One RTX 3090 Can Run an AI Helpdesk for a 40-Person Office and Pay for Itself in 5 Months

Cloud models are smarter than local ones, and nobody will argue with that. But "smarter" isn't the only metric. Subscriptions keep getting more expensive, limits on the top model run out in half an hour, and your data goes to the provider. For accountants, lawyers and anything involving personal data, this is a practical constraint, not an abstract risk. At an open lecture on November 6, Sasha Andreev, CTO of the studio "Flag" (Flag), showed what a sober calculation looks like: when a graphics card in a desktop PC is enough, and when it isn't.

An Agent with Tools Instead of a "Model That Knows Everything"

The first case is an ordinary non-IT office of 40 people: 1C, accounting, sales, management. A manager pesters the accountant, the accountant drowns in bank statements, the director wants a list of debtors "right now." Each task is simple, but together they create constant noise.

Andreev suggests not teaching the model everything, but giving it tools: access to 1C, Bitrix24, electronic document management (EDO) and internal regulations. The model doesn't reinvent invoice payment from scratch; it calls an ordinary automation script. Such scripts are tested, scale well, and don't "reason" for the thousandth time with a chance of making a mistake. The scheme is simple: a request lands in a queue (Bitrix, Planfix, Trello — it doesn't matter), the agent picks it up and executes it, and the result is returned to the person.

The same logic works in his development workflow. A bug or client request lands in the queue, a local Qwen 27B in agent mode goes into git, commits, merges and deploys. A stronger cloud model (most often Codex) checks the result as a reviewer. The goal is finished work, not a victory of the local network over the cloud.

The boundary is also clear. A local model can call an already written 1C handler. It cannot develop a 1C configuration: the context of agent sessions there is huge, and the cloud copes noticeably better.

The Math: The Card, Tokens and Quantization

On an RTX 3090, Andreev gets about 150 tokens per second. In a 40-person office, 2–3 requests typically run at the same time; the rest are queued or at lunch. An average dialogue takes about 20 thousand tokens, and with a trip into 1C about 60 thousand. The context of his configuration is up to 262 thousand tokens. For a helpdesk, that is enough.

The economics look like this. A workload of about 2,000 calls per week in the cloud costs roughly 296 thousand rubles a year in Yandex AI Studio and 267 thousand in Cloud.ru (GigaChat). Andreev bought his 3090 for 75 thousand. Right now such cards go for about 120 thousand on Avito, and at that price payback comes in around five months. An important caveat: this is only the card, without the rest of the hardware and electricity, not the total cost of ownership.

To fit a 27-billion-parameter model into 24 GB of VRAM, it is quantized. At 16 bits it weighs about 54 GB; in Q4 format only about 13.5 GB. On benchmarks the quality loss is around one and a half points. An analogy with the number π: for everyday calculations 3.14 is enough, but you can't round it to three, or quality falls apart. The KV cache, that is, the "summary" of the context, is quantized too. Speed depends heavily on the engine (llama.cpp, Ollama and specialized kernels), and benchmarks are usually measured on an empty context. In real use with a full window, speed is lower.

What to Do in Practice

The order Andreev recommends: first choose the model, quantization and tasks, then run the same model in the cloud at a provider, and only then buy hardware. The yardstick is neither the token price nor the card price, but the cost of successfully completed work. If the workload is repeatable and the model copes, your own hardware pays off. If not, you need a hybrid or the cloud. Cheap tokens that give a bad result end up being the most expensive.

A few more takeaways from the lecture:

• Start with analytics: pain points, 3–4 priority cases, a pilot, and measuring the effect.
• A hybrid is the most common option. The local model executes, a cloud model or a rigid script verifies. Human oversight doesn't scale to 40 employees; verification must be automated.
• Without a digital foundation (CRM, task tracker, wiki, messenger), agents have nothing to hold on to.
• Model weights can't be "read," so treat them as third-party software and keep a control perimeter around them.

The limitations should also be stated honestly. All figures relate to one setup and one workload, and this is the speaker's personal experience, not an independent benchmark. Prices for used cards are rising: in summer 2025 a 3090 cost about 55 thousand. Andreev himself says he regrets buying only one card, and the price of this path is constantly having to figure out quants, models and engines. Still, for repeatable office routine, a consumer graphics card is no longer exotic but a working tool.