AI for finance: a Kazatomprom workshop
On 11–12 June 2026 I ran an AI seminar for Kazatomprom’s finance directors, covering Excel, financial models, budgets and reporting.
My approach combines participants’ own tasks with tools I use myself. I prepared the examples with two finance professionals beforehand, collected more requests on day one and worked through them with the room on day two.
- A talk on what AI already does in finance, with a source behind every number
- A practicum: P&L model, dashboard, report consolidation and reconciliation on participants’ tasks
- Every output comes with a way to check it: a person owns the numbers
AI in finance: from chat answers to finished tasks
Day one started with what has changed in software engineering. AI can now read files, write code, run checks and fix its own output. The same way of working is reaching other professions: a finance professional can ask an agent to compile a report, reconcile data or build a model.
-
The person sets the task
Hands over the files, the requirements and what a finished result looks like.
-
The agent does the work
Reads the data, runs the numbers, creates a file and checks it.
-
The person signs off
Checks totals and assumptions. Responsibility for the decision stays with a person.
Credit analysis
An agent can gather financial statements, calculate ratios and draft a credit memo. The analyst checks the sources, weighs the risks and makes the call. Moody’s and Anthropic are building tools like this for credit analysis and counterparty screening.
Moody’s: Q1 2026 prepared remarksMonth-end close
- time to close the month
- −7.5 days
- general ledger detail
- +12%
In a study of 277 accountants, AI use went hand in hand with a faster month-end close and more detailed books. These are research findings, not a promise for every company: each team has to measure the gain on its own process.
Stanford GSB researchAnd the room had its own questions
- Isn’t it risky to ask it to find errors? It will invent some
- Who is responsible if AI gets a report wrong?
- How do we double-check the numbers?
- Can we work like this with our data?
- Why didn’t it ask what the product is?
- Can it go straight to PowerPoint?
- Is it free or a subscription?
- Is the 20-dollar plan enough?
- Can it forecast the tenge rate to year end?
- Why is this better than Google?
- Does it work on a phone?
- Won’t we forget how to think?
- How do I install Claude in Excel?
- Is the file mine or the service’s?
- Which metrics show it got better?
- Will AI take our jobs?
Day two: AI on real finance tasks
On day two we moved into Excel. I demonstrated the tools on my laptop while participants changed the requirements and checked the outputs. Tasks ran in parallel: while one was processing, we started the next.
We prepared the practicum with two finance professionals and added requests from the room on day one. Every demonstration used training budgets, procurement data and reports, never real company figures.
From assumptions to a coffee shop P&L
We started with a business everyone understands: a coffee shop in Almaty. The input was an Assumptions sheet with floor area, rent, average ticket, margin, payroll and a startup budget of KZT 30 million. These were training assumptions, not the business plan of a real café.
The task for Claude in Excel was a single line: a full-year P&L with three scenarios and live formulas, so that changing an assumption recalculates the model.
Build a financial model for next year: downside, base and upside scenarios
One line of request produced a working Excel model. Next we tested whether its formulas could be trusted.
The room set a trap. Claude found the ×0.8 multiplier
Once the model was built, participants asked what would happen if we planted an error on purpose. I added a 0.8 multiplier to the profit formula and opened the file in a fresh chat. That chat had no history of building the model: the AI did not know what I had changed, or where.
Claude flagged the unexplained multiplier and spelled out its effect: the formula deducted only 80% of the tax on revenue. The recording shows not a vague remark about file quality but the exact formula and the cell where the error sits.
I could not fool the model: it found the planted error in a fresh chat. The room could check the finding against the exact formula themselves.
From Excel to an interactive dashboard in 8 minutes
The next input was a ready-made file, fake_cfo_pl.xlsx: monthly P&L data, scenarios, business units and budget. The analysis already lived in Excel. The task was to turn the table into something easy to explore and compare.
The AI built an HTML page with charts, filters and recalculating metrics. The room asked to show values inside the chart segments, and we made that change from a screenshot. Then we produced a PowerPoint version from the same model.
There is a file in the folder. Build a good-looking web dashboard from it, with filters
“There it is, the wow effect,” someone in the room said. The dashboard made the existing Excel analysis visible: the data could be explored right there in the meeting.
Branch reports and procurement reconciliation
The next examples came from day-to-day finance operations. Here the source files and the checks matter as much as the output.
Four branches sent reports in different formats and currencies. The European file used half-years while the others used quarters, and every table started on a different row. The model had to bring the data into one shape and reconcile the totals against the sources. Below is an excerpt from one of the input files.
A separate verification sheet compared row counts, period totals, actual revenue and budget with the source files. The saved workbook shows the values on both sides and a PASS status. During the demo the model also checked the formulas in 1,988 cells.
A participant described the budget-versus-procurement reconciliation by voice. We agreed on what to compare: item names, quantities, ENS TRU codes, dates and terms. The training example used simplified data. The frame shows the brief: what was approved and where the discrepancies are.
Stack
What we explored and tested over two days
We moved from the broad AI conversation to files, formulas and real requests from a finance team. For every example we covered both the output and how to verify it.
To keep practising on your own, read AI for finance professionals: tasks and prompts. What to automate across a team, and how to stay in control, is covered in AI in company finance.
Financial model
Built a three-scenario annual P&L for a coffee shop from assumptions, then found a deliberate formula error in a fresh chat.
Interactive reporting
Turned an Excel file into a dashboard with filters and charts, revised it at the room’s request and produced a presentation.
Consolidation and reconciliation
Combined four branch reports with control totals and worked through the brief for comparing a budget with the procurement plan.
FAQ
Yes. Every example at this seminar was a training one: the coffee shop model, the P&L behind the dashboard, the branch reports and the procurement requests were prepared in advance. The tasks mirrored real finance work, and no company figures went into a chat.
Claude in Excel for models and formula checks inside the workbook, Claude Code for the dashboard, consolidation and reconciliation over a folder of files, and ChatGPT Deep Research for market analysis with sources. Tasks were briefed both in text and by voice.
The way you would check a new hire’s work. Ask for a separate reconciliation sheet with control totals against the source files, review the formulas rather than just the total, and hand the finished file to a fresh chat with no history for review. At the practicum, a fresh chat found the 0.8 multiplier we had planted in a formula.
Yes. We build the programme around your work: we pick examples from participants’ own tasks in advance and follow your data rules and the tools your company allows. See how it works on the corporate AI training page.
Corporate AI training built around your team’s work
Bring this workshop to your team with a programme built around your industry and processes. We will pick tasks with participants in advance, follow your data rules and work through each task together, from the brief to checking the output.