Blog · Published · Updated
AI training for FP&A: writing variance commentary
Short answer
AI will draft a comment on a budget variance, but it does not know a cause that is missing from the data. Training for FP&A teaches people to keep the checked numbers, catch added explanations and ask the department a precise question. At Syntalith it costs from €1,150 net a day. With 4 controllers saving an hour and a half a week, it pays back in 1.4 months.
- net per day of closed training for the team
- from €1,150
- people in one training group
- up to 15
- after this time, we measure the effect on the same tasks
- 30 days
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
AI in work with text: statistics 2026
1.4 mo.
payback of €1,150 training when 4 controllers save 1.5 hours a week on commentary. €817 a month is left after licences. Scenario.
40%
According to a study published in Science (2023), ChatGPT cut the time 453 professionals spent on writing tasks by 40%.
Source: Noy and Zhang, Science (2023) ↗
25.1%
According to a Harvard Business School study (2023) of 758 consultants, tasks within the model's capabilities were done 25.1% faster with AI.
15.1%
According to Eurostat (2025), 15.1% of people aged 16 to 74 in the EU used generative AI tools for work.
32.7%
According to Eurostat (2025), 32.7% of people aged 16 to 74 in the EU used generative AI tools in the previous 3 months.
Figures checked: 22 October 2026
What to write when you know the difference but not the reason
In the exercise the plan is 200 hours of work and the actual is 250 hours, for the same period and the same unit.
- Plan
- 200 h
- Actual
- 250 h
- Difference
- 250 − 200 = 50 h
Variance
50 / 200 = 25%
The numbers are the exercise's assumption. The calculation stays in a checked spreadsheet, and the model only puts the result into words.
When AI adds that overtime was the cause, the controller stops at that sentence. Nothing in the table confirms it.
In the next part of the exercise a fictitious timesheet register appears. It confirms that the extra 50 hours were classed as overtime. We still do not know why they were ordered, so the question to the department reads: „Why were the extra 50 hours shown in the register as overtime ordered? Please point to the record that confirms this reason.”
What the participant learns on their own commentary
- 01DataThey check the period, the unit and the source of the plan and the actuals.
- 02ReaderThey decide what commentary the reader of the report needs.
- 03DraftThey ask the model for a text based on the named materials.
- 04CheckThey go back to the sources: the numbers, the direction of change, the sentences about causes.
- 05QuestionThey write a question to the department where confirmation is missing.
The trainer discusses the decisions on the participant's own text. If a number lacks a reference to the spreadsheet, the trainer shows where the reader loses the means to check it. If a hypothesis has turned into a fact, they go back to the sentence and the material that was meant to support it.
What the training costs and when it pays back
- Time saved
- 4 people × 1.5 h a week × 4.33 = 26.0 h
- Value of that time
- €907
- Licences
- $100 (€89)
- Left after licences
- €817 a month
Payback of €1,150 training
1.4 months
The hours saved are the scenario's assumption, which we check by timing the tasks before the training and after 30 days. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. ChatGPT Business Standard at $25 a seat with monthly billing, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
We time the commentary from the last month-end close, before the training and after 30 days. Controllers who already have licences bring the cost in the count down to the training alone.
How to fit learning between month-end closes
One session is enough for a shared example and a review of corrections. If the goal is monthly reporting, spread the work: the first session before the close, the participant's own commentary in the normal cycle, a review after the close. At the end the controller gets material from another period in which an explanation is missing or two sources contradict each other.
When the system needs attention
In its preview documentation for Variance analysis in Finance in Copilot, Microsoft describes analysing a pivot table and source data in one workbook and editing the generated commentary. When a controller gathers information by hand from several systems every month, an integration is needed. The scope is described on the AI in finance and controlling page, and we agree the workshop programme as part of team training.
