Latest episode
AI agents for finance that can show where every number came from.
An agent that cannot read last quarter's decision will make it again in a different way and sound equally certain both times. We make sure the model it reads can answer it.
AI agents for finance take the repeating steps of the close and record what they used.
An AI agent for finance is software that works through a task with several steps, such as matching intercompany balances or drafting a variance note. It records what it read at each step, so a person can check it. Constancia is an EPM consultancy that prepares the data model the agent reads.
Before an agent touches a number the board reads, we ask 3 questions of the model it will use. Has anyone read the data model, who carries the cost if it is worse than expected, and is the foundation built or assumed.
Where the mapping from each ledger to the group chart lives in 1 analyst's workbook, an agent asked to map a new acquisition has nothing to read. It guesses, and the guess looks exactly like every other line.
5 agent tasks in the close each depend on something in the model, and the fifth is a person.
Awaiting image
- 01
Account mapping
The agent proposes where a new ledger's accounts sit in the group chart, and it needs the chart with 1 owner per line.
- 02
Intercompany matching
The agent pairs the balances between entities and lists the differences, and it needs both sides posted to the same period.
- 03
Reconciliation checks
The agent flags the balance that moved unlike last month, and it needs 13 months of history on the same hierarchy.
- 04
The variance note
The agent drafts why actuals differ from plan by entity and line, and it needs the plan and the actuals on 1 structure.
- 05
The sign-off
A named person in finance signs anything that reaches the board pack, and the agent's record of what it read sits beside the signature.
The same agent gives 2 answers on 2 models, and the difference is the record.
A model with a record
The agent cites the decision from last quarter and applies it again.
Every mapping, every elimination rule and every definition sits in the model with a date and an owner. The agent reads the record before it acts, and the reviewer sees what it read.
- Mappings with a date and a name against each
- Definitions written once and read by every report
- A log of what the agent used, kept beside the result
A model with no record
The agent makes last quarter's decision again in a different way and cannot say why.
The mapping lives in a workbook, the rule lives in 1 person's memory and the definition changed in the autumn without a note. The agent reads whatever is in front of it and answers with the same confidence.
- A mapping that exists only in an export
- A rule the person who wrote it took with them
- An answer with nothing behind it to check
We put AI agents on the close for a finance function where 4 things hold.
- 01
A close across several entities
The agent's work is matching and mapping between ledgers, and 1 entity has nothing to match.
- 02
A platform the agent can read
OneStream (EPM software) or Abacum (FP&A software), with the model owned by finance.
- 03
A dated event the board is watching
An audit clock, an integration or a board that has asked for AI by a date it has already announced.
- 04
A person who will sign
A group controller or head of FP&A who will put their name to what the agent proposes, period by period.
A single-entity business with a 3 day close has no matching for an agent to do, and its spreadsheet still does the job.
SensibleAI or Abacum Intelligence, and the platform that runs your close decides which.
OneStream (EPM software)
SensibleAI agents on the consolidated model
- For a group that closes and plans on OneStream
- Agents read the same entities, accounts and periods the consolidation uses
- The record of what each agent used stays in the platform
Abacum (FP&A software)
Abacum Intelligence on the planning model
- For a finance team whose plan lives in Abacum and whose close is settled elsewhere
- Commentary and forecast drafts read the drivers the business owns
- Finance keeps the model, and the agent reads it in place
Constancia is certified on both platforms and an official partner of both vendors.
Questions finance teams ask before they let an agent near the close.
How can agentic AI be used in finance?
Agentic AI in finance works on the repeating steps of the close and the forecast: matching intercompany balances, proposing account mappings, checking reconciliations and drafting variance notes. Each task has an answer a person can verify against the ledger, and the judgements stay with the finance team.
Which AI agent is best for finance?
The best agent is the one that reads your data model directly and records what it read. OneStream (EPM software) ships SensibleAI for its consolidated model, and Abacum (FP&A software) ships Abacum Intelligence for its planning model. We implement both and rank neither against tools we have not run.
Is there a ChatGPT for finance?
General assistants exist, and the finance platforms now carry agents of their own. A general assistant sees the spreadsheet pasted into it, and a platform agent sees the consolidation with its history and its owners.
What are the 7 types of AI agents?
The standard textbook lists 7 types, from simple reflex agents that react to 1 input up to learning agents that improve from feedback. The agents used in finance are the goal based kinds, given a task with a checkable end and a record to write to.
What is the best AI for finance?
No AI fixes a model whose definitions disagree, so the best one reads a model that has been settled first. Once the definitions have owners, the choice between the 2 platform agents is smaller than it looks.
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.