Create a Use Case Portfolio: How an AI CoE Prioritizes What Matters Most

An AI center of excellence should not behave like an idea factory. Its job is to turn scattered enthusiasm into a disciplined portfolio of bets that are aligned to strategy, feasible to deliver, and safe to scale. That matters because AI can create very large economic value, but the opportunity is unevenly distributed: McKinsey estimates that a relatively small set of business functions accounts for most of the value from generative AI, which means selection discipline is more important than sheer volume of experiments.

That is why the best AI CoEs start with a portfolio mindset instead of a backlog mindset. A backlog treats all requests as roughly equal until delivery constraints force choices. A portfolio treats every proposed use case as an investment decision that must compete for funding, data, technical capacity, and risk tolerance. Google Cloud describes effective AI strategy as a combination of clear vision, focused implementation of the right use cases, and consistent tracking of results, while Microsoft’s business-envisioning guidance similarly frames use-case selection around business objectives, measurable success, stakeholders, and feasibility.

A strong portfolio process also reflects the reality that AI use cases are never just technical artifacts. The OECD’s lifecycle framing shows that AI spans planning and design, data collection and processing, model building and validation, deployment, operation, and monitoring. NIST’s AI Risk Management Framework makes the same point from a governance angle: risk and trustworthiness need to be considered throughout design, development, use, and evaluation rather than bolted on at the end.

For a CoE, then, prioritization has one practical purpose: identify the few use cases that can create meaningful business outcomes in a reasonable time, using data the organization can actually trust, with risks the business is prepared to manage. The most useful scoring systems are not the most complex ones. They are the ones that force leaders to compare initiatives consistently across business value, delivery difficulty, data readiness, risk, and time to impact.

Start With Identification, Not Scoring

Before an AI CoE scores anything, it needs a clean way to define a use case. The minimum standard should include the business problem, the target user, the workflow being changed, the measurable outcome, the process owner, the required data, and the human decision points that remain in place. Google’s guidance on defining AI business use cases begins with measurable business goals and pushes teams to decide whether the need actually requires generative AI, another AI technique, or no AI at all.

That last point is more important than it sounds. Many weak portfolios are filled with “AI-shaped” ideas rather than real use cases. A proposal such as “use a chatbot for HR” is not a use case; “reduce average policy-answer handling time for HR by 30 percent while keeping escalation accuracy above target” is. Microsoft’s framework emphasizes business objective, key results, and accountability precisely because use cases need to be grounded in a decision context rather than in a technology label.

Once the CoE has a standard intake format, it should cluster ideas into a portfolio taxonomy. Typical categories include employee productivity, customer experience, revenue growth, risk and compliance, operations, and product or service innovation. This prevents the portfolio from being dominated by whichever department is loudest and makes it easier to compare near-term efficiency ideas with longer-horizon strategic opportunities. McKinsey’s work on value concentration across functions supports this approach: some domains simply have more repeatable AI value than others, so categorization improves pattern recognition.

At this stage, the CoE should reject or defer any proposal that fails three gating questions. Is the problem tied to a business metric? Is there a credible process owner and executive sponsor? Does AI materially improve the current state compared with automation, analytics, search, or workflow redesign alone? If the answer to any of these is no, the use case should not enter formal scoring yet. A smaller, sharper pipeline beats a large list of vague possibilities every time.

Score the Five Dimensions That Actually Matter

The first scoring dimension is business value, and it should carry the highest weight. This includes direct revenue upside, cost reduction, productivity gains, service-level improvement, experience gains, and strategic importance. Value should be estimated in concrete terms whenever possible: hours saved, conversion lift, reduced error rates, faster cycle times, lower attrition, or better compliance outcomes. Google explicitly recommends a business-value-driven decision approach, and McKinsey defines use cases in terms of measurable outcomes tied to specific business challenges.

The second dimension is complexity, which captures how hard the use case will be to deliver and operate. Complexity is broader than model difficulty. It includes systems integration, workflow redesign, user adoption, process exceptions, model evaluation burden, vendor dependencies, and ongoing support needs. A simple summarization assistant built on well-contained content may be low complexity; an agentic workflow that touches multiple enterprise systems and requires human review, auditability, and escalation paths will score much higher. Google’s prioritization advice around actionability and feasibility fits naturally here.

The third dimension is data readiness, and many portfolios underestimate it. A use case may look attractive on paper yet stall because the necessary data is fragmented, poorly governed, inaccessible, unlabeled, or not trusted by the business. Google’s data-foundation guidance argues that strong data foundations are fundamental to AI success, and its broader AI-readiness guidance ties scale directly to data capability and operating discipline. In practice, data readiness should assess data availability, quality, access rights, freshness, lineage, governance, and whether the needed context already exists in a usable format.

The fourth and fifth dimensions are risk and time to impact. Risk should cover legal, regulatory, privacy, security, safety, bias, explainability, reputational exposure, and operational failure modes. NIST’s AI RMF and its generative AI profile both emphasize that trustworthiness and risk management are not optional side checks but core decision criteria across the lifecycle. Time to impact matters because an AI CoE needs visible wins to earn credibility; initiatives that can produce measurable value in one or two quarters often deserve a premium over ideas that may be powerful eventually but require a year of foundation work before they prove anything.

Use a Weighted Model So the Portfolio Reflects Strategy

A practical method is to score each criterion on a 1-to-5 scale and apply explicit weights. For many organizations, a good starting model is: business value 30 percent, complexity 20 percent, data readiness 20 percent, risk 15 percent, and time to impact 15 percent. In this setup, higher scores should always mean “better for prioritization,” so complexity and risk are reverse-scored: a low-complexity use case gets a 5, while a high-complexity one gets a 1. The weighted score becomes a simple way to compare very different ideas using one shared language.

The formula can be written as: Priority Score = (0.30 × Business Value) + (0.20 × Ease of Delivery) + (0.20 × Data Readiness) + (0.15 × Managed Risk) + (0.15 × Time to Impact). That formula is intentionally transparent. The point is not statistical precision; the point is disciplined trade-off discussion. A CoE should be able to explain why a use case ranked highly without hiding behind a black-box scoring spreadsheet.

To make the model more useful, pair the weighted score with thresholds. For example, any use case below a minimum risk score or below a minimum data-readiness score cannot proceed to pilot, even if its total score is high. This avoids the common mistake of allowing attractive business cases to leapfrog foundational concerns. NIST’s guidance is especially relevant here because it treats governance, mapping, measurement, and management as coordinated activities; a high-value idea with unmanaged risk is not portfolio-ready, it is just unfinished.

After scoring, the CoE should map each use case into one of four portfolio buckets: quick wins, strategic bets, foundation first, and do not pursue now. Quick wins score high on value, readiness, and speed. Strategic bets score high on value but may need more change management or technical build-out. Foundation-first items depend on data cleanup, governance, or architecture work before they can responsibly move forward. The final bucket matters most culturally: a mature CoE is willing to say no, not because an idea lacks ambition, but because timing and readiness do not support it yet.

Turn Prioritization Into an Operating Rhythm

Scoring once is not enough. The portfolio should be reviewed on a regular cadence, often monthly for intake triage and quarterly for deeper re-ranking. Market conditions change, data foundations improve, sponsors move on, regulations evolve, and lessons from pilots reshape assumptions. NIST’s lifecycle view and OECD’s end-to-end framing both support this kind of continuous reassessment rather than one-time approval.

The review forum should be cross-functional. An AI CoE can coordinate the process, but it should not score alone. Product or business leaders own value assumptions. Data leaders judge readiness. Architecture and engineering teams assess feasibility and integration load. Risk, legal, privacy, and security functions evaluate controls and exposure. This kind of shared review reflects OECD’s emphasis on responsible AI across the value chain and Microsoft’s emphasis on accountability and executive sponsorship.

A healthy portfolio also balances horizons. If every item is a quick win, the organization may show activity but miss strategic transformation. If every item is a moonshot, the CoE will lose credibility before it proves value. A useful target is a barbell portfolio: enough near-term initiatives to generate visible returns and enough longer-term bets to reshape important workflows or customer experiences. Google’s strategy guidance on focused implementation and results tracking fits this balance well, because it links ambition to a clear roadmap rather than to scattered experimentation.

Finally, the CoE should measure portfolio performance, not just project completion. That means tracking how many use cases moved from idea to pilot, from pilot to production, how fast value appeared, how many were paused for data or risk reasons, and which assumptions proved wrong. Over time, those metrics improve the scoring model itself. The portfolio becomes smarter because the organization stops guessing what matters most and starts learning it from evidence.

What Good Prioritization Looks Like in Practice

In practice, the highest-ranked use cases are often less glamorous than the first ideas people pitch. They tend to sit where business pain is obvious, workflows are frequent, outcomes are measurable, and human oversight is already natural. Think knowledge retrieval for service teams, drafting support inside constrained workflows, case summarization, document classification, or next-best-action support for employees. These are not the only valuable AI opportunities, but they often outperform splashier concepts because they combine clear value with manageable complexity and faster time to impact.

By contrast, the weakest candidates often share a familiar pattern: unclear ownership, fuzzy metrics, poor data access, broad workflow disruption, and unresolved risk. They may still be worth pursuing later, especially if they align tightly to long-term strategy, but they should not crowd out more executable opportunities. A portfolio is not a popularity contest for visionary ideas. It is a governance mechanism for sequencing value creation.

The CoE’s real contribution, then, is not merely to collect use cases or to evangelize AI. It is to create a repeatable decision system that helps the business choose where to place its attention, money, and trust. When the scoring model is clear, the thresholds are explicit, and the review cadence is disciplined, prioritization becomes far less political and far more strategic.

That is how an AI CoE prioritizes what matters most: define the use case clearly, score it consistently, challenge it with data and risk reality, and place it in a portfolio designed to produce both momentum and durable advantage. The organizations that do this well are not the ones chasing the most AI ideas. They are the ones turning the right AI ideas into measurable outcomes, at the right time, with the right controls.

Source: Create a Use Case Portfolio: How an AI CoE Prioritizes What Matters Most