Claude Sonnet Helps C-Suite Build AI Director for Critical Decisions in 8 Hours

In one working day, a team of CEO, CTO, CIO, and AI engineer assembled not a chatbot, but a management AI assistant capable of arguing with leadership and holding its ground under pressure. At the closed intensive Snow BASE, the CAITO project, Chief AI & Technology Officer, was supposed to make decisions in a crisis case for the retail division and change opinion only when facts change. The hackathon in Sochi was organized by AI Talent Hub of ITMO University and South HUB.

Participants were given eight hours to solve the case: a large retail business faces simultaneous collapse of its recommendation system due to seasonal data drift, infrastructure is already operating at its limit, cloud spending is growing faster than revenue, and new requirements under Federal Law 152-FZ increase the risk of fines. On the table before the board of directors lies an unpleasant choice: scale the system, freeze development, or postpone launch, with only 14 days to decide. The CAITO team built a management model around this task in which AI should not simply answer questions, but maintain balance between the interests of CEO, CFO, COO, and the technical block.

This is precisely where ordinary LLMs often break down. If the model simply continues the dialogue, it begins to adapt to the last and most persistent interlocutor: the CEO pushes for growth at any cost, the CFO demands immediate ROI and cuts investments, the COO reminds about SLA and operational constraints. As a result, instead of a management position, you get a mirror of current pressure.

So the team chose a maximally pragmatic approach: first build robust single-shot reasoning, where each move requires one model call, and only then check if a more complex agentic loop is needed. This mode provided predictable response latency at the level of several seconds, one structured JSON output, and simpler debugging under hackathon conditions. Claude Sonnet was chosen as the model, and the service itself was built on Bun and TypeScript with access to Cloud.

ru Foundation Models through an OpenAI-compatible API. The CAITO architecture rested on three pillars. The first was a system prompt with a strict mandate: the assistant must first formulate a solution, then arguments, separately record metric conflicts, and not change position without new data.

The second was workflow.yaml, where internal roles, their weights, and order of consultations were specified: first facts from ML and economics, then operational constraints, and only then management policy. The third was long-term memory.

In it were separately stored immutable facts of the case and a living history of decisions made, assumptions, KPIs, and position shifts. This allowed the system to remember context and explain why opinion remained the same or changed. During the presentation, the team showed three scenarios.

In the first, CAITO was confronted with contradictory data and checked which sources it relied on; to reduce hallucination risk, the response began to show where key figures came from. In the second scenario, the assistant was pressured by the CEO demanding immediate action, but the system maintained its frame and responded that without updated facts, only risks could be clarified, not the solution rewritten. In the third scenario, pressure came in waves: first new data, then emotional attack, then another batch of information.

Here CAITO had to distinguish real situation change from repeated pressure and reconsider position only based on facts. In parallel, the team managed to assemble an alternative — a multi-agent pipeline of ten specialized roles with separate task routing. On individual metrics, such a scheme showed better analytical breakdown, especially where it was necessary to carefully distinguish new signals from old pressure.

But within eight hours, the main advantage turned out to be not the richness of architecture, but its reliability. At the final presentation, the winner was chosen by leaderboard, where 70 percent of the score came from automation and 30 percent from the jury; quality of management decisions and stress resistance were assessed, as well as functionality, security, stability, UX, and cost. The main single-shot solution brought the team first place.

From this case comes a quite practical conclusion. For AI that should participate in management decisions, what matters more is not the number of agents, but clear mandate, transparent escalation rules, and memory of previously made assumptions. An impressive multi-agent orchestra can provide depth, but under deadline often loses to a simple, explainable, and robust scheme.

The next step for CAITO is agentic workflow with function calling, asynchronous role calls, dynamic RAG, and full tracing. But already now the project demonstrates something more important: AI can take on a significant part of preparing the management position, while final responsibility and verification of non-obvious factors still remain with humans.