How do you trust a figure that came out of an AI tool?
Analysts and finance staff now ask AI tools for totals, growth rates and variances, and the answers come back as tidy tables. Unless a tool did the calculation, each figure in those tables is a prediction. A model that writes a sound analysis plan can still get a comparison of 2 decimals wrong in the same reply.
This course sets up a way of working where the AI computes with code or spreadsheet formulas the analyst can read, and every figure is reconciled to the source before it reaches a report. Participants also learn which data may be uploaded, how to record AI use for audit, and which analyses stay with a person.
After this course
- Can you explain why a language model can get simple figures wrong and tell a predicted number from a calculated one?
- Can you get the AI to compute with code or spreadsheet formulas you can read and run yourself?
- Can you brief an AI for analysis with a data dictionary, one question, a method and an output format?
- Can you verify AI-generated numbers by recomputing, reconciling totals and tracing results to source rows?
- Can you prepare a dataset for upload so personal data stays out of tools not approved for it?
- Can you record AI use in each report and keep decisions about individuals with a person?
Skills you practise
- Briefing for analysis
- Tool-based calculation
- Reconciliation
- Source tracing
- Data minimisation
- Audit notes
- Bias awareness
Course structure
The course is 1 day of 6 teaching hours, in 2 blocks of 3 hours with a lunch break. Online, it runs as 2 live sessions of 3 hours on the same day, not recorded. Exercises use anonymised or synthetic datasets with the same structure as the team’s own reports, agreed with the client before the course.
The morning shows where language models go wrong with numbers. Participants ask the same questions twice: once for a typed answer, once for code or a spreadsheet formula they run themselves. They compare the 2 results line by line. They then write a full brief for a recurring analysis: data dictionary, question, method and output format. The morning ends with preparing a dataset for upload, removing or aggregating personal data, and checking which approved tool may take it.
The afternoon starts with verification. Participants recompute the main figures, reconcile totals to the source, trace a sample of results back to source rows and hand the method to a second checker. Each check is timed, so the team sees which ones are fast enough to run on every report.
The last 75 minutes cover the AI-use note for reports and the analyses that stay with a person, including decisions about individuals under Article 22 GDPR. The day ends with the team’s briefing template, reconciliation checklist and AI-use note format in final form.
About two thirds of the time is hands-on, in the AI tool your organisation has approved.
1Where language models go wrong with numbers45 min
- Sound reasoning next to wrong arithmetic
- Arithmetic slips in long calculations
- Confident figures with no calculation behind them
- Telling a predicted number from a calculated one
A model can lay out a sound analysis plan and still add up a column wrongly in the same answer. Check every figure, however well the reasoning around it reads.
Caution. Recompute every figure taken from a chat answer before it goes into a report.
You leave with: A list of the participant's recurring calculations, each marked: compute with a tool, or keep by hand.
2Making the AI compute with code and formulas75 min
- Code execution in your approved tool
- Spreadsheet formulas you can inspect
- Reading code well enough to check it
- What to do when the tool cannot run code
A model produces numbers the way it produces words, by predicting the next piece of text. A figure is a prediction unless a tool computed it, so ask for code or a formula and run it.
Caution. Read the code or formula before you trust the result, because a wrong cell range gives a clean answer that is wrong.
You leave with: 3 recurring calculations from the participant's work, each rebuilt as a formula or short script with a 1-line explanation.
3Briefing an AI for analysis45 min
- A data dictionary: what each column holds
- One business question per brief
- Method and assumptions stated up front
- The output format for the report
The model sees only the data and words you give it in the chat, and it guesses what a column holds from its name. A data dictionary replaces those guesses with the definition of each column.
Caution. In June 2024 a Dutch court (ECLI:NL:RBGEL:2024:3636) used ChatGPT to estimate solar panel lifespan and the price per kWh in a damages calculation. A figure with no stated inputs or sources leaves the reader unable to check it.
You leave with: A briefing template for analysis, with the 4 fields filled in for one of the participant's recurring reports.
4Uploading datasets safely45 min
- Personal data in rows and free-text fields
- Removing or aggregating identifiers before upload
- Approved tools only, and where files are kept
- Files from outside sources that carry hidden instructions
A dataset from outside can carry instructions in its text fields, and a model may follow them and leak what it was given. Upload only the columns and rows the analysis needs.
Caution. Personal data uploaded to a chatbot is processing under GDPR, and in August 2024 the Dutch data protection authority reported breach notifications after staff entered patient data and customer addresses into chatbots.
You leave with: A data preparation checklist: which columns to drop, which to aggregate, and which approved tool may take the result.
5Verifying AI-generated numbers75 min
- Recomputing key figures independently
- Reconciling totals to the source
- Tracing a sample of results to source rows
- A second checker for the method
- Asking 'are you sure?' and why it fails
A figure is ready to use once someone has traced it to the source. Make each check quick enough to run on every report.
Caution. A study published in 2024 found that models struggle to correct their own reasoning without outside feedback, so the check runs against the source data.
You leave with: A reconciliation checklist for the team: which totals to reconcile, how many rows to trace, and who signs off.
6Recording AI use for audit45 min
- What to record: tool, data, prompts, checks
- Which steps the AI did and which a person did
- An AI-use note for each report
- Keeping prompts and code with the working files
The AI-use note states which steps the AI did, from suggesting a formula to running the whole calculation, and which a person did. A reviewer then knows what to check.
Caution. In October 2025 Deloitte Australia agreed to partially refund the Australian government for a report containing AI-generated errors, and the revised report disclosed the AI use.
You leave with: An AI-use note format for the team's reports, tested on one real report.
7Analyses that stay with a person30 min
- Decisions about individuals and Article 22 GDPR
- Bias learned from past decisions
- Scores and rankings of people
- Group-level analysis as the default
Decisions about individual people stay with a person, and AI analysis works at group level by default. Where AI output on one person is used at all, it is one input to that decision, never the verdict.
Caution. In October 2018 it was reported that Amazon had dropped an experimental CV-scoring tool after it penalised CVs containing the word 'women's'.
You leave with: A list of the team's analyses that involve individuals, each marked: group patterns only, or no AI.
Where language models go wrong with numbers
Ask the model to mark each figure as calculated or estimated.
What's the average order value by region? [pastes table]
Using only the table below, give for each of the 4 regions: total revenue, number of orders and average order value. Write each average as 'revenue / orders = result'. Mark each figure 'calculated' or 'estimated'. Return a table with 4 rows. [table]
Making the AI compute with code and formulas
Ask for a formula or code, then run it yourself.
Calculate year-on-year growth for each product line.
Write 1 spreadsheet formula for year-on-year growth per product line. Data: column A product line, column B revenue 2025, column C revenue 2026, rows 2 to 48. The formula goes in D2 and fills down. Show the result as a percentage with 1 decimal, and 'n/a' where column B is 0. Explain the formula in 1 line.
Briefing an AI for analysis
Give the data dictionary, question, method and output format first.
Here's our sales data. What can you tell me?
Data: the attached CSV has 1 row per invoice for 2026. Columns: inv_date (invoice date), region (North, South, East, West), net_eur (amount excluding VAT), cust_type (new or existing). Question: did revenue from new customers in North fall from Q2 to Q3? Method: sum net_eur by quarter and cust_type for North, using code you show me. Output: a table with 2 rows and 3 columns, then 3 sentences on the result and 1 on what the data cannot answer.
Uploading datasets safely
Remove or aggregate personal data before you upload a dataset.
Here's the full absence file from HR. Find patterns by team.
The attached file has 1 row per team per month with 3 columns: team_code, month, absence_days. It holds no names, employee numbers or reasons for absence. Show the 5 teams with the largest rise in absence_days from Q2 to Q3 as a table, with the calculation in code.
Verifying AI-generated numbers
Reconcile every total to the source before you use a figure.
Double-check your numbers.
For the table above, give: the grand total of net_eur across all rows of the source file, the sum of your 4 regional totals, and the difference between them. Then list the source row numbers behind the North Q3 figure, up to 20 rows, so I can check them in the spreadsheet.
Recording AI use for audit
Record the tool, data, prompts and checks in the working file.
Add a note saying we used AI for this report.
Draft an AI-use note for the report below, under 80 words, with 5 fields: tool as shown in the interface, data used (file name and date), what the AI did, what a person checked and how, and who signed off. Write 'to complete' in any field I have not given you the facts for.
Analyses that stay with a person
Ask for group-level patterns and keep decisions on individuals with a person.
Rank these 200 applicants and tell me who to reject.
Using the anonymised application data attached (no names, ages or photos), count how many applicants meet each of the 4 published criteria. Return a table with 1 row per criterion. Our recruiters review each application and make every decision.
The team leaves with a verified analysis workflow in 3 parts. The briefing template sets the data dictionary, question, method and output for every AI-assisted analysis. The reconciliation checklist says which totals to reconcile, how many rows to trace and who signs off. The AI-use note format records the tool, the data, what the AI did and what a person checked.
In the week after the course, the team runs its next recurring report through the workflow and attaches the AI-use note. Reviewers and auditors can then see how each figure was produced and checked. Each participant also receives an attendance record for your AI literacy file.
- Teaching time
- 6 hours on 1 day, in 2 blocks of 3 hours with a lunch break.
- Before the day
- Confident spreadsheet use, regular use of an AI assistant, and access to an approved AI tool, ideally one that can run code.
- After the day
- A verified analysis workflow: a briefing template, a reconciliation checklist and an AI-use note format for reports.
Completion
You complete the course by working through each exercise on your own tasks. There is no exam and no grade. Each participant receives an attendance record for your AI literacy file (not a certificate).
Who is this course for?
Analysts, controllers, finance and operations staff, and reporting teams who work with data and already use an AI assistant. Confident spreadsheet use is the only technical requirement. Participants read code during the course; nobody has to write it from scratch.
It fits teams that produce recurring reports for management, auditors or regulators and want one standard for AI-assisted figures before those reports go out. It also fits teams asked to analyse people data, such as absence or recruitment figures, who need to know where AI stops.
Choose AI for Daily Work: Prompts, Assistants and Automations if most of your work is writing and admin. Participants who are new to AI take AI Basics at Work and then AI for Daily Work before this course.
This course is not part of a path. Book it on its own.
- Level
- Advanced
- Before you start
- Confident spreadsheet use, regular use of an AI assistant, and access to an approved AI tool, ideally one that can run code.
- Group
- 1 to 12 people, online, live or at your office
This course is right for you if you
- produce reports with figures for management, auditors or regulators
- ask AI tools for totals, growth rates or variances
- use spreadsheets with confidence
Unsure whether it fits? Ask us. We reply within 1 working day.
Practical information
- 1 day, 6 teaching hours, with a lunch break
- Online: live in 2 sessions of 3 hours on one day, not recorded; times agreed in the intake call
- At your office: 09:30 to 16:30; you provide a room, a screen and internet access
- In English, in the AI tool your organisation has approved
- Taught live by a practitioner with more than 15 years of commercial and technical leadership in software and data companies across Europe and Asia-Pacific, including maritime, energy and fintech.
- Attendance record for your AI literacy file
- Online, live: 1 day in 2 live sessions of 3 hours, not recorded. Book it as a private day for your team of 1 to 12 people, on your own Teams, Zoom or Meet.
- At your office: 1 day (6 teaching hours) for 1 to 12 people, travel within the Netherlands included.
Dates and booking
Choose the format, an open weekday in the calendar and the number of participants, then reserve on this page.
The calendar and booking form load with JavaScript. Without it, ask us for a date on WhatsApp, by phone on +31 6 20 51 73 96 or at advisory@ai-business.school.
Investment
| Item | Online, live | At your office |
|---|---|---|
| Price | €399 per person | €2,900 for the day, for your group |
| Including 21% Dutch VAT | €482.79 per person | €3,509 |
| The training day | Included | Included |
| A 30-minute intake call on your tools and data rules | Included | Included |
| Course materials and templates | Included | Included |
| Attendance record for your AI literacy file | Included | Included |
| Travel within the Netherlands | Not needed | Included |
| Group | 1 to 12 people | 1 to 12 people |
Prices exclude VAT. Online, live, you pay per participant. At your office, one price covers your whole group, from 1 to 12 people; for a larger group, book a second day. You reserve on this page; the payment details appear when you do.
Questions
Can ChatGPT or Copilot calculate reliably on our financial data?
Only when the calculation runs in a tool, such as code execution or a spreadsheet formula, and a person reconciles the result to the source. A figure typed out in a chat answer is a prediction until someone recomputes it.
Is it safe to upload company datasets to an AI tool?
Use only a tool your organisation has approved for that data, and remove or aggregate personal data the analysis does not need. Personal data entered into a chatbot is processing under GDPR, so purpose limitation, data minimisation and security rules apply.
Do analysts need Python skills for this course?
No. Participants ask the AI for code or spreadsheet formulas and learn to read them well enough to check what they do. Confident spreadsheet use is enough.
How should we document AI use in reports for auditors?
The course gives you an AI-use note format with 5 fields: tool, data, what the AI did, what a person checked, and who signed off. Teams attach it to each report that used AI.
Can we use AI to score or rank employees or job applicants?
Decisions based solely on automated processing that have legal or similarly significant effects on a person fall under Article 22 GDPR. The course shows how to keep such decisions with a person and use AI for group-level analysis.
Does this course support our AI Act AI literacy duty?
It supports your Article 4 AI literacy duty. The AI Act does not require a certificate; you keep internal records, and each participant gets an attendance record for your AI literacy file.
Last reviewed 21 October 2026.
Related courses
AI for Daily Work: Prompts, Assistants and Automations
Build a tested assistant or simple automation for a recurring task, with a check step and an owner.
€399 per person online · €2,900 a day at your office, up to 15 people
Shadow AI and Safe AI Use
Use AI at work without leaking data or falling for deepfake scams, and leave with the input for your AI policy.
€399 per person online · €2,900 a day at your office, up to 15 people