Skip to main content
ON DEMANDEveryone builds. IT approved.Watch the replay
Sigma Computing
Fundamentals

AI agents for finance and FP&A: Key use cases and how to set them up

Neil Coleman
Neil ColemanSenior Solution Architect
November 19, 2026
15 min read
AI agents for finance and FP&A: Key use cases and how to set them up

Financial planning and analysis (FP&A) teams spend 46% of their time on data collection and validation, the highest figure in five years. They still need to find time to plan, close, report, and advise the business on compressed timelines, using data that lives across an ERP, a warehouse, a planning tool, and many spreadsheets.

AI agents for finance and FP&A change how that work gets done. Sigma handles the calculations; agents decide what deserves attention, draft the explanation, and trigger the next step.

Key takeaways

  • The workflows that pay off with AI agents for finance are repetitive and data-heavy, including forecasting and rolling projections, variance analysis and commentary, headcount and workforce planning, and expense management and approvals.
  • Finance agents fail with scattered data, inconsistent metric definitions, no governed writeback path, and missing audit trails. Fix the data foundation, permissions, and audit controls before selecting a model or framework.
  • After setting up AI agents for finance, keep a person in the loop for judgment calls. Use human-in-the-loop approval for anything touching the ledger, and human-on-the-loop monitoring for reversible, lower-stakes work.
  • MCP connectors let agents read business rules and context from Confluence, Slack, and other systems within Sigma, while figures stay in the governed warehouse.

What are AI agents for finance and FP&A?

An AI agent for finance and FP&A connects to financial data, reasons through a multi-step task, and takes an action, such as building a forecast or routing an approval. An agent carries forward a goal set by a person, uses data that person scoped, and runs independently only because someone configured it to.

A copilot, for example, can draft variance commentary when you ask for it. An agent works on a schedule: Sigma calculates the variances, and the agent reviews them, identifies which ones need commentary, drafts the explanation, routes it for review, and flags items that need a human decision.

4 traits that define AI agents for finance

A finance agent applies finance-specific concepts and controls. Four characteristics separate the ones that hold up in a controller's review from the ones that don't:

  1. They reason over financial data structures, not generic tables. Sigma’s data model encodes the chart of accounts, general ledger (GL) hierarchies, fiscal calendars, and period logic such as close periods and fiscal year-end. The agent reads these through its governed data sources, so it can tell a budget line from an actual and a closed period from an open one.
  2. They treat segregation of duties and approval controls as first-class constraints. The person who proposes a change shouldn’t be the one who approves it. An agent proposes and role-based permissions and approval actions route the proposal to a different role for sign-off before anything is saved or sent.
  3. They produce audit-ready outputs traceable to source transactions. Sigma lets users drill from any figure an agent surfaces to the underlying transactions. Agent queries and actions can be logged. That gives a controller, auditor, or regulator a record they can walk through months later.
  4. They write back to systems of record under governed, permissioned paths. The agent calls Actions that write forecast inputs, scenario assumptions, flux commentary, and approvals to the warehouse through Input Tables. The scoped path inherits row-level security, so nothing lands in an unsecured file that has to be re-keyed later.

Agents combine these traits in different ways depending on the workflow. A copilot is often the right starting point, but it still leaves the last mile of the workflow to a person.

Common use cases for AI agents in finance and FP&A

The workflows where finance agents deliver the most value share the same structure: they are repetitive, data-heavy, and decision-supporting rather than decision-making. Sigma handles assembly (pulling actuals, calculating variances); the agent handles triage, explanation and exception handling while a person retains judgment over what the numbers mean and what to do about them.

  • Agents handle forecasting and rolling projections. 29% of organizations need more than 10 days to produce a forecast. Sigma holds actuals, driver logic and assumptions; the agent runs the defined forecast, states assumptions, and presents results for review and approval.
  • Agents automate variance analysis and commentary. Sigma calculates budget-versus-actual and period-over-period variances from GL data, Sigma agents identify drivers, and draft first-pass narratives. Useful commentary names the line that missed the plan, the amount, and whether the run rate changed.
  • Agents support headcount and workforce planning. Personnel costs are often a major part of organizational budgets. Sigma provides the budget-to-actual variances on people costs and flag delayed backfills, unplanned compensation adjustments, and benefits increases that would otherwise take days to trace. Agents explain drivers and flags delayed backfills or comp adjustments
  • Agents route expense management and approvals. Traditional expense auditing commonly examines only a subset of transactions. AI audit agents review 100% of submissions for duplicates, fake receipts, and policy violations using your rules plus its reasoning, then route out-of-policy items to the right approver and fast-track compliant ones through Actions.

Each of these workflows only pays off if the agent's output is reliable, and reliability is where most finance AI pilots fall short.

Why AI agents for finance don't consistently produce reliable results

Most finance AI agent pilots stall for the same reason: the data foundation underneath them is not ready. Among leaders whose AI projects hit setbacks, 38% cited poor data quality or limited data availability as a direct cause of failure. Data quality, permissioned access, and audit controls have to be in place before an agent runs its first task.

Financial data scattered across spreadsheets and disconnected systems

When "revenue" resolves differently across three systems and half the reporting logic lives in undocumented spreadsheets, an agent has no reliable ground truth. It reads whatever it is pointed at and returns almost-right numbers, and almost right is dangerous when the decisions involve real money.

Half of organizations work across eight to 10 different sources and reporting systems, and nearly all still rely on planning spreadsheets for planning and reporting. Standardized metric definitions, documented lineage, and a single governed source have to come before the agent.

No governed path for an agent to write back to the numbers

An agent that can only read stops at a summary. A person then re-keys the output into the plan, recreating the manual step the agent was meant to remove. Governed writeback combines technical plumbing with financial controls. An agent that can both recommend a financial action and execute it, with no independent approval step, eliminates the segregation of duties that controls exist to enforce.

Incidents involving autonomous agents with unchecked access underscore why writeback needs a scoped, permissioned path that keeps proposal and approval in separate roles.

Missing audit trails for automated decisions

Without a log of what the agent read, what it decided, and who reviewed it, no one catches its errors before they reach a board deck or a filing. That matters more when the model is wrong: 86% of CFOs say their finance team has encountered hallucinated data while using AI. "The AI said so" doesn't survive a Sarbanes-Oxley (SOX) walkthrough, and neither does an agent whose reasoning can't be reconstructed after the fact.

How to set up AI agents for finance to deliver consistently reliable results

Reliable output is decided before the agent runs its first task. The steps below map to the failure modes above: naming the decision and preparing governed data prevent almost-right answers, scoping permissions and approvals preserves segregation of duties, and logging every action makes the work audit-ready.

1. Define the decision and set a baseline

Start by naming the management decision the agent supports. Without a named decision, there's nothing to measure the agent against. Document current cycle times, error rates, and rework volume for that decision before launch, so improvement can be measured rather than assumed once the agent is running.

2. Prepare governed data and metric definitions

An agent is only as good as the data it reads and the system of record it writes to. Standardize metric definitions first in Sigma, so "gross margin" means one thing regardless of how a prompt is phrased, and require documented data lineage and data freshness before the agent touches a dataset.

3. Scope permissions and approval gates

Give the agent the minimum access the task requires. List which processes it may perform and which it may not, based on compliance risk and financial harm. Read-only analysis and anomaly detection carry low risk. Drafts and field updates that a person reviews carry medium risk. Anything that approves transactions or moves money carries high risk and requires a human approval gate.

Then decide, per action, whether a person approves before execution or monitors after. Human-in-the-loop review pauses the agent at defined checkpoints and waits for sign-off. Use it for disbursements, journal entries, and anything touching the ledger. For reversible, lower-stakes work, human-on-the-loop monitoring allows the agent to act while a person retains the ability to intervene afterward. Design these checkpoints into the agent's architecture from the start.

4. Configure and test the agent

With the decision, data, and guardrails set, configure and test the agent. In Sigma, a builder defines the agent’s instructions, the workbook data it can read, and the tools it can call, such as Actions that write to Input Tables, MCP connectors, and warehouse agents. Keep calculations, like combining actuals and budget and computing variances in the workbook, so they stay deterministic, and reserve the agent for judgment tasks like deciding which variances need commentary and drafting the explanation. For complex processes, split the workflow into scoped agents, or have a Sigma agent orchestrate warehouse agents, so each one has a narrow job and limited access. Build an evaluation suite of known scenarios, such as variances that should and shouldn’t be flagged, and check the agent’s selections and drafted commentary against it before each release. Pilot on one bounded workflow, such as monthly flux commentary for a single cost center, before broadening scope.

5. Log and monitor every run

Log the data sources and decisions behind each run, along with any human override. That log lets a controller trace any output back to source transactions when the board asks why a number changed. Monitor continuously, since set-and-forget assurance fails as model outputs drift.

How Sigma enables AI agents for finance and FP&A

Sigma gives finance and FP&A teams a governed foundation for building and running AI agents on the cloud data warehouse (BigQuery, Databricks, Amazon Redshift, Snowflake, and others) where their financial data already lives. Because Sigma connects directly to the warehouse your team already secured, finance teams get support for reporting across billions of rows and inherited row-level security.

Governed AI tools sit inside a spreadsheet-like interface, where business users can add numbers, comments, and approvals the same way they would in a spreadsheet. And because the warehouse holds more than finance data, an agent built on Sigma can reason over financial results alongside sales, product, operations, and customer data, for a cross-functional view that tools limited to finance-system data can't provide.

For a finance agent, that foundation supports the move from pilot to production. Permissions, audit, lineage, and change management stay in one control plane, so agent artifacts use your existing governance from the moment they're built.

Sigma Agents that build forecasts and diagnose variances

A builder configures a Sigma Agent inside a workbook with plain-English instructions, the tables and columns it can read, and the actions it can run. Configured for finance, a Sigma Agent can build a forecast, draft variance commentary that drills from a consolidated P&L to payroll-level detail, flag transaction anomalies, and route exceptions for review. It can answer questions conversationally, propose and wait for approval before writing or sending anything, or run autonomously on a schedule.

A governed foundation for building and running your own agents

Sigma generates SQL for agent queries and executes it in the connected cloud data warehouse, keeping source warehouse data in place. A Sigma Agent uses the caller's permissions and is restricted to data available to that caller. The AI runs through your own stack, whether that is Databricks model serving, Snowflake Cortex, or customer-selected models.

The queries and actions it takes can be logged, creating a traceable record in a format compliance teams already understand. Users can also inspect the underlying tables behind workbook values, so finance can validate results before they land in a board deck.

Writeback through Input Tables for plans and approvals

Forecast inputs, scenario assumptions, journal entry drafts, flux commentary, and sign-offs are written back to the warehouse through Input Tables, in the same workbook where the team views results. Input Tables write new data to a separate schema, so new records go there, preserving original warehouse data and an auditable change history.

Sigma Actions can chain approvals, alerts, and writeback off a single trigger, such as a variance threshold, so a flagged item can route to the right approver without anyone leaving Sigma or bouncing a file between tools.

MCP connectors that bring business context to the numbers

The warehouse holds the numbers, but the rules and reasons behind them usually live somewhere else. Allocation methods sit on a Confluence page, a capitalization policy sits in SharePoint, and the explanation for a missed forecast is buried in a Slack thread. MCP connectors let a Sigma agent reach those systems through the Model Context Protocol (MCP), an open standard for connecting AI agents to external tools and data. An admin adds an MCP server once and grants access to specific users or teams. Builders then attach the connector to the agents that need it.

MCP extends the agent's reach without weakening governance. Figures still come from governed warehouse tables, while MCP supplies the rules and explanations around them. Any action in an outside system, such as creating a ticket or posting a message, goes through the same approval gates as writeback.

Build AI agents for finance and FP&A with Sigma

Finance agents fail because the foundation underneath them isn't ready. Almost-right numbers, missing approval gates, and reasoning that cannot be reconstructed later are the failures a controller, auditor, or regulator will surface first.

Building on Sigma addresses these risks from the start: security and permissions are inherited from the warehouse your team already secured, writeback can run through scoped access and approval paths designed to preserve segregation of duties, and agent actions are logged in one control plane. The work goes from question to outcome without leaving the warehouse.

Pick the one workflow that costs your team the most days each month and start there. Get a demo or try Sigma free.

FOLLOW SIGMA

Related articles

Get started