An AI project register: goal, owner, and stop conditions
Let's discuss your project →A lightweight register helps teams see which AI experiments are active and who decides whether they continue. According to the NIST AI RMF 1.0, inventory and accountability sit within Govern. This sheet is a Dedaloop editorial tool, not a certification or a complete compliance framework.
One sheet for each experiment
Start with projects already underway, even small ones. For each one, write the goal, the responsible person, the users involved, the data allowed, the tools in use, the human review step, the success criterion, and the stop condition. Add the date of the next review and where the test results are stored.
Avoid phrases like “increase productivity” if they do not point to an observable task. It is more useful to specify “prepare a draft reply about internal procedures, to be checked before sending.” The owner should be able to explain who corrects errors and who can stop the experiment.
A simulated example
Consider an assistant that looks up guidance in a fictional manual for new colleagues. The pilot uses only that manual, does not access email, and does not modify accounts. A tutor checks every response before it is used.
The sheet can define the share of answers supported by the manual and the total review time as outcomes to watch. A response that invents an operational permission or exposes content outside the authorised scope can be a suspension condition to agree before testing. These are not measured results or universal thresholds; they are examples of criteria to define.
Involving the people who use the system
Collect observations from people with different roles, digital skills, and accessibility needs. Offer a simple way to report unclear answers without needing to understand how the model works. Record the issue and the decision made, and avoid turning the register into a collection of unnecessary personal data.
At review time, choose explicitly whether to continue, narrow the task, fix it, or stop it. If the model or the documents change, note that and repeat the relevant checks: a past assessment does not automatically describe the new version.
The first step can be a short meeting with the people managing the experiments already in place. The useful output is a list of owners and next decisions, not a catalogue of tools to buy. Only then does it make sense to choose which project deserves more resources.