Step 1. A data policy
Split data into three classes: may go outside, only to a Russian cloud, only inside the company perimeter. An agricultural holding from this section’s case studies runs webinars for staff and will approve an internal AI policy by the end of 2026. Result of the step: one page of rules signed by the head of the company.
Step 2. One task with a measurable result
Take a task with a steady flow: sorting requests, checking contracts, answers from a knowledge base, document drafts. Write down the metric before the start: minutes per document, the share of requests handled without an operator, the number of errors. A telecom operator put a response time of up to 400 ms and refactoring accuracy of at least 85% into its tender: this is what a measurable task looks like.
Step 3. 50 test requests
Collect 50 real requests with expected answers. This is your exam for any model, cloud or local, and it matters more than any ranking. A retail chain ran 3 scenarios on 3 models 3 times each and got its answer in 27 runs.
Step 4. A trial on a rented card
Rent an RTX 4090 from about $0.50 an hour or an H100 from $3 an hour in a Russian cloud with hourly billing, install Ollama or vLLM and run the 50 requests on two or three open models: 8–9 billion, 30 billion and 70 billion parameters. Measure quality and speed. A week of trials costs from $87 to $500 and replaces a blind purchase of $12,000.
Step 5. The model and the engine
Choose the smallest model that passed the exam: it needs less memory and runs faster. Choose the engine by the number of users: LM Studio or Ollama for one person, vLLM for a department. Pin the versions of the model and the engine: an update is a separate event with a repeat exam.
Step 6. Hardware by calculator
Enter the model, the context and the number of users into the memory calculator and get a list of suitable hardware. Then use the monthly cost calculator and the break-even point against the cloud. Benchmarks as of 29 September 2026, by our calculation: a Mac Studio with 128 GB at $11,000 for 1–3 users, a workstation with an RTX 5090 at about $13,000, and a server with two 96 GB cards at about $30,000 for a department. If there is no payback and the data may go outside, stay with the cloud or rented hardware.
Step 7. Agents through one interface
Agents connect to the model through one interface, so you can change the model with a setting. Move heavy deterministic processing into code, as a bank’s IT manager did: he moved deterministic processing into separate applications and left the rest to the agent. Keep chains short: in the retail chain’s test, all three models solved a two-step task in every run, but on 15 steps none of them, local or cloud, gave a fully verifiable result.
Step 8. Monitoring and maintenance
Monitor temperature, memory errors and response time from day one. Back up the settings and the document base. Have a failure plan: a spare card, a rental contract or a switch to the cloud for data that is not restricted. A responsible person at 10–20% of a full-time role for a department server, by our estimate. A repeat exam on the 50 requests after every update.
Timeline and our role
By our estimate, steps 1–5 take 2–4 weeks, step 6 depends on hardware delivery, and steps 7–8 take another 2–4 weeks. We do steps 2–5 and 7 in a pilot of 14 days from $1,499 on your own or rented hardware, and an AI department of 3–5 agents on a local model in 4–6 weeks from $4,999; you buy the hardware based on our calculation.