Claude Sonnet helps executives build an AI officer for tough decisions in 8 hours

In a single working day, a team made up of a CEO, CTO, CIO and an AI engineer built not a chatbot but a management AI assistant that can argue with leadership and does not give up its position under pressure. At the closed Snow BASE intensive, the CAITO project (Chief AI & Technology Officer) had to make decisions in a crisis case for a retail division and change its mind only when the facts change. The hackathon in Sochi was organized by the AI Talent Hub of ITMO University and South HUB.

Participants had eight hours to solve the case: at a large retail business, the recommendation system is degrading because of seasonal data drift, the infrastructure is already running at its limit, cloud costs are growing faster than revenue, and new requirements under Federal Law 152-FZ raise the risk of fines. The board of directors faces an unpleasant choice: scale the system, freeze development or postpone the launch, with only 14 days to decide.

The CAITO team built a management model around this task in which the AI must not simply answer questions but hold a line between the interests of the CEO, CFO, COO and the technical side. This is where ordinary LLMs often break down. If a model simply continues the dialogue, it starts adapting to the last and most insistent interlocutor: the CEO pushes for growth at any cost, the CFO demands immediate ROI and cuts investment, the COO points to SLAs and operational constraints. The result is not a management position but a mirror of the current pressure.

So the team chose a deliberately pragmatic approach: first build a stable single-shot reasoning setup, with one model call per turn, and only then check whether a more complex agentic loop is needed. This mode gave a predictable response latency of a few seconds, one structured JSON output and simpler debugging under hackathon conditions. The team used Claude Sonnet as the model and built the service on Bun and TypeScript, with access to Cloud.ru Foundation Models through an OpenAI-compatible API.

CAITO's architecture rested on three pillars. The first is a system prompt with a strict mandate: the assistant must first state the decision, then the arguments, record any metric conflict separately, and not change its position without new data. The second is workflow.yaml, which defined the internal roles, their weights and the order of consultation: first facts from ML and economics, then operational constraints, and only then management policy. The third is long-term memory. It stored the immutable facts of the case separately from a live history of decisions made, assumptions, KPIs and shifts in position. This let the system keep context and explain why its opinion stayed the same or changed.

At the final presentation, the team showed three scenarios. In the first, CAITO was confronted with contradictory data and checked for which sources it relied on; to reduce the risk of hallucinations, answers began to show where the key figures came from. In the second, the CEO pressed the assistant to act immediately, but the system kept its frame and replied that without updated facts it could only clarify the risks, not rewrite the decision. In the third, the pressure came in waves: first new data, then an emotional attack, then another batch of inputs. Here CAITO had to tell a real change in the situation from repeated pushing and revise its position only on facts.

In parallel, the team also managed to build an alternative: a multi-agent pipeline of ten specialized roles with separate task routing. On some metrics this setup gave better analysis, especially where new signals must be finely separated from old pressure. But within eight hours, the main advantage turned out to be not architectural richness but reliability.

In the final presentation, the winner was chosen from a leaderboard in which automated scoring accounted for 70 percent and the jury for 30 percent; the criteria were the quality of management decisions and stress resistance, as well as functionality, security, stability, UX and cost. The main single-shot solution took the team first place.

The takeaway from this case is practical. For an AI that must take part in management decisions, the number of agents matters less than a clear mandate, transparent escalation rules and memory of previously made assumptions. An impressive multi-agent orchestra can add depth, but under a deadline it often loses to a simple, explainable and stable design.

The next step for CAITO is an agentic workflow with function calling, asynchronous role calls, dynamic RAG and full tracing. But the project already shows something more important: AI can take on a large part of preparing a management position, while final responsibility and the checking of non-obvious factors still stay with a human.