Nearly two-thirds (65%) of the 4,470 leaders around the world surveyed for the Enterprise AI Maturity Index in partnership with Oxford Economics identified an investment return on AI. Yet only 23% gained more than 15%. C-suites worry about regulations, data security, and intellectual property.

Solving these problems will require AI governance to be clear and AI asset KPI’s to be tracked over time. Some companies are already doing so and considering AI as a strategic asset, but more work remains before AI receives the same care it deserves with other strategic assets.

There should be a focus on AI lifecycle management, keeping model relevant, sustainable and measurable against business value at all times. This can make sure AI models meet their original intent and evolve with the data, technologies, and business goals.

AI ideas

The AI asset lifecycle starts with a single idea repository for the automations, predictions, and recommendations problem space that can be addressed by AI. This is important so nothing gets lost. This also makes it easy for other teams to share and iterate on ideas.

A marketing team for instance might submit a proposal for an AI chatbot to promote products awareness, a customer service team submits a proposal for better customer service, and a data science team submits a proposal for an AI algorithm to promote products. It makes it very convenient to look at, amalgamate, and apply them all together by having them all organized in one place.

In combination with this storehouse, the ideation and innovation flow is much smoother and more efficient which eventually produces the next generation of AI assets that are actually going to have a positive impact.

Demand prioritization

After the ideas are assembled, they can be evaluated in terms of feasibility, impact and compatibility with the larger strategy. Then a first availability analysis can tell you whether there are enough signals to produce predictive or prescriptive information.

It’s the first-place strategic value is discovered and enables organisations to invest in AI programmes that matter most.

A marketing department designing a product recommendation model, for instance, could target long-term strategic objectives like loyalty and brand adoption. Practically, this framework will allow you to have more specialized marketing campaigns and personalized content that results in real-time attention and conversions that directly impact day-to-day marketing operations.

Development cycle

The process of development starts with the requirement analysis and then involves an iterative process of building models, testing and rewriting again and again. It includes teams of cross-functional data scientists, product managers and domain specialists working together to define goal, scope, and results.

Planning: Teams need to determine business and data requirements, and goals of success (accuracy, recall, savings). When you are designing and testing the model, be sure to record the system and user test results in full detail so that you can use it for model development on future AI projects.

Here also several strategic and tactical business value definitions can be defined and documented. Once model fit has been verified the model can go live in production to consume real data and create value.

Operationalization

If you’ve put an AI/ML model into production, then it is a great start, but it doesn’t always ensure the model will deliver value on a consistent basis. Operationalization tracking the model, implementing it in business processes, and always providing value to users.

It’s important to have proper monitoring practices and AI guardrails to make sure models are being used and operating as intended. Execution, Response time, and accuracy tracking allow teams to stay ahead of performance and resolve problems in real-time.

Secondly, data provenance, in which data sets are transparently traced from ingestion to AI ready structures (engineered features, embeddings, etc) is essential for evaluators, regulators, and users.

Model drift is identified and retraining can be performed at an earlier stage, while monitoring business value drift ensures that it is aligned with the strategy, and provides models updates with early intervention.

We should check prompts and output in generative AI models for offensive content, data leaks, hallucinations, errors and toxicities. These things provide monitoring in generative and Agentic AI creation an important element.

Maintenance

When the monitoring system is properly in place, AI assets need regular maintenance to work properly. AI models must change with business requirements.

Essential upkeep is keeping track of user comments in a central feedback database. Collaboration, identifying problems and creating more openness can help teams iterate on all models.

Then, AI incident management can help teams quickly manage issues, prioritize release, and avoid issues in the future.

Retirement

The final step in the lifecycle of an AI asset is retirement – when a model no longer serves its purpose or is too expensive to run. This phase needs to be a deliberate one, with the reasons why the model is being retired assessed and the lessons learned addressed.

Consider how well the model performs, as losing accuracy or relevance could be a sign it is time to retire. When they invest in models in a purposeful way, they are allocating resources to something else that delivers value.

Dedicated control of the AI asset lifecycle will enable long-term business value, and enable corresponding AI investments in the strategic as well as tactical direction. This helps companies to evolve at a faster pace, keep up with the times, and comply with new norms that governments and regulators like:

• U.S. White House AI Executive Order.
• EU AI Act
• NIST standards (National Institute of Standards and Technology).
• DARPA’s Explainable AI (XAI) mandates Defense Advanced Research Projects Agency (DARPA).

Our AI assets and workflows are organized by AI Control Tower on Now Platform. This includes Now Assist Guardian and innovation management, demand management, agile management, incident management, creator workflows, governance, risk and compliance applications.