Latest episode
What is agentic AI? It is software that acts on a goal and records its decisions.
Agentic AI is software you give a goal, a set of tools and permission to act. Finance teams use it on 4 jobs that come round every month.
Agentic AI is software that takes several steps towards a goal without being prompted for each one. An agentic system reaches a specific goal with limited supervision. In finance the goal is usually a reconciliation, a mapping or a first draft of the commentary.
- The platforms' agents
- OneStream (EPM software) sells SensibleAI Agents, and Abacum (FP&A software) sells Abacum Intelligence. Anaplan, Pigment and Board each have agents on their planning platforms.
- The constraint
- An agent reads the definitions held in the data model. If the model has 2 definitions of gross margin, the agent gives 2 answers and sounds sure of both.
- Where we sit
- Constancia is an EPM consultancy. We survey the data model before an agent goes on it, then fix the price of making the model ready.
Agentic AI in finance works through a task in the close or the plan step by step, and a named person signs the result.
An agent takes a task in the close or the plan, works through it and records each decision it makes. A named person in finance then signs the result.
The task has a goal, such as matching the intercompany balances for the period. The agent chooses the steps it takes to get there.
OneStream's SensibleAI Agents work on long financial tasks and turn a question into a real calculation. There are 4 of them: a finance analyst, a search agent, deep analysis and a forecast agent.
Abacum Intelligence runs AI inside the Abacum model. Every answer comes from how the business actually works, so the same question gets the same answer next time.
Finance teams use agentic AI on 4 monthly jobs, and a person can check every answer.
- 01
Mapping a new ledger
The agent proposes the mapping from an acquired entity's accounts to the group chart and lists the accounts it could not place.
- 02
Reconciliations
The agent matches intercompany balances and bank items, clears the matches within tolerance and hands the rest to a person.
- 03
Variance commentary
The agent drafts why actuals differed from plan by entity and by line, and an analyst edits the draft before the pack.
- 04
Forecast questions
The agent answers questions about drivers, accuracy and scenarios, working from the figures in the actual forecast.
The sign-off stays with a named person in finance for anything the board reads.
Finance teams use 3 types of AI, and an agent is a language model with tools and a record.
Statistical models forecast from history, and language models draft text. An agent is a language model with tools, a goal and a record of what it did.
All 3 types read the same data model, so they all pick up its faults. The agents inside an agentic system are machine learning models that copy human decisions to solve problems as they happen.
Anaplan combines fixed calculations with probabilistic AI on 1 platform. In a group close the arithmetic always gives the same result, while the agent's judgement can vary, so a person checks the judgement.
An agent needs 4 things in the model before it touches the close, and a record of each decision afterwards.
Before it touches the close, an agent needs 1 entity list, 1 chart of accounts, 1 calendar and an owner for each definition it reads. A definition with no owner is left to the agent, which settles it differently each month.
The fifth need is a record of every decision the agent has made. An agent with no record repeats its past decisions and sounds just as sure the second time.
Say 1 entity's policy sets an intercompany tolerance of £5,000 and another's sets £500. An agent clears a difference between those 2 limits in the first entity and escalates it in the second.
A chatbot and an agent differ on 6 counts, and the agent keeps a record of what it read for the audit.
| A chatbot | An agent | |
|---|---|---|
| What it is given | A prompt and whatever file is pasted in. | A goal, a set of tools and permission to use them. |
| How many steps | 1 answer to 1 prompt. | As many steps as the goal takes, in an order it chooses. |
| What it reads | The document in front of it. | The ledgers, the model and the policies it is connected to. |
| What it records | The conversation. | Each step, what it read and what it decided, for the reviewer. |
| Who signs | Nobody, and the answer is pasted onwards. | A named person in finance, on anything the board reads. |
| Where it fails | On the fact it was never given. | On a definition with 2 owners or none. |
Agentic AI in a 12-entity group works from the definitions the group has agreed, and our survey lists them.
In a multi-entity group an agent reads the consolidation, the ERPs and the plan. The agent works well on the entities where all 3 agree.
Constancia is an EPM consultancy, and we survey those 3 systems before an agent goes on them. The survey lists every definition with 2 owners or none, because the agent will answer twice on each one.
We set the price of fixing those definitions after the survey, and the price is then fixed. In a group that closes in 3 ERPs and a workbook, the survey starts with the workbook.
Finance teams often ask 3 more questions about AI agents, and we answer each one here.
How is AI being used in the finance industry?
Banks and insurers use AI for fraud detection, credit scoring and reading documents. Corporate finance teams use it for reconciliations, mappings and commentary. Those jobs come round every month, so they suit an agent.
What are common AI agents?
Finance software usually has 4 kinds of agent: an analyst that answers questions on the model, a document search agent, a matching agent and a forecaster. OneStream (EPM software) has 4 agents of this kind in its SensibleAI range. Board has a controller agent for the close.
Which is the best AI agent right now?
The best agent reads your group's own model and keeps a record of each step. A demonstration on another company's data cannot show either of those. You can test an agent on 1 month of your own intercompany matching before you buy.
Our agentic AI service and 2 more articles on AI in finance take the subject further.
Agentic AI for finance
We get the 4 things an agent reads ready before the close, for a fixed price.
Read about the serviceAI use cases in finance
Some AI uses last in a group, and others fail in the second month.
Read the use casesGenerative AI for finance
A language model drafts finance text well, but it cannot check its own work.
Read about generative AI
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.