Research into business AI
What makes an AI action safe enough to execute?
Stonefold investigates when external controls make AI-driven business operations safer, what assumptions those controls depend on, and what operational costs they introduce.
The aim is useful evidence for businesses deciding how to deploy AI agents. The gateway is an experimental instrument; its value must be tested against existing application controls and other approaches.
Independent research · open-source Python prototype · no production validation
Purpose
Test the guarantee, not just the decision
A gateway can correctly return “allow” while the resulting business action is still wrong. We want to understand the conditions under which enforcement actually protects a workflow.
Authority and approval
Does the action executed match what a person approved? What happens when the amount, recipient, or underlying record changes?
Coverage and bypasses
Which credentials and network paths must be controlled? Can another agent, integration, or human invalidate the assumptions?
State and failure
Do controls survive concurrent requests, retries, stale records, and a timeout after an action may already have succeeded?
Operational cost
How much legitimate work is delayed or refused? When is a simpler application control sufficient?
Research direction
Compare against competent alternatives
The next studies should compare a properly configured application, an established enforcement approach, and the Stonefold prototype where it differs. An unrestricted agent alone is not an adequate baseline.
Results should describe what happened in the downstream system, the policy and deployment assumptions, and the cost of blocked or delayed work. A finding that existing controls are sufficient is as useful as a finding that an additional layer helps.
These are the standards for the next research phase, not a claim that comparative studies have already been completed. Novelty and related work must be assessed for each question.
Publications
Fresh evidence, with a clear scope
New research notes will state the question, method, observed results, and limitations. Where an experiment supports a finding, the publication will include the configuration and instructions needed to reproduce it.
No studies from the current research programme have been published here yet. This site currently describes the questions and intended method; it does not present them as completed findings.
Status · November 2026
A research-only project
Stonefold is a research project. Its Python prototype is an experimental instrument for examining controls on AI-driven business actions. No independent comparative result or production safety guarantee is claimed.
The prototype can test explicit rules under stated assumptions. It cannot establish that a business policy is correct, make every model output accurate, or control actions outside its deployment boundary.
Feedback from people deploying AI into real workflows is welcome, especially corrections, counterexamples, and cases where existing controls already solve the problem.