AAISM lesson 3: AI assets and data life cycle

AI asset and data life cycle management means knowing every AI asset the organisation owns or depends on, classifying it, assigning an owner, protecting it while it is in use and retiring it properly. Subtopic 1C and the supporting task on identifying, inventorying and classifying data and assets give it a clear position: nothing else in the programme works until the inventory exists, because policy, risk assessment and monitoring all need something to point at.

What counts as an AI asset

Classic inventories list servers, applications and data stores. An AI inventory adds items that are easy to overlook.

To these add the supporting items: embeddings and vector stores, evaluation results, and the documentation of data sources and limitations. Shadow AI, meaning tools adopted by business units without review, belongs in the inventory too. Discovery therefore needs more than a survey: network and cloud logs, expense data and browser or application telemetry reveal use that nobody reports.

The life cycle

The figure shows five stages, and each has a management question.

1. Identify and inventory. Record what the asset is, where it runs, what data it touches and who depends on it.
2. Classify and assign an owner. Apply the organisation’s classification scheme, described in the data classification topic. An AI asset generally inherits the highest classification of the data it was trained or grounded on, and a named owner is accountable for it.
3. Protect and monitor. Apply controls proportionate to the classification: access, encryption, integrity checks, logging and performance monitoring.
4. Review. Revisit classification, ownership and fitness at set intervals and on triggers such as a new data source, a model update or a change of purpose.
5. Retire and dispose. Withdraw the asset, revoke access, handle the data under retention rules and keep the evidence required.

Underneath sits the data life cycle: collection, preparation, use for training or retrieval, storage, sharing, retention and disposal. Provenance (where the data came from and under what permission) and integrity (evidence it has not been altered) must be recorded from the start, because they cannot be reconstructed later.

Ownership and the retirement problem

Each asset needs one accountable owner, usually a business role, plus a custodian who operates it. Data owners decide what data may be used for AI. Without that decision, data collected for one purpose drifts into another.

Retirement is where AI differs most. A model that has already learned from a record keeps that influence after the record leaves the dataset. Retiring a model therefore raises questions that a file deletion does not: Which datasets, embeddings and logs derived from it must also go? Do retention or deletion obligations extend to the model? Is retraining needed? Practical answers include documenting lineage so that affected models can be identified, setting retention for training data deliberately, and deciding at the outset which models would need retraining if data had to be removed.

A worked example

A 400-person insurer runs a discovery exercise. The security manager expects a dozen systems and finds forty: two fine-tuned models, a vendor claims summariser, several retrieval assistants built by analysts on shared drives, and public tools used with personal accounts. Each entry records the owner, data touched, vendor and purpose. A retrieval assistant indexing HR folders is reclassified as holding restricted data, which triggers access limits and logging. One abandoned prototype still has a live endpoint and read access to claims data. It goes through retirement: access revoked, endpoint closed, index deleted, and the decision recorded.

Easy to confuse

• Inventory versus register of risks. The inventory lists assets; risk assessment uses it later.
• Data owner versus system custodian. The owner decides on use; the custodian operates.
• Deleting data versus removing what a model learned. Only the first is simple.

Key points

• Inventory comes first; discovery must look for shadow AI.
• Assets include models, datasets, prompts, agents and endpoints, plus embeddings and documentation.
• Classification follows the most sensitive data involved, and every asset has a named owner.
• Review is triggered by change as well as by the calendar.
• Retirement must address derived data and models, not just the source files.

Check yourself

This lesson covers subtopic 1C of Domain 1 (31%); the governance and programme management recap summarises it for revision.