TRENDING

TECHNOLOGY NEWS FROM AROUND THE WORLD

→
The 2024 EU AI Act: Why a business needs an AI system inventory
Business Intelligence · 2024
Business Intelligence2024

The 2024 EU AI Act: Why a business needs an AI system inventory

A practical business example for recording an AI system’s purpose, ownership, and human oversight in the context of the 2024 EU AI Act.

Archive retrospective · Topic year 2024

Yazılımcının GünlüğüPublished: Nov 18, 20263 min read
Reading progress0%

A model name is not a system inventory

EU Regulation 2024/1689 established a risk-based framework for AI systems and uses. For a business, counting purchased models is less informative than recording which decisions are affected. A model used to draft sales copy and one used during a customer application can create different responsibilities. Record the purpose, affected users, data source, type of output and the person able to interrupt the workflow.

Trace a change in business authority

Suppose a purchasing team tries a service that extracts totals from invoices. A staff member initially verifies the result. Allowing the service to place invoices directly into a payment queue changes its business effect. Record the new authority and identify who can reverse the action. Decide how a mistaken extraction will be corrected and when the supplier is notified. A vendor name cannot explain the scope of a deployment.

Machine learning pipelineConceptual sketch · summary of the article's technical flow

Test whether human oversight works

An approval step is ineffective when the reviewer lacks evidence or an alternative. Put an incorrect value in a sample and observe how the reviewer detects it. Check that the automation can be paused, that a disputed result is not repeated and that the version can be identified later. Record the time available to review. A button alone does not establish that human oversight works in practice.

Update the record when the use changes

Identify who reviews changes to provider, model, data or purpose. Reuse the existing inventory entry and highlight what changed in access, authority and outcome. An applicable legal classification depends on the actual context and organizational role. The inventory is a practical input to that review. A software team should be able to identify the deployment, and a business team should be able to explain how its output is used.

Test model answers against evidence

Fluent output is not proof of accuracy. Define the data available to a model, its permissions, tools and failure behavior. Evaluate stale documents, conflicting sources, denied requests and missing retrieval results, not just simple examples. Keep citations and version identifiers so an output can be audited later.

Operate with cost, privacy and rollback in mind

Context length, retrieval, caching, tool calls and retries influence the true cost of an AI application. Avoid sending private data unnecessarily to external providers. Set clear logging retention and human review paths. Compare a new version against a stable test set and preserve a rollback route for incorrect automated actions.

Be the first reader ✎ Write a comment

Comments

No comments yet. Be the first to contribute.

Discussion question from the editor Who approves the transition from AI suggestions to automated actions in your team?

Alex Morgan

Reading “The 2024 EU AI Act: Why a business needs an AI system inventory” raised this question for me: data lineage and model correctness need verification.

Sample editor reply

Editor note: Product claims and measured outcomes should be assessed separately. A small pilot with defined metrics is a useful first step.

Robin Bennett

We should also discuss maintenance cost. An easy rollout does not guarantee trouble-free upgrades.

Cameron Hayes

I would start with a pilot and measurable success criteria rather than assuming a launch means proven value.

✎ Write a comment

Your first and last name will appear publicly alongside your comment.