really shortens the work and where it only looks impressive on the slide. We transfer this knowledge to your case.
KG Electronics · Boiler Manufacturer
12 000
Telemetry backend: 12,000 devices send data, the system aggregates and alerts. It's not RAG, but it's proof that we can harness data at scale and keep it in production for 18 months. AI sits on such a data layer.
View the full case studyAutomation · Related Services
n8n + AI
Often, AI is one element of a larger process: email classification, and then the machine sends it further. If the main goal is to carmate repetitive tasks, start with carmation, we will add AI where it makes sense.
See carmationsTwo paths, two different company profiles.
Ready chatbot / SaaS AI
Dedicated RAG on your own server
A ready-made chatbot is enough for a simple Frequently Asked Questions. Dedicated implementation is done when the answers must be based on your knowledge, the data is sensitive, or AI is to be part of the process and not a widget in the corner of the page.
What happens after implementation, and who we are.
After implementation
Optional packages. We will discuss the details during a meeting after implementation.
See also
30 seconds on JSON Crew
Software development company with a niche in the digital transformation of B2B sales. Product configurators, panels, carmation and AI for the sales process and operations. Three main case studies (Akpil, Forest, internal JSON Hub platform) and portfolio implementations (KG Electronics, Metal Roofing, heat recovery center, modular houses).
RAG (Retrieval-Augmented Generation) connects the language model to your knowledge base. Instead of responding with general knowledge from the Internet, the model first searches your documents and then responds only based on them, with a link to the source. Thanks to this, he knows your price list and procedures, and when there is no basis in data, he says he doesn't know, instead of making things up.
Hallucinations cannot be eliminated, but they can be limited and detected. We use RAG (source-based responses), force document citations, measure accuracy on your real questions and set a threshold below which the model says it doesn't know instead of guessing. You assess the quality on the prototype before you pay for the full construction.
Depends on requirements. When the data is sensitive (GDPR, personal data, finances), we place the model on your server and the data does not go outside. When there are no such restrictions, we use cloud models, which are cheaper and faster. We make the decision in the first stage, the data entrustment agreement is signed before the start.
We match the task, budget and GDPR requirements, not the other way around. Cloud models (e.g. Claude, GPT) when quality counts and there are no data limits. Open models, placed on your server when the data must stay with you or the volume is large. The architecture is built in such a way that it is possible to change the model when prices or the supplier's policy change.
Individual valuation after determining the scope in the first stage. Fixed price after defining the use case, schedule 30 / 30 / 30 / 10. We price a simple Frequently Asked Questions assistant on ready-made documents separately, cheaper than full implementation with integration into the system. In addition, there is the cost of querying the model, which we show in advance and monitor after implementation.
This is a normal situation, not an obstacle. The first stage is a data review: what you have, what condition it is in, what needs to be sorted before indexing. Sometimes a Sample of key documents is enough for MVP, and we add the rest later. If there is not enough data for AI to make sense, we will say so directly instead of yesing the project by force.
Describe what you want the AI to do and what data it should operate on. We will respond within the next business day.
