Skip to content

AI use cases in finance that hold up all have a checkable answer.

A controller can mark a model's account mapping right or wrong in a minute. A summary of the board pack has no answer a controller can mark like that.

The AI use cases in finance that work are the ones the ledger can check.

AI use cases in finance are the tasks in a group's close and forecast that a model can do and a named person can mark. Finance teams asking about AI today have a board mandate for it, with a date attached. At Constancia, an EPM consultancy, we count 6 use cases that pass.

A group of 12 entities on 3 ERPs matches intercompany balances across 66 entity pairs each month, by arithmetic. Each pair either agrees or shows a difference with 2 invoices behind it. Intercompany matching has a checkable answer, and every use case on our list shares that shape.

3 subsidiaries in a group can book the same customer under 3 different names. A model asked for revenue by customer then returns 3 customers and a confident total.

The finance team has no way to check that total, and the use case fails at its first board meeting.

6 use cases in the close and the forecast have an answer the group controller can mark.

  1. 01

    Mapping an acquired ledger to the group chart

    The model proposes a group account for each line of the new ledger from the mappings it has seen, and the controller accepts or corrects each.

  2. 02

    Matching intercompany balances

    The model pairs the balances across the 66 entity pairs, flags the 4 that differ and shows the invoices behind each difference.

  3. 03

    Checking the balance sheet reconciliations

    The model reads each reconciliation, names the ones whose supporting balance moved since the sign-off and leaves the rest as they were signed.

  4. 04

    Drafting the variance commentary

    The model drafts why each line moved against plan, entity by entity. The analyst rewrites the 2 it got wrong before the pack goes out.

  5. 05

    Watching the close calendar

    The model reads the calendar as tasks finish, names the entity late on day 3 and shows which upstream task it is waiting on.

  6. 06

    Refreshing the forecast

    The model refreshes the forecast from last month's actuals and the agreed drivers, and the head of FP&A signs the version the board reads.

A board memo drafted from a pasted spreadsheet is the most demonstrated use case, and it has no answer to mark.

Drafting and summarising are the best-known AI use cases, and neither has an answer to check.

A model can summarise a 40 page board pack into 1 page in 20 seconds. The summary keeps every inconsistency in the original 40 pages. The reader cannot tell which figures the model weighed and which it skipped.

Asked for the CFO's commentary on gross margin, the same model writes 3 fluent paragraphs from the first margin definition it finds. A group's consolidation and sales report can define margin in 2 different ways, and the draft picks 1 without saying so.

We put drafting last on the list, after the numbers underneath are settled. Done first, drafting puts a well-written paragraph on top of a figure no one has traced.

A vendor demo of a summarising feature takes 4 minutes and needs none of the group's data, so CFOs have usually seen 5 of them. A demo of the mapping use case needs the group's own ledgers and takes a week to arrange. Most CFOs have seen none.

Every use case on the list fails the same way when the model has no record of past decisions.

The mapping model needs the mappings from the last 5 acquisitions to propose the 6th. In most groups those mappings sit in 5 workbooks on 5 different drives. The model never sees them and starts from the group chart alone.

The matching model needs the intercompany balances at invoice level, with the counterparty named the same way on both sides. Entity 7 might call entity 12 by its trading name while entity 12 uses its registered name. The model then reports a spelling as a difference.

The mapping for a 2021 acquisition can sit on a former controller's drive until the auditor asks for it. That workbook holds the record the model needs, so the record gets built before the model goes in.

The data model sets the order of the use cases, and mapping goes first because the others read its output.

Mapping goes first because the matching, the reconciliations and the commentary all read the group chart the mapping produces. Commentary written before the mapping describes a chart with 2 names for the same account.

Matching and reconciliations go second, because they produce the settled balances the forecast and the commentary read. The calendar use case runs alongside from the first month, since it reads task status alone.

Each use case gets a named person who marks its answers for the first 3 periods. The CFO reads the mark rate each period.

A controller correcting 40% of mapped lines in month 1 and 4% in month 3 shows the model learning the group's chart. A model stuck at 40% is reading a chart with no record behind it.

The forecast refresh goes last because it reads everything above it: the mapped chart, the matched balances and the signed reconciliations. Refreshed first, the forecast would rest on last quarter's unmatched intercompany balances.

We survey the group's consolidation, ERPs and spreadsheets before we name the use case that goes first, and we fix the price after that survey. The survey shows which of the 6 use cases the data can carry this year, and which need the record built first.

See more on our

Subscribe on YouTube
  • Latest episode

    How an EPM implementation works, and why the software comes last

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.