How much autonomy to give AI: recommend, approve or act

When it comes up

When a finance AI moves from drafting to doing: posting entries, applying cash, sending reminders, preparing or releasing payments, or making decisions that affect customers, suppliers or employees.

Your options

Recommend only

Fits when: Consequences of error are high or hard to reverse, accuracy on your data is unproven, or the decision is strategic (targets, guidance, capital).

Watch out: People may rubber-stamp recommendations; savings are smaller because people still execute.

Act with human approval

Fits when: Volume is moderate, actions are reversible with effort, and you want speed with a control point.

Watch out: Approval fatigue; approvers need context and time, or the control becomes a formality.

Act alone within limits, with monitoring

Fits when: High volume, low value per item, reversible actions, measured accuracy on your data, and strong monitoring and sampling.

Watch out: Errors repeat at scale; needs limits, logs, sampling, a kill switch and change control on prompts and models.

How to decide

Rate the consequence of a wrong action: amount, reversibility, who is affected, regulatory exposure.
Measure accuracy and overrides on your own data at the lower level first.
Set limits like a delegation-of-authority matrix: amount, type, counterparty.
Define evidence: logs, approvals, sampling results.
Promote one level at a time when evidence supports it; demote after incidents.
Keep strategic decisions at recommend: targets, guidance, capital allocation, adverse decisions about people.

Trade-offs

More autonomy brings more savings and speed but larger, faster errors; less autonomy keeps control but limits value and can create rubber-stamp approvals that look like control without being one.

Real examples

Esker's 2026 Global Finance AI Trust Index found 32% of finance leaders use AI to recommend actions that teams execute manually and 29% allow AI to act with human approval; 78% of CFOs (vs 11% of other finance leaders) say AI should not execute decisions around setting revenue targets. Source: Esker — New research finds finance leaders want more proof and control before giving AI greater decision-making authority (Sep 24, 2026) ↗
In Moffatt v. Air Canada (2024), a tribunal held the airline liable for its chatbot's wrong statement about bereavement fares and rejected the argument that the chatbot was a separate legal entity responsible for its own actions. Source: Civil Resolution Tribunal (British Columbia) — Moffatt v. Air Canada, 2024 BCCRT 149 ↗ · McCarthy Tétrault — Moffatt v. Air Canada: a misrepresentation by an AI chatbot ↗
Australia's Robodebt scheme raised debts relying heavily on averaged income data, persisted despite criticism until November 2019, and in 2020 debts raised through averaging were reduced to zero after a class action settlement; the Royal Commission recommended more review and oversight of automated decision making. Source: APRA — Royal Commission into the Robodebt Scheme: overview of the final report (2024) ↗

What to check first

The consequence and reversibility of a wrong action in this process.
Measured accuracy and override rates on your own data at the current level.
Whether logs, limits and a kill switch exist and have been tested.
Whether the decision affects people adversely or is strategic.

Learn and use

Sources

Last reviewed 2026-10-01.