Costbl: cost estimation from drawings

About the client

Costbl is a FAVORIT-line product for metalworking and manufacturing. Goal: replace manual Excel costing and turn a drawing into clear cost of goods plus a document pack (proposal, request, invoice).

The project ran for more than five months. This was not “ship a chat in one sprint”: we tried approaches and research tracks, brought in analysts and engineers, and worked by OKRs. In the end we locked a working approach and built a multi-agent system.

The challenge we started with

A drawing is not a ready calculation: parts, operations, dimensions and material must be extracted
When data is missing, the system must ask rather than "guess silently"
Cost must be reproducible and clear to an engineer/analyst
Long R&D: different approaches, OKRs, analysts and engineers working together
From a chat MVP to a full workflow graph without losing the meaning of the calculation

A long road: 5+ months and two stages

Not a linear "feature after feature". We tried different approaches and research, involved analysts and engineers, and checked progress against OKRs. In the end we fixed the architecture and built a multi-agent system.

Stage 1 - AI chat on a drawing (prototype)

The first working pipeline: the user uploads a PDF and writes "make a calculation". The system parses the file, finds parts and operations, and states honestly what is missing (material, dimensions, roughness). The screenshots below show stage 1 / the prototype, not the final product.

Stage 2 - full calculation (as it is now)

As in production now: a multi-agent workflow and live cost calculation from a drawing. The video below shows stage 2 specifically: a real drawing and cost calculation, not a prototype.

Multi-agent system

The result of the research is a multi-agent pipeline: drawing reading, geometry, calculation and documents are linked into a reproducible workflow. The approach was not found on the first attempt - it came through iterations with engineers and analysts.

Video

Stage 2: real drawing and cost calculation

This is already stage 2 (not a prototype): a live run of drawing → calculation → cost. Press Play, no autoload.

Gallery

FAVORIT: SaaS cost estimation from drawings - screenshot 1

FAVORIT: SaaS cost estimation from drawings - screenshot 2

How it works

User scenario

Uploads a drawing / describes the product in chat
Clarifies material, dimensions and missing parameters
Views the calculation path on the graph (stage 2)
Gets the final cost and documents (quote / request / invoice)

Team / admin scenario

Manages OKRs and the quality of calculation scenarios
Brings analysts and engineers together around the pipeline
Develops graph nodes and subscription limits

Technology

Duration: Over 5 months · 2 stages
Stage 1: AI chat · PDF drawing parsing
Stage 2: Chat + graph · multi-agent
Product: costbl.ru

Timeline: More than 5 months: stage 1 (chat) → stage 2 (graph + multi-agents)

Result and business effect

Launched Costbl (costbl.ru) - SaaS cost estimation for metal parts from drawings. 5+ months, two stages: chat with PDF parsing → full multi-agent calculation graph. The page includes screenshots of both stages and a video of a real calculation.

A fit if you need

You need cost estimation from drawings, not Excel - AI agents
You want a transparent calculation path (graph / workflow) - Web applications
You are ready for R&D with engineers and analysts under OKRs - AI agents

Frequently asked questions

Questions about this type of project, the same ones clients ask before a similar build.

What happens if the agent gets it wrong?

A mistake stays a draft. Anything touching money, stock or new counterparties needs a person to confirm by default. When the agent is unsure, it does not act; it creates a task for a person. Every action is in the log and can be reversed.

Who approves what the agent does?

You decide. At the start we agree where the line is: for example, the agent creates CRM tasks on its own but only drafts invoices and supplier orders for a manager to approve. Once the team trusts the output, the line can move.

Do we have to change 1C or our CRM?

No. We connect to what you already run: via API, 1C data exchange or file export. If a system has no way to accept data from outside, we will tell you during the estimate.

How is this better than a template-based bot?

A template bot breaks when a supplier changes the email format. An agent understands free text and tables in different layouts, so you do not have to force partners into a template.

Video