Choose the smallest pattern that can deliver the required outcome reliably. Use a prompt for individual, judgment-led work; a workflow for stable rules and recurring events; an application for structured interaction and records; and an agent when the system must pursue a delegated goal across changing context and tools. An agent is not the default destination of an AI use case. It is one operating pattern with additional uncertainty, cost and ownership.
Start with the work, not the interface
A conversational interface can hide very different architectures. A chat can run one prompt, start a deterministic flow, update an application or delegate a goal to an agent. Selecting the pattern from the interface therefore produces poor decisions.
Describe the work first. What starts it? Which inputs are structured or ambiguous? Which decisions require interpretation? What may the system change? Which records must persist? What happens when information is missing or a step fails?
The answer often reveals that only one part of the process needs AI. The rest may be safer and cheaper as conventional software or automation.
Four patterns solve different operating problems
Prompt: assist a person with one bounded transformation
A prompt is appropriate when a person remains responsible for the task and can judge the output before using it. Summarizing a document, producing alternative wording, extracting themes or drafting a meeting agenda can fit this pattern.
The prompt may include approved context and a requested structure. It should not quietly become an operational process with shared dependencies and no owner. When many people rely on the same prompt, it may deserve packaging as a governed skill or application.
Workflow: execute known steps and rules
A workflow is appropriate when triggers, sequence, decisions and outputs are sufficiently stable to define. It is strong where precision and repeatability matter: route a request, validate required fields, create a record, send a notification or request approval.
AI can exist inside a workflow. A prompt may classify unstructured text or produce a structured summary, while deterministic steps validate and act. The presence of a model does not turn the entire workflow into an agent.
Application: support structured interaction and persistent work
An application is appropriate when users need screens, roles, records, state, validation and a repeatable multi-step experience. Applications make the work visible and controllable. They are often the right home for case management, data correction, approvals and exception handling.
Generative features can assist inside the application, but the data model and process remain explicit. This is valuable when users need to see, edit and understand the state rather than delegate the whole objective.
Agent: pursue a delegated goal under changing conditions
An agent is justified when the path cannot be fully known in advance. It may need to interpret intent, choose tools, sequence subtasks, respond to intermediate results and ask for help or approval.
That flexibility creates an operating burden. The organization must define authority, evaluate variable behavior, monitor tool use, manage cost and own the service. If the task follows a stable path, agentic orchestration may add uncertainty without adding value.
Microsoft’s agent architecture guidance distinguishes dynamic, hybrid and deterministic orchestration, while its technology-planning guidance recommends checking existing software and non-agent code before building a custom agent. That supports a least-complexity decision, not an agent-first decision.
Use the uncertainty-authority matrix
Two dimensions expose the central trade-off.
Uncertainty asks how much interpretation is required at runtime. Does the system face ambiguous language, variable inputs or a path that depends on intermediate findings?
Authority asks what the system may change. Does it only suggest, or may it send, approve, modify, purchase or delete?
| Low authority | High authority | |
|---|---|---|
| Low uncertainty | Prompt assistance or simple workflow | Deterministic workflow or application with strong validation |
| High uncertainty | Prompt, retrieval assistant or bounded agent | Hybrid design with explicit controls, approvals and limited agent authority |
High uncertainty plus high authority is not automatically the best agent opportunity. It is the highest control burden. Split the work so interpretation occurs before a deterministic validation or human decision whenever possible.
A worked pattern decision
Consider incoming customer emails that request a delivery-date change. The message is unstructured, but the permitted change rules are precise.
A pure prompt can draft a response but cannot reliably update the order. A fully agentic solution could interpret the email, inspect the order and decide what to do, but it would place high authority behind probabilistic interpretation.
A hybrid is stronger. AI extracts the order number, requested date and reason into a structured proposal. Deterministic checks confirm customer identity, order status, inventory constraints and contractual cutoffs. An application shows exceptions to a service employee. A workflow performs the approved update and sends a templated confirmation.
The architecture uses AI where ambiguity exists and deterministic components where correctness matters. No single pattern needs to own the entire process.
Ask five questions before choosing
- Can a person complete the task with one bounded AI output? Start with a prompt.
- Can the path and rules be defined in advance? Prefer a workflow.
- Do users need persistent state, records and structured interaction? Use an application.
- Must the system choose steps or tools from changing context? Consider an agent.
- Which consequential steps can be separated from interpretation? Build a hybrid with deterministic and human controls.
The questions should be answered with a real task example, not an abstract capability description.
Failure modes of choosing too much or too little
Over-engineering appears as an agent with many tools that follows nearly the same path every time. It costs more to evaluate and operate than the deterministic process it replaced.
Under-engineering appears as a shared prompt used for a business-critical process. Users copy data manually, apply inconsistent judgment and create no persistent record. The prompt worked for one person but never became a controlled service.
Another failure is forcing the whole process into one platform component. Good solutions often compose an application, workflow, model call and human decision. Clear boundaries matter more than architectural purity.
Complexity must earn its place
Evaluate options using reliability, operating cost, change frequency, user experience, auditability, support skills and failure consequence. Include the cost of evaluation and ownership, not only build speed.
The best design is not the one with the most intelligence. It is the one that places interpretation, rules, records and authority in the components best suited to them.
Amplified Pi uses the smallest-reliable-pattern rule before engineering begins. We help teams define the decision, compare architectures and deliver the composition that turns AI opportunity into an operable result. A useful next step is to map one process into input, interpretation, decision, action and record, then assign each stage deliberately.