The question of which work to give to an agent is, in practice, the most consequential decision in an agentic program — more consequential than the choice of model, more consequential than the choice of orchestration framework, and far more consequential than the choice of vector database. A poorly chosen use case — one where the agent's capabilities don't match the task's requirements, where the risk of autonomous error is too high for the value generated, or where the organization's data and integration readiness is inadequate for what the task demands — will fail visibly and expensively, damaging not just the specific deployment but the organization's appetite for the broader program. A well-chosen portfolio of use cases, by contrast, generates evidence of value, builds organizational capability, and creates the institutional confidence that sustainable agentic programs are built on.
The Portfolio Mindset
Use case selection for agentic AI is a portfolio management problem, not an optimization problem. The goal is not to find the single highest-value use case and execute it perfectly; the goal is to assemble a portfolio of use cases that collectively demonstrates value across multiple dimensions — quick wins that prove the concept, workhorses that generate steady business value, and longer-horizon bets that position the organization for future capability. A portfolio mindset also recognizes that use cases interact: the data prepared for one agent can often power another; the integration work done for one deployment reduces the cost of the next; and the organizational learning generated by one successful agent creates the human capital that enables the next.
The portfolio should be actively managed, not accumulated. A use case that made sense eighteen months ago may no longer make sense today — because the cost structure of agentic AI has changed, because the organization's data readiness has improved and better use cases are now accessible, or because the original deployment has revealed that the task is more complex than it appeared and the agent's performance is not improving toward the required threshold. Regular portfolio reviews — quarterly for a mature program — should assess each use case against its original business case, flag underperformers for remediation or retirement, and identify new candidates that have become viable as the program's capabilities have expanded.
Taxonomy of Agentic Use Cases
Use cases for enterprise agents cluster into a taxonomy that has emerged from the collective experience of early adopters across industries. Research and synthesis tasks — gathering information from multiple sources, summarizing it, and producing a structured output — were among the first enterprise use cases to prove out, because they are relatively forgiving of imperfect accuracy, the output is reviewed by a human before it is acted upon, and the value proposition is clear: the agent can do in minutes what would take a human analyst hours. Process execution tasks — following a defined workflow to complete a business process, such as processing an invoice, onboarding a vendor, or generating a compliance report — are higher-stakes but also higher-value, because they displace labor at scale rather than augmenting it.
Monitoring and alerting tasks — continuously watching a data source, identifying events that meet defined criteria, and triggering a response — are a natural fit for agentic AI because the task is repetitive, time-sensitive, and well-defined. Decision support tasks — preparing the information and analysis that a human decision-maker needs to make a decision, without making the decision itself — represent the boundary between augmentation and automation: the agent does the legwork, the human retains the judgment. And interaction management tasks — handling conversations with customers, employees, or external partners on behalf of the organization — are the most visible and the most sensitive, because errors in these interactions have immediate reputational consequences.
What Makes a Use Case Agent-Appropriate
Not every task benefits from agentic automation. The characteristics that make a task well-suited for an agent can be summarized in five criteria. First, multi-step structure: the task requires a sequence of decisions and actions, not a single query-response pair. Second, tool access requirement: the task requires accessing external data sources, systems, or services that cannot be incorporated into a single context window. Third, sufficient error tolerance: the consequences of an agent error are bounded and recoverable — either because the agent's outputs are reviewed before they are acted upon, or because the actions themselves are reversible. Fourth, clear success criteria: it is possible to define and measure what a good outcome looks like, which is necessary for both operational monitoring and governance. Fifth, data and integration readiness: the data the agent needs is accessible, clean, and classified appropriately for the agent's permission level.
A task that fails one of these criteria is not necessarily unsuitable for agentic automation — but the failure point should be addressed explicitly before deployment, either by engineering a solution (adding human review to address an error tolerance problem) or by accepting the residual risk in the governance charter. A task that fails multiple criteria is a strong candidate for deferral until the underlying readiness gaps have been closed.
"The question is not whether agents can do the task. With enough prompt engineering, they can do almost anything. The question is whether they should — and whether the organization is ready for them to do it unsupervised at scale."
Building and Governing the Portfolio
A governed use case portfolio begins with a formal intake process: a lightweight template that captures the task description, the proposed agent architecture, the expected value, the risk assessment, the data and integration readiness status, and the success criteria. The intake process is not a bureaucratic gate — it should take no more than a day to complete and should be designed to help teams think through the questions they need to answer before deployment, not to slow them down. The governance council reviews intake submissions and makes one of three decisions: approve for pilot, defer pending readiness improvement, or decline.
Approved pilots enter a structured governance regime: a named business owner, a technical owner, a data owner, a defined pilot scope and duration, a set of evaluation metrics that will determine whether the pilot advances to production, and a human oversight arrangement that is commensurate with the pilot's risk level. Pilot outcomes feed back into the portfolio management process: pilots that succeed provide evidence for scaling; pilots that fail — and some will — provide the learning that improves the selection and design of subsequent use cases.
Companion: for a worked example of a twenty-five-case portfolio across nine industries — every entry sourced — see The Case Atlas. The atlas is sortable by industry, deployment phase, and ROI tier, and three rows link out to deeper narrative treatments in Report №03.