AI seminar for the Kazatomprom finance function
In June 2026 Kazatomprom’s corporate centre brought the finance directors of its subsidiaries together for an off-site called “The finance director of the future”. I was invited to run the AI block: a talk on day one and a practicum on day two. This was my second training for the group. In 2025 I ran a two-day programme for staff from several departments. This time the audience was narrower: finance people, their Excel files, budgets, procurement and reporting.
- 01 A talk on what AI already does in finance, with a source behind every number on the slides
- 02 A morning of practice: P&L model, dashboard, report consolidation and reconciliation on participants’ tasks
- 03 The rule to take home: one process, measure before and after, a human owns the result
Everyone has heard of AI, few use it
At the start of the talk I asked who uses AI at work every week. Few hands went up, and most of them meant fixing a text in ChatGPT. Yet a finance director’s day is almost entirely made of things AI can already do: spreadsheets, reconciliations, reports, models and explanations of numbers.
The group also has strict security rules, and external services are not allowed everywhere. So the seminar was not about selling a tool. It had to show how the work itself is changing and how to check AI output well enough to sign under it.
- Explain what changed in a year: AI no longer just answers questions, it gets work assigned.
- Show examples from banks and finance teams with numbers and sources, not a news recap.
- Walk through where people get burned by AI and why a human is still accountable.
- Give a plan people can start on the next day.
And the room had its own questions.
First what is happening in the world, then hands-on
Day one was an overview without the hype. Every claim on the slides had a source: Anthropic and METR research, Stanford and MIT, Moody’s, JP Morgan, Citi. Where a number sounded too loud, I said so. Day two was practical: “I am more of a practitioner. Yesterday I showed what is happening in the world. Today I will show what I actually do, on your tasks.”
- 01
What changed in a year
AI stopped being a chatbot: it now breaks a task into steps, writes small programs and works for hours. Programmers felt it first, finance is next.
METRClaude for ExcelFinance is next - 02
Examples from finance
A credit memo takes 2 minutes instead of 40 hours. Accountants with AI close the month 7.5 days earlier. The head of JP Morgan built a swaps analysis in 20 minutes.
- 03
Where people get burned
Deloitte refunded the government $440,000 for a report with invented quotes. Lawyers filed 1,353 cases that AI made up. The takeaway: AI is an intern, and an intern’s work gets reviewed.
- 04
A plan and three habits
Pick one process, measure how long it takes now, hand it to AI and measure again. AI first, then your hands. Brief it like a new hire. Check it and teach it your way.
- 05
Practicum on tasks from the room
I asked participants to send tasks in advance. Two came from real work: reconciling the budget with the procurement plan and analysing the sulfuric acid market. All data was synthetic, so no real figures left the company.
P&LDashboardConsolidationReconciliationMarket analysis
Every result comes with a check
I ran the demo from my laptop while the room suggested edits and set traps. Tasks ran in parallel: while one was computing, the next one started. That is part of the lesson too: with an agent you no longer wait, you distribute work.
The point of the practicum was not the polished output but the ways to verify it. The model kept a log of its steps on a separate sheet, reconciled totals against the source files and reviewed work in a fresh chat with no context. The room deliberately broke one formula by multiplying net profit by 0.8, and the model caught it on its own.
A one-year P&L model for a coffee shop in Almaty, three scenarios. One line of prompt, live formulas in tenge. The model found its own error in cumulative profit and fixed it.
An audit in a clean chat: no calculation errors, but three methodological gaps. No depreciation, cumulative profit is not payback, and break-even is about 107 cups a day, not 120.
A subscription business model dictated by voice. When the room asked why the model did not clarify the product, we told it to ask questions first, and it switched to a clarification mode.
A web dashboard from an Excel file in 8 minutes from a one-line prompt: filters by scenario and business unit, charts, a mobile layout. Then the same model produced a PowerPoint version.
Consolidating reports from four branches in different formats and currencies. The model kept a log sheet and a reconciliation sheet: 0 formula errors in 1,988 cells, every control total matched the sources.
Reconciling the budget with the annual procurement plan by item, quantity, procurement code and timing. A participant described the task by voice the day before, and the transcript went to the model as the brief.
A memo on whether a sulfuric acid plant in Kazakhstan makes sense, via Deep Research: 429 searches, criteria first, then analysis, the score at the very end. We showed it as a format to copy, not as finished research.
What changed
Finance staff saw AI on their own tasks
Not on abstract examples from a deck, but on P&L models, budgets, procurement and branch reports they know from their work.
They got a way to check the output
A step log, a reconciliation sheet, an audit in a fresh chat and links to sources. With these, an AI draft can be accepted instead of redone.
A plan for tomorrow, not just inspiration
One process, a before and after measurement, three habits. Participants received certificates, and the next meeting of the finance club was set for December.
Want a seminar like this for your finance team?
I will build the programme around your processes: budgets, procurement, reporting, your accounting system and your data rules.