Latest episode
Generative AI for finance produces fluent answers, so the checks sit underneath it.
A generative model can write board commentary in 20 seconds, even when the figure underneath is wrong. Finance teams need 4 checks between the ledger and the paragraph.
Generative AI for finance writes the words, and the numbers it writes about come from the ledger.
Generative AI for finance means a language model that drafts commentary, summarises the pack and answers questions about it. Constancia is an EPM consultancy, and we think the output is only as good as the 4 checks under it.
Take 2 board reports that define gross margin 2 ways, 1 after logistics and 1 before. A model writes about both margins with equal fluency and raises 0 error messages.
The problem surfaces when a non-executive director asks which margin the paragraph meant.
In some groups 3 regional finance teams each keep their own headcount number. A model asked for group headcount picks 1 of the 3, and the other 2 regional controllers read it in the board pack as fact.
A generative draft passes 4 checks between the ledger and the paragraph before it reaches the board.
- 01
The figure traces to a ledger
Every number the model writes about links back to the consolidated cell it came from and the ledger line beneath that.
- 02
The definition has 1 owner
Gross margin, headcount and revenue by customer each carry 1 written definition and 1 named person who would defend it to the auditor.
- 03
The draft records what it read
The model keeps a list of the reports, periods and cells it used, and the reviewer sees the list beside the paragraph before signing anything.
- 04
A named person signs the paragraph
The group controller or the head of FP&A puts their name to the commentary, and the model's draft is an input to that signature.
The auditor will ask about any draft that passes only 3 of the 4 checks.
A generative model fails quietly, and finance often finds the error in the board meeting.
A failed consolidation throws a validation error and stops the close on day 3, in front of the controller. The controller fixes it before anyone outside finance sees a number.
A generative model given a figure nobody traced still writes 3 paragraphs on time. The close finishes on day 5 with the error inside the pack.
The paragraph reaches the audit committee 6 weeks later. A member asks why the margin commentary on page 3 disagrees with the segment note on page 14.
Those 2 pages came from 2 processes reading 2 definitions, and the pack gave no sign of it. The group controller then has to work out which report the model read, from a draft with no list attached.
The usual fix people reach for is a better prompt. A prompt cannot change what the model was given to read. The fix belongs in the data model, as 1 definition with 1 owner.
A quarter's commentary may have 3 authors, and 1 of them a model. Without the 4 checks, nobody in the meeting can say which paragraphs the model wrote.
Generative AI in finance works best on numbers that were settled before the model saw them.
Variance commentary works well on a consolidated result with 1 definition per measure. The analyst checks each sentence against a traced figure.
The model writes the first draft. The analyst then spends an hour on the 2 variances that need a phone call.
Questions over the board pack work for the same reason, when each answer carries the cell reference beside the number. A question about what moved in Germany gets the figure, the period and the line it came from.
A general assistant given a pasted pack answers the same question with a paragraph and no cell reference.
The annual report narrative saves the most rekeying, with each figure in the text linked to its cell. The model drafts the sentence around the linked figure, and the figure never passes through the model.
A group that runs the first 2 uses for 3 quarters sees the same pattern each time. The draft arrives in seconds and the checking takes an hour. That hour goes on 2 lines out of 42, and the controller's name stays on all 42.
In all 3 uses the model writes words around a number it did not produce. A named person checks the words against the number and signs the 20 second draft.
EPM and FP&A platforms now include generative features that read the model where it sits.
OneStream (EPM software) has generative features called SensibleAI. SensibleAI reads the consolidated model directly, with its entities and history.
Abacum (FP&A software) has Abacum Intelligence, which drafts commentary on the plan and the actuals. The commentary sits in the same tables the finance team plans in.
A platform feature passes checks 1 and 3 by design, because it reads the cell and lists what it read. Checks 2 and 4 stay with the finance team.
A general assistant working from a spreadsheet export relies on the person pasting for all 4 checks.
Constancia is an official partner of OneStream and Abacum, with consultants certified on both. We switch these features on once the survey shows every definition has an owner.
Definitions with 2 owners or none go on the roadmap first. The drafting comes after them.
Our AI for FP&A service puts these features on a planning model, and 2 more articles go further.
AI use cases in finance
The 6 use cases in the close and the forecast where a controller can check every answer.
Read the articleWhat agentic AI is
Software that works towards a goal and records each decision it takes on the way.
Read the explainerAI for FP&A
We put forecasting, commentary and scenarios on a planning model your finance team owns.
Read the service
See more on our
Subscribe on YouTube
Every Sunday, 6am
The Sunday Close.
Our weekly newsletter on AI in finance.
Yours to enjoy with your Sunday morning coffee.
- AI in finance observations from Alex (CEO) and Brad (COO)
- Newest video from our YouTube channel
- AI in finance trends we have spotted each week
Tell us the date you are working to.
The people who would do the work read every message and reply inside 2 working days.