Consulting

Accountable AI Agents for Operations Leaders

Independent analysis by Starro. Not client work.

A robotic hand about to press a key on a console while a human hand rests beside a stop switch

Objective

Controls for AI Agents That Act in Business Systems

AI agents now act inside business systems: they approve, pay and change records. Most organizations deploying them lack a mature governance model, and approval thresholds alone are a familiar but weak control.

Which controls make an AI agent accountable, and how much review and audit does each one cost?

Findings

Automated Checks Give Most of the Protection

An amount threshold alone is a costly control: approving every invoice above $2,000 leaves $57k of expected loss per 10,000 invoices, against $187k under full autonomy, and takes 168 reviewer hours. Automated checks on every action cut loss to $68k for 22 hours.

Checks plus a $1,000 threshold match the loss of a $500 threshold with 24 % fewer reviewer hours. Audits are for detecting change: 620 audited actions per period detect a rise in error rate from 2 % to 4 % with 90 % power.

Audit sample per review period needed to detect a rise in agent error rate, at 90 % power and 5 % false alarm.

Approach

Autonomy Ladder Sized on a Synthetic Payables Agent

A five-level autonomy ladder built from familiar internal controls: limits of authority, approval, automated checks, logging and audits. The controls are sized on a synthetic accounts-payable agent with a 2 % error rate, simulated over one million invoices, with sensitivity to the check catch rate and to errors that grow with invoice amount.

Autonomy ladder for AI agents, from level 0, which only suggests, to level 4, which acts with audit only.

Validation

One Million Simulated Invoices and Sensitivity Runs

How we checked the analysis:

  • Expected loss and reviewer hours come from simulating one million invoices for each control set.

  • Results hold across a range of check catch rates and for errors that grow with invoice amount.

  • Audit size is set for 90 % power at a 5 % false-alarm rate.

Recommendations

Grant Agent Authority as You Would to People

Grant an agent authority the way operations already grants it to people: limits, checks on every action, approval where the stakes justify it, and audits sized to detect drift. Automated checks give most of the protection for a fraction of the review time. The controls map to OWASP, the NIST AI RMF, ISO/IEC 42001 and COSO.

Tools used:

Monte Carlo simulation of one million invoices; binomial audit sizing