Multi-agent systems, where several AI agents work on one process under an orchestrator, are turning into a separate IT budget line: Axenix estimates that Russian financial institutions will spend 10-20 billion RUB on AI agents in 2026. The study came out on 19 November. For a mid-sized business the takeaway is not the size of bank budgets but the order of steps: one agent on one process first, a team of agents later.
What Axenix calculated
The estimate starts from the generative AI market in banking and insurance: 12 billion RUB in 2025 and 14.4-19.2 billion RUB in 2026, assuming 20-60% growth. Of that, 25-35% goes on buying agent solutions from vendors (3.6-6.7 billion RUB), while in-house development costs roughly twice as much (7.2-13.4 billion RUB). Sber's investment in its own platform is not included.
The analysts describe three maturity levels for agent systems:
- A single agent handles one narrow task.
- An orchestrator runs several specialised agents across an end-to-end process.
- A dynamic system: agents assign tasks among themselves within business rules.
The largest financial institutions are now moving from level one to level two. Two or three systemically important banks are piloting it, and by the end of 2027 seven to nine organisations will be testing multi-agent setups. Axenix expects level three in 2028-2029, and not before 2030 for payment operations.
What a multi-agent system costs a bank
According to Axenix, the first pilot reaches production in 6-9 months and costs 50-100 million RUB: a team of 7-8 specialists, compute and LLM platform licences. Such a system handles 40-60% of operations on its own, and people do the rest. Experts in a ComNews review from 18 November add a constraint: the client company is legally and financially responsible for an agent's mistakes, so significant decisions need human sign-off.
A multi-agent system does not start with architecture. It starts with one process where an agent already works and the cost of a single operation is known. A team of agents makes sense on top of something that has already paid off.
What it means for a mid-sized business budget
A mid-sized company does not need a 50-100 million RUB bank pilot, and copying one makes no sense. The path is the same as the banks', just shorter: level one on the process with the most manual routine. Our first AI agent starts at 97,500 RUB, and a team of agents is quoted separately once the first one has shown results. Common starting points are chat and sales plan control (AI sales head agent) and a summary of income, costs and cash runway (AI finance agent).
Measure the money with three metrics from the same review rather than the share of operations "run by agents": cost per operation including the model, infrastructure and human checks, time to result, and error rate. If the cost per operation has not dropped after the first agent, a second one will only multiply spending. We covered the rollout order in our piece on rolling out AI agents in a business.
