M2 handed 40% of its marketing copy to AI without losing content quality

M2 has shown a practical setup in which AI doesn't replace the editorial team but takes over a significant share of routine content. The company already trusts the model with about 40% of its product and marketing copy, but the final system is more than a single button press: quality depends on strict rules, precise prompts, and mandatory human involvement in complex and creative tasks.

The AI editorial project started because content volume kept growing. The editorial team had to handle more work for marketing, sales, and support without bloating headcount or outsourcing everything. M2 first weighed the standard options: hire another copywriter, outsource part of the work, or build an internal AI tool. It chose the third as the fastest and cheapest, but set a constraint from the start: the model had to write in the brand's voice and follow the internal editorial policy at least as well as a junior writer.

The main problem turned out to be not choosing a model but explaining the company's standards to it. M2 didn't train a separate model on its own text corpus; it built a system of prompts instead. In the first phase, the team packed the core Tone of Voice rules, syntax constraints, stop words, and product naming requirements into one large system prompt. That zero-shot approach didn't work: the model lost focus, wrote blandly, and sometimes mixed styles. The team then switched to a few-shot setup, adding good and bad examples of real texts to the prompt alongside the instructions. That produced predictable results and let the team build separate templates for news, social media, and other formats.

Under the hood, the AI editorial system runs on Videocat, a corporate platform with llama.cpp and Open WebUI components, and the case described uses Google Gemma 3 as the working model. But the specific stack matters less here than the architectural principle: the tool stays flexible so other models can be plugged in and tested for different tasks. This removes dependence on a single provider and lets the team balance generation quality, speed, and cost.

As a result, M2's model doesn't just write drafts and adapt texts for different channels; it also helps generate ideas, extracts information from images, and explains its own reasoning. That last part matters especially for employees outside the editorial team: when the model justifies its word choices and structure, it acts as both an assistant and a teaching tool. At the same time, the company deliberately does not try to turn AI into a full replacement for an editor.

M2 draws clear lines of responsibility. Routine, templated, and high-volume tasks can go to the machine almost entirely, but in-depth pieces, creative concepts, slogans, and material that demands emotional precision still stay with people. This balance helps beat blank-page anxiety, speeds up drafting, and takes load off the editorial team without sacrificing the human feel of the text. Security is a separate piece: the AI tool runs inside the internal network perimeter, and employees are trained on which data they can and cannot upload to the system. That detail matters for any company that wants to use generative models in real workflows rather than in a sandbox.

The takeaway from M2's case is fairly down-to-earth, which is what makes it useful: models can already take over a substantial chunk of routine content work, but they can't build a strong editorial team on their own. A company first needs its own standards, a clear Tone of Voice, and real editors who can tell good writing from merely correct writing. Only then does AI become a force multiplier for the team rather than a threat to jobs. For the market, it's another signal: demand will shift not from writers to machines, but from ordinary content producers to specialists who can design workflows, tune prompts, and bring a machine draft up to brand standard.