What makes AI agentic?
The difference is between a filter and an investigator. Rules, OCR, and classic automation each handle a step: extract a field, flag a mismatch, or route an approval. An agent works the case. It reads the invoice and delivery note, queries the ledger, checks the vendor's history, chooses the next check based on what it found, and drafts the correction or dispute. When it cannot decide, it should stop, list the candidates, and attach the evidence rather than guess.
The flag-don't-guess model explains this control in practice. The guide to AI-generated invoice fraud shows why evidence outside the ledger matters.
How far along is adoption?
Survey expectations are strong and specific: 14% of finance executives expect more than 75% of procure-to-pay activity to run agentically within three years, another 24% expect 50–75%, and 35% expect 25–50% (SSON R&A, Q1 2026). These figures measure expectations, not current deployments. Production agents today tend to concentrate in narrow, evidence-heavy processes, and the gap between survey optimism and running systems is where diligence belongs.
The motivation is concentrated too: 60% of finance transformation is driven by efficiency and cost reduction (SSON R&A, Q1 2026). That pressure can make autonomy sound attractive when finance actually needs broader coverage with clear control.
What stands in the way?
Legacy technology and systems complexity lead by a wide margin: 43% of executives name them the biggest barrier to finance transformation, ahead of organizational resistance at 26%, unclear ROI at 17%, and talent gaps at 8% (SSON R&A, Q1 2026). The barrier is integration-shaped rather than model-shaped. That is why read-only entry points can move first: an audit of exports needs no production integration and can return findings while a longer IT roadmap is still being decided. The ERP remains the system of record.
Does process ownership solve it?
Ownership helps, but it is not sufficient. Eighty-nine percent of GBS executives say a Global Process Owner has a positive influence on performance, and 77% already use one for procure-to-pay (SSON R&A, 2026). Ownership without enabling automation does not resolve leakage, while automation without ownership creates a system nobody answers for. The two work together when a named owner can inspect every action.
What has to stay human?
The decisions. Agents investigate, with every case carrying its source data, actions, reasoning, and evidence. Approval gates hold critical actions for a person or, once trust has been earned, a reviewing agent configured by the finance team. Machines investigate. Finance decides.
This is the architecture behind Controlling & Close. The low-risk way to evaluate it on your own data is a read-only lookback audit across 12–24 months of exports.