Faster business workflows with KobiGPT’s model upgrade

At X Mind Solutions, we have completed the migration from Gemini Flash 3.5 to Gemini Flash 3.7 in KobiGPT with a few of our pilot customers. KobiGPT now uses Gemini Flash 3.7. Our logs from real-world usage show that the model change accelerates not only text generation but also workflows connected to company systems. For us, the value of this development lies less in a technical performance improvement than in reducing the time users wait for answers to their questions.

In calls without tool use, generation speed increased from approximately 42 tokens per second with Gemini Flash 3.5 to approximately 117 tokens per second with Gemini Flash 3.7. This means text generation is approximately 2.8 times faster. However, this measurement alone does not capture the full KobiGPT experience. Answering a business question often requires accessing relevant records, filtering data and analysing it, in addition to generating text.

In real-world workflows using tools that involve ERP, CRM, RAG and other company systems, average response time fell from 39.8 seconds to 12.7 seconds. We thus observed an approximately 68% reduction in total waiting time. These figures were obtained from actual KobiGPT usage traffic, not laboratory tests. We interpret the results within that scope: the measured averages are not a guarantee that every query will be completed in the same amount of time, but concrete evidence of the change in the workflows we observed.

For example, when a user asks which products sold the least last month, KobiGPT does more than prepare a text response. It retrieves sales records through the ERP connection, filters the relevant date range and analyses products alongside their sales volumes. When needed, it cross-references stock data to rank products with low sales performance. It then presents the result clearly. The speed improvements we measured in similar tool-based workflows translate into shorter overall waiting times for this multi-step process.

Access to KobiGPT is not limited to web and mobile; users can also ask questions about company data via WhatsApp. Examples include checking whether there is a production disruption, requesting a summary of the past week’s production report or listing customers with overdue payments. KobiGPT connects to company systems in line with the user’s permissions, finds and analyses the relevant data, and sends the answer to WhatsApp. This removes the need to switch to another application to ask a question.

With the model upgrade, we are achieving lower model costs, faster agent-based workflows and faster ERP and CRM interactions. From the user’s perspective, our focus is on delivering answers based on actual company data with less waiting. We see web, mobile and WhatsApp access as part of this experience. Businesses interested in exploring KobiGPT’s existing ERP and CRM connections can view the supported systems in the Integrations section of our website.