What are AI agents in finance - and how much of "agentic finance" is real?

Workflow automation follows predefined rules deterministically - reliable, bounded, no autonomy. A copilot assists a human in the loop, suggesting and drafting while the person decides. An agent pursues a goal across multiple steps and chooses some of its own actions. The marketing blurs these because "agent" sells; the evaluation question is what the thing actually decides and acts on without a human. Most valuable finance "AI" today is copilot-grade (suggest, human approves) or smart automation - and that's appropriate, because unsupervised autonomy over money and the books is exactly where risk concentrates.

Real today: extracting and coding invoices with confidence-routed human review, drafting variance commentary, surfacing anomalies, answering spend questions from structured data, and chasing information through defined workflows. Theater: fully autonomous month-end close, an agent that "runs your AP" end-to-end without human gates, and any demo where the agent confidently executes financial actions with no approval step. The tell is the approval gate - real finance AI keeps humans on the irreversible actions; theater removes them to look impressive.

Autonomous agents operate toward goals with minimal prompting and growing independence - an active frontier in the broader AI world. For finance, the right posture is informed caution: the capability is advancing fast, but unsupervised autonomy over payments, journal entries, and vendor data is precisely where the risk is highest. Care enough to track it and to ask vendors hard questions about autonomy and gates; don't hand an autonomous agent your books because a demo impressed you.

Ask three questions: What can it do without a human (autonomy level)? What systems and actions can it touch (action scope)? What requires approval before taking effect (gates)? A well-designed finance agent has narrow scope, mandatory gates on financial actions, and full logging. Broad scope plus high autonomy plus weak gates is the risk profile to reject regardless of how capable the demo looks.

Keep a human gate permanently on the irreversible, high-blast-radius actions: releasing payments, posting material journal entries, and changing vendor banking details. These are exactly the targets of fraud and the sources of restatements, and "the agent did it" is no defense. Agents can prepare, suggest, and route these; a human should authorize them - forever, not just during calibration.

For core AP and analytics, buying agentic capability inside the platform that already holds your data and controls is usually the lowest-risk path - the gates, audit trail, and security come with it. Building internal agents makes sense only with real engineering and governance capacity, and carries the data-exposure and control risks you'd otherwise inherit pre-solved. "Wait" is reasonable for autonomous frontier capabilities; "wait" on confidence-routed AP assistance mostly means leaving efficiency on the table.

To be useful it needs invoice data, vendor records, ERP structure, approval routing, and sometimes payment context - which is also a map of everything an attacker or a malfunctioning agent could damage. The access list *is* the risk assessment: scope it to the minimum needed, log every action, gate the financial ones, and confirm the data isn't leaking to model training or sub-processors. Useful and dangerous are the same access list viewed from two directions.

Stampli describes its AI as embedded, agentic intelligence that operates across the procure-to-pay lifecycle - extracting data, applying accounting logic, routing approvals, matching documents, and flagging issues without waiting for prompts - explicitly framed as "not a chatbot" and never as autonomous finance. The distinction Stampli draws is the one that matters for evaluation: the AI acts continuously on the routine and surfaces work, but humans confirm, correct, and approve before anything posts. Agentic capability under human control, not autonomy over the books.