Slide contents

1. 1. Executive Decision Agent

   Hands-on in Yandex AI Studio: from knowledge base and search to calculations, options and an executive brief

2. 2. Alexander Polomodov

   Technical Director & Fellow, T-Technologies

   Architecture, engineering R&D, AI adoption.

   Closes HSE series with agent loop.

   Focus: agent for executive decisions.

3. 3. Bring course topics into one loop

   One managerial scenario joins the course

   Platform — AI Studio: models, retrieval, tools, API.

   Data — Connect company docs and KPIs as facts.

   Economics and control — Calculate impact, risk, payback and explicitly show assumptions.

4. 4. 01. What we build

   Executive Decision Agent: an agent that prepares a board-level management brief

5. 5. Not a chatbot: decision support

   Facts, options, checks before launch

   Context — Uses strategy, SLA, KPIs and security policy.

   Calculation — Evaluates economics, risk, payback and assumption sensitivity.

   Recommendation — Offers launch, pilot or postpone options.

6. 6. Question for the agent

   Realistic request, safe data

   Launch AI support next quarter?

   Brief: options, economics, risks, checks.

   Internal docs are primary facts.

   Missing data → assumptions/questions.

7. 7. Final answer: short and executive-ready

   Decision, trade-offs, next steps

   Brief structure

   Decision in one paragraph.

   3 options: launch/pilot/postpone.

   Economics, risks, dependencies.

   Management standard

   Separate facts, estimates, recommendations.

   Show sources and unknowns.

   End with checks before decision.

8. 8. 02. Yandex AI Studio

   Which platform capabilities we use in the lab and why the agent needs them

9. 9. AI Studio provides the agent prototype loop

   In the lab we use a minimal subset of the platform capabilities needed for the decision

10. 10. Knowledge base grounds the agent

    Lab pack: github.com/polomodov/2026-05-23-hse-agent-lab

    What to upload from the repository

    company_context.md and support_kpi.csv.

    security_policy.md and investment_rules.md.

    What to verify

    Facts from documents, not generic knowledge.

    Named sources; missing data visible.

11. 11. Web search is only market context

    Internal data and rules stay primary

    Good use

    Check market context and adoption risks.

    Separate external observations from internal facts.

    Bad use

    Replace company KPIs with internet averages.

    Make financial conclusions without internal inputs.

12. 12. Calculation must be a tool

    The function returns numbers to the brief

    Parameters — Inputs describe load, automation, cost, implementation and risk.

    Result — The function returns saving, net effect, payback and recommendation.

    Fallback — No function calling → Code Interpreter.

13. 13. 03. Agent assembly

    From system instructions and knowledge base to a calculation tool and first run

14. 14. System instruction defines behavior

    Clear rules beat long prompt

    Prepare a CEO/COO/CIO executive brief.

    Facts from knowledge base first.

    Separate facts, assumptions, risks and options.

    Missing data → owner questions.

15. 15. Participants get a small simulated company

    Short docs keep focus on the agent

    Context — Lower support load; protect VIP/PII.

    KPIs — 120k tickets/month, 62% common, 210 RUB.

    Constraints — Common cases only; disputes/VIP to humans.

16. 16. Minimal calculation schema

    Stops economics-by-vibe

    Input: tickets/rate/cost/implementation/AI-cost/risk.

    gross_saving = tickets × rate × cost.

    net_effect = saving - ai_cost.

    payback = implementation / net_effect.

17. 17. Launch prompt checks loop

    Documents + search + calculation

    Launch AI support assistant next quarter?

    Prepare a one-page board brief.

    Compare launch, pilot and postpone.

    Show economics, risks and checks.

18. 18. 04. Quality checks

    The agent is ready only after scenarios where it fails safely

19. 19. Test stable behavior

    Each scenario changes the conclusion

    Three runs

    Optimistic: high coverage, low risk.

    Cautious: data not ready, pilot.

    No launch: risk > benefit.

    Compare

    Does the recommendation follow inputs?

    Does the agent reveal gaps?

    Next steps, owners, revisit conditions.

20. 20. How to evaluate the result

    We judge decision usefulness, not text polish

    Facts: internal docs, no source substitution.

    Calculation: reproducible economics.

    Risk: guardrails, stop conditions, approval.

    Format: concise and executive-readable.

    Good agents aid decisions; bad agents write confidently

21. 21. Human approval needed

    Safety is part of the product

    Approve budget, timelines or impact.

    Recommend launch without risk/process owner.

    Use PII against security policy.

    Change SLA, communications or escalation rules.

22. 22. 05. Decision review

    Final debrief: where the agent helps and where it must be constrained

23. 23. Lab plan

    Build the minimum working loop

    0-20: why agent, task, lab docs.

    20-40: agent and system instruction.

    40-70: knowledge, facts, calculation.

    70-90: scenarios, guardrails, debrief.

24. 24. What each team should have

    Working agent + path to org rollout

    One-page brief for AI support launch.

    Knowledge base + context + calculation tool.

    Facts ≠ estimates; risks are visible.

    Platform backlog: knowledge, tools, evals, access, cost.

    Bridge from learning to AI-platform work

25. 25. Links and materials

    Docs and lab pack

    Yandex AI Studio agents/search/function docs.

    Lab pack + guide on GitHub.

    Previous lectures in the series.