A direct cash gap or selling at a loss of 126,720 rubles threatens a plant when commercial proposal calculations are delegated to a universal language model. In the industrial sector, response speed decides the outcome of a deal: an application for the supply of rolled metal and metal structures on the real market lives for a matter of hours. The customer sends a specification to several suppliers simultaneously and contracts with the one who is the first to issue an offer with precise parameters, while in production, preparing such a calculation takes from 4 to 8 hours of working time of a qualified estimator.
An attempt to cut these hours by connecting a ready-made LLM to request processing breaks the unit economics of deals. The blind application of probabilistic text generators for production calculation leads to a systematic underestimation of cost. The incoming stream of specifications is far from structured tables: in reality, about 90% of incoming applications represent informational chaos from scans, photos of drawings, motley tables, and messenger messages. Enterprise management expects that multimodal models will instantly turn this array into a ready-made calculation. However, financial and physical calculation cannot be trusted to text neural networks.
A common objection from developers and optimizers sounds simple: configure a detailed prompt, upload examples, and the model will issue exact figures without involving specialized engineers. A controlled experiment on the GPT-4o model shows the opposite. When submitting a real request to calculate the theoretical mass of a batch (16P channel, 300 meters long), the system generated a plausible response with formulas, indicating a mass of 2.50 tons.
Correct batch mass: 300 m × 14.20 kg/m = 4,260 kg = 4.26 tons.
In reality, the theoretical mass of 1 linear meter of the 16P channel is 14.20 kg. The model's error was 1.76 tons on the downside, which means underestimating the batch weight by 41.3%. At a base purchase price of 72,000 rubles per ton, the cost of rolled metal according to ChatGPT's estimate amounted to 180,000 rubles instead of the real 306,720 rubles. A plant that sent such an offer to a client either sells the batch at a loss or disrupts the supply.
Why does the language model make such a gross miscalculation?
It operates as a probabilistic token predictor rather than a deterministic calculator. The model does not calculate physical quantities, but extracts statistically plausible numbers of neighboring standard sizes from the training sample. Instead of an exact standard, it substitutes a plausible weight, creating the illusion of error-free calculation.
Layer Isolation Instead of Blind Generation
Eliminating hallucinations in commercial calculations requires abandoning end-to-end generation. A working solution involves dividing request processing into three strictly isolated loops, where neural networks are denied access to arithmetic. At the first stage, a multimodal model (Vision API) analyzes document layout and extracts unstructured data, after which normalization and multiplication are transferred to deterministic code and rigid standards databases.
The multimodal model (Vision API) analyzes the document layout.
The promise to close plant estimate preparation with a single universal chat prompt turns into direct holes in working capital. Instead of instant automation, risk-free processing requires linking specialized recognition with rigid regulatory directories, where mathematics remains with classical algorithms.
The exact cost of implementing a hybrid system compared to manual work by estimators in the source materials is not disclosed, which leaves open the question of capital expenditures for developing such a loop.
What’s behind it
A general-purpose text generator simply substitutes plausible figures for physical standards. Keeping the business from selling metal at a loss requires isolating the AI inside a three-stage workflow that strips it of direct arithmetic. The drive to cut quotation times yielded a flawlessly phrased commercial proposal that simply forgot to bill the customer for nearly half the steel.