Part of the AI in finance series.
An AI agent in FP&A is software that is given a goal and takes a sequence of actions to reach it, deciding for itself which steps to run and when it is finished. That is a materially different thing from an AI model that answers a question you have already framed. Group finance teams that have seen AI-assisted forecasting demos are now being sold agents - and the distinction matters because the failure modes, the data requirements, and the governance obligations are entirely different.
What is the actual difference between an AI model and an AI agent?
A model responds to a prompt; an agent acts on a goal. If you ask an AI tool to draft variance commentary on a revenue shortfall, that is a model answering a question you framed. An agent, by contrast, is given a standing instruction - say, flag any planning variance above a materiality threshold, identify its driver, and route it to the responsible owner - and then runs without being asked each time. The human defined the goal; the agent determines the steps.
In practice that means: the agent watches the plan against actuals continuously, applies a materiality test it has been configured to use, queries the underlying data to identify what moved and why, and raises it. A human may never touch that loop unless the finding crosses a threshold that warrants escalation. The FP&A analyst is pulled in at the conclusion, not at the beginning.
What genuinely changes in FP&A?
Three things shift in a way that periodic AI-assisted analysis does not capture.
Continuous monitoring replaces periodic review. Traditional FP&A is structured around the month-end cycle because humans and manual tooling run in batches. An agent that can read from a live data warehouse is not constrained by that rhythm. A material movement in operating cost in week two of the period does not wait until the management accounts are produced to surface; it surfaces when it happens. The compression of the loop between a number moving and someone knowing why is the headline operational change.
Driver investigation runs without a human framing each query. When an analyst investigates a variance today, they bring a hypothesis to the data - they suspect it is volume, or price, or a one-off accrual, and they query accordingly. An agent applies a defined decomposition logic to the same data without that framing step. That removes a layer of investigative latency, but it also means the quality of the conclusion is entirely dependent on the quality of the decomposition rules and the underlying data model. If the cost centre mapping is wrong, the agent's confident answer is confidently wrong.
The planning process becomes a surveillance function as well as a planning function. Agents blur the boundary between the plan-and-reforecast cycle and the reporting-and-control cycle. That is useful - but it requires the EPM data model to carry both workloads cleanly, which not all current implementations do. Our earlier piece on AI in FP&A forecasting covers where the model-quality bar sits for assisted forecasting; agentic work sets it higher.
What does an agent demand in return - and where do pilots stall?
Agentic deployments stall for four reasons that are worth naming plainly before a vendor demo.
A defined goal and a bounded action set. An agent without a precise objective drifts. Before deployment, the finance team needs to specify: what constitutes a material variance, what data sources the agent may query, what actions it may take (flag, escalate, reforecast a line, do nothing), and what it may not touch. That specification work is the hard part of the project, and vendors often understate it.
A data model that is good enough. Every conclusion an agent reaches is only as reliable as the data it reads. If your actuals feed is incomplete at T+1, if cost allocations shift between periods, or if the hierarchy used in the agent's decomposition logic diverges from the hierarchy in your consolidation tool, you will get plausible-sounding findings that are wrong. Fixing the data model is not a project that runs in parallel with the agent pilot - it is a prerequisite. The same lesson applies in the close context, as we covered in the piece on why AI financial close pilots stall.
An audit trail that matches human standards. Every action an agent takes - every query it runs, every variance it flags, every escalation it routes - needs to land in the same audit log as a human action would. Finance functions are regulated, and the argument that an automated finding is outside the normal approval and review framework will not hold with an auditor. Build the audit trail into the design, not the remediation.
A named owner. Somebody in finance has to own what the agent does. That means owning its configuration, reviewing its outputs on a defined cadence, and being accountable when it raises something incorrectly or misses something it should have caught. Ownership without sufficient authority to change the agent's configuration is not real ownership - the person responsible needs access to the goal definition and the action set.
Which platforms are moving and what should finance teams track?
Planning Sentinel-style agents - persistent, goal-driven monitors sitting inside EPM environments - are arriving in the major platforms. Vendors including those in the CCH Tagetik and Anaplan ecosystems are building or acquiring this capability. The sensible posture for a group finance team right now is to track what is arriving in the platforms already on contract, run a scoped pilot against one process with a clean data model, and set the audit and ownership requirements before the pilot starts. Waiting for the category to fully mature is a reasonable hedge; waiting without tracking is not.
The Finance Value Score Maturity Matrix scores your FP&A function on the 1-5 scale from Manual to AI-embedded. Most group finance functions that have seen AI demos are sitting at level 3 (Integrated) or early level 4 (Automated) in forecasting. Agentic capability is the frontier between level 4 and level 5. The value at stake in moving there is worth costing before the next vendor conversation.
Common questions
What is the difference between AI-assisted FP&A and agentic AI in FP&A?
AI-assisted FP&A tools respond to questions a human frames - for example, drafting commentary on a variance the analyst has identified. An agentic AI is given a goal, such as monitoring plan versus actuals and investigating material variances, and determines for itself which steps to take and when the task is complete. The human defines the objective; the agent runs the process.
What does an FP&A agent actually do that an AI model does not?
An FP&A agent monitors data continuously, applies a materiality test, investigates the driver of a variance without waiting for a human to frame the query, and routes findings to the responsible owner - all without being prompted each time. The compressing of the loop between a number moving and the finance function knowing why is the core operational change.
Why do agentic AI pilots in finance stall?
The four most common causes are: an objective that is too loosely defined, a data model that is not clean enough for the agent's conclusions to be reliable, an audit trail that does not meet finance governance standards, and no named owner accountable for what the agent does. These are design and data problems, not technology problems.
Are agentic AI tools available in EPM platforms now?
Planning Sentinel-style agents - persistent, goal-driven monitors embedded in EPM environments - are arriving in the major platforms, including those in the CCH Tagetik and Anaplan ecosystems. They are worth tracking for teams already on those platforms rather than waiting for the category to mature further.
What data quality is needed before deploying an FP&A agent?
The agent's conclusions are only as reliable as the data it reads. Actuals feeds need to be complete and timely, cost allocations need to be stable across periods, and the hierarchy used in the agent's decomposition logic must match the hierarchy in the consolidation environment. Data model remediation is a prerequisite for an agentic deployment, not a parallel workstream.