A Roadmap for Trustworthy Enterprise Adoption
Trusting AI is peculiar. It’s not something you build with internal memos or announcements. Trust develops gradually and often in silence, built from systems exhibiting predictable behaviour over time.
AI TRiSM is about providing a formal process to develop trust, evaluate risks, and manage security. The essence of it is protecting data in addition to continually monitoring the way models of the system perform throughout their lifecycle, not just at initiation but also for the life of the model.
A place to start practically is with the data lineage. If a model creates a wrong recommendation, can you trace it back to the data that created the model? A lot of organisations can’t do it easily—therefore, there is risk involved.
Check out: Human-Centered AI Integration
Another example is the drift of a model. A pricing model, for instance, has worked well for 6 months, but the pricing model may not work effectively because the economy has changed. Monitoring those errors is essential; if there isn’t any monitoring, then the errors will compound over time and, by the time they are noticed, the decisions have already been affected by the errors.
Implementing AI TRiSM is not entirely a technical implementation. There will be a requirement of business working together with legal and data science teams to have a common understanding of what “acceptable risk” means, which is going to create some discomfort in having that conversation; however, it needs to be had.
Having transparency is also part of this process. There is not always a need for full explainability of every decision but possibly critical decisions in heavily regulated environments.
In summary, AI TRiSM is more about resiliency than perfectionism. Systems will fail and models will drift and the objective is to identify the problems early and act intelligently; respond to the problem in real time rather than react after the problem has occurred.
