Contributed opinion · TechRadar Pro
Five questions to test whether an AI stack is truly under your control
Varun Sharma sets out five tests for whether an enterprise truly controls its AI stack—from exact running state and critical dependencies to legal authority and safe exit.
Varun SharmaAuthor · 11 August 2026
This is contributed opinion by Varun Sharma. TechRadar Pro publication does not constitute independent validation, endorsement or assurance of NEUVIOR, PHARMORIS or any enterprise AI system.
Prove what is actually running
A model name or dashboard is not an inventory. Genuine control begins with an exact, reviewable account of the model, configuration, data path, dependencies and environment in operation.
Control the dependencies around the model
Weights matter, but so do hosting, orchestration, observability, permissions, update paths and the parties able to change them. The weakest dependency can define the real control boundary.
Preserve a credible exit
The ability to pause, switch or dispose of a system safely is part of control. If exit rights exist only in a contract but cannot be exercised operationally, the organisation remains dependent rather than sovereign.
Questions this record can answer
The facts, without the halo effect.
What are the five control tests?
Can the organisation prove what is running, control model weights and critical dependencies, control infrastructure and operations, demonstrate legal and contractual authority, and stop, switch or dispose of the system safely?
Why is vendor access not the same as control?
Access to a service can leave the organisation dependent on another party for model changes, infrastructure, contractual rights, continuity and exit. Control must be evidenced across the whole operating stack.
Does the article certify an AI stack?
No. It is a contributed framework for asking better questions, not an audit, certification, regulatory opinion or product endorsement.