AI in business processes
What it looks like in practice
Figures from our projects where AI already works every day:
In all three cases the AI does the routine work: reads drawings, scans and request texts, moves data and assembles documents. A specialist checks the result and makes the decision.
How we bring AI in
- We start with one task. We look for an area where people copy data from sheet to sheet or check things by eye: parsing requests, preparing files, reconciling documents.
- Measure first, promise later. We build a reference set from your real documents and run several models on it. For the property management company this was more than 150 requests labelled by hand.
- We don't take the agent's report on trust. The result is checked by a separate script or a second agent: in the design bureau the model is checked against the drawing by mass, cross-sections and shape.
- Mistakes become rules. Every failure we analyse becomes a rule, a test or a verified function. The design bureau already has more than 40 such rules.
- Red lines are written down in advance. In the laboratory the AI parses documents and checks formatting, but never touches measurement results.
Which tasks we hand to AI
- Incoming documents. Scanned requests, free-form emails and receipt photos become a draft with flags where a person is needed.
- Request classification. The type of work, urgency and responsible team are determined from the request text and the company's own catalogue.
- Engineering documentation. 3D models from ESKD drawings, drawings, bills of materials, files for laser cutting, fit and electrical insulation checks.
- Template documents. Work reports from a technician's voice note, commercial quotes, formatting to the company standard.
- Answers from the knowledge base. The assistant answers from your regulations and instructions and links to the exact clause of the document.
Data and 152-FZ
Personal data does not reach the model. In the laboratory the system blacks out names, phone numbers, emails, bank details and signatures before sending and checks the page again, with a regression test for every redaction error found. In the property management company the text of requests is anonymised.
We choose the model to fit the requirements: Russian (GigaChat, YandexGPT), foreign, or local on the company's own server, so drawings and documents never leave the network.
We train the team
A system works only if people use it. Along with the rollout we teach staff to work with AI agents: everyone gets their own course on a private platform, lessons built around their tasks and hands-on practice on a real project. More in the “Team training” section.
Projects with AI
FAQ
Where should a company start with AI?
With one task that staff now do by hand and from memory: parsing incoming documents, moving data between systems, a routine check. On the call we pick such an area, build a prototype in 5–9 days with no prepayment and measure how much time it saves.
What happens if the AI makes a mistake?
Decisions that affect money, deadlines or safety are made by an employee. The agent prepares a draft and flags what it is unsure about. Before launch, accuracy is checked against a reference set of your real documents, and every error found becomes a rule or a test.
Which models do you use?
We choose the model to fit the task and the data requirements. For projects with personal data or government contracts we use GigaChat or YandexGPT and keep the infrastructure in Russia. Some tasks can go to a local model on the company's own server, so documents never leave the network.
How is personal data protected?
Personal data does not reach the model: before sending, the system finds and blacks out names, phone numbers, emails, bank details and signatures, then checks the page again. If anything is left, the page is not sent to the model.
Do employees need training?
Yes, otherwise the system will not take root. Along with the rollout we teach the team to work with AI agents: everyone gets their own course on a private platform and hands-on practice on a real project, and the mentor sees who has read what and where they got stuck.
How much does bringing in AI cost?
It depends on the task and the volume of data. We estimate after a call and show the first result on a prototype before any contract. Payment is made in stages, against signed acceptance acts.
Tell us which step in your process is done by hand and from memory. We will show how to hand it to an agent without losing control.