From AI Ideas to Workflows That Matter
OpenAI’s Identifying and scaling AI use cases and A practical guide to building agents address the same organizational question from different ends: how does a company move from noticing attractive uses for AI to operating a workflow that creates value without surrendering judgment or control? The answer is not a catalog of clever prompts or an assumption that every task should become autonomous. It is a progression from finding friction, to choosing worthwhile work, to mapping the work as a system, and only then deciding how much control an AI system should receive.
That progression matters because enthusiasm and maturity are not the same thing. The first guide is about discovering and scaling use cases. The second explains what is required when a workflow needs an agent—an AI system that can manage a sequence of steps, select tools, and act on a user’s behalf. Read together, the guides offer a way to make a more defensible judgment: which work should be assisted, which can be made repeatable, and which, if any, is suitable for bounded autonomous execution.
The problem is not a shortage of AI ideas
The pressure to act is real. The use-case guide says that 39% of U.S. adults had used AI within two years, compared with 20% internet adoption over the same period. It also attributes to BCG findings that AI leaders saw 1.5 times faster revenue growth, 1.6 times higher shareholder returns, and 1.4 times better return on invested capital than less advanced peers. At the same time, the guide cites McKinsey’s finding that 92% of companies planned to increase AI investment, while only 1% believed their investments had reached full maturity.
Those figures establish a tension, not a guaranteed business case. The BCG and McKinsey findings are reported through OpenAI’s guide rather than presented with their underlying samples or methods. The guide’s own experience—300 successful implementations, more than 4,000 adoption surveys, and over 2 million business users—is also a statement about OpenAI’s work, not a neutral census of every organization. The useful conclusion is narrower: companies have reasons to experiment, but many still lack a reliable way to turn experiments into mature value.
A large idea list does not solve that problem. A technically impressive use case can be difficult to maintain, disconnected from the work that matters, or redundant with a tool the company already trusts. Conversely, a modest assistant that removes a recurring bottleneck can spread through a department because it fits an existing routine. The central decision is therefore not “Where can we add AI?” It is “Which part of work is worth changing, and what evidence would show that the change is helping?”
OpenAI’s first guide puts leadership and employee initiative together rather than treating them as substitutes. Leadership is needed to encourage adoption and support use cases that matter beyond one person. Employees are often closest to the tedious, blocked, or ambiguous parts of their jobs, so they can identify opportunities that a central technology team would miss. The organization needs both direction from above and observations from inside the work.
A progression from opportunity to value
The use-case guide organizes that movement into three steps:
- Understand where AI adds value. Look for work that is repetitive and low-value, held up by a skill bottleneck, or difficult to start because the path is ambiguous.
- Teach foundational use cases. Give employees a shared language for common forms of assistance, such as creating content, researching, analyzing data, coding, developing ideas, and automating routines.
- Collect and prioritize candidates. Compare opportunities by their likely impact and the effort required to make them useful at the intended scale.
The sequence is important. If a team begins with a product feature, it tends to search for a problem that fits the feature. If it begins with the work, it can ask a better set of questions: What consumes attention without requiring much judgment? Where does progress wait for a specialist? Where does uncertainty prevent a person from taking the first step? The guide’s “super-assistant” framing is useful here because it describes AI as extending people’s capacity across many tasks, not as automatically replacing the person responsible for the outcome.
The six primitives make the next conversation more concrete. A marketing team might use content creation to draft a campaign, research to understand a market, data analysis to estimate an opportunity, ideation and strategy to shape a brief, coding to build a visualization, and automation to repeat a report or localize the result. These are not six separate departments or six promises of readiness. They are recurring patterns that help a team describe what kind of help a task needs. A real workflow can use several of them in sequence.
This shared language also makes discovery less dependent on specialist vocabulary. OpenAI says its analysis of more than 600 customer-sourced use cases found that most fell into these six types. That is a first-party classification, and the guide does not provide the sampling method or claim that every use case fits cleanly. Its value is practical: once employees can recognize the patterns, they can propose candidates in terms that others can compare.
The third step is selection. The guide’s Impact/Effort Framework separates high-impact, low-effort opportunities from self-service experiments, high-value projects that require substantial investment, and high-effort work with little apparent value. That distinction prevents two common mistakes. A quick win should not be burdened with the governance of a major transformation, while an ambitious project should not be rejected merely because it is harder than a personal assistant. The guide recommends leadership support for initiatives that affect whole departments and quarterly reassessment because capabilities and economics can change.
In other words, discovery creates options; prioritization creates commitment. A candidate’s position in the framework is a working judgment, not a permanent label or a financial forecast. Reported customer examples—including Tinder’s use of AI for specialized coding tasks and Indeed’s months of testing for a job-recommendation explanation—illustrate possible value, but they do not prove that the same result will appear elsewhere. Local workflow, data, existing tools, and review requirements still decide the case.
Why workflows are the bridge to agents
A task list is useful for discovery, but it can hide where value and risk actually accumulate. The guide’s department workflow example follows a marketing effort from researching market trends, to analyzing customer data, to brainstorming a launch strategy, to generating messages and campaign assets, and finally to localization and channel optimization. Each step produces context for another step. Improving one step may help, but the greatest opportunity—or the most serious failure—may occur at the handoff between them.
Workflow mapping makes those handoffs visible. It asks what information enters each task, what output it produces, who reviews it, what permission is needed for the next action, and what happens when the result is incomplete or wrong. It also prevents a subtle category error: a sequence of AI-assisted tasks is not automatically an agent. Several independent tools can help people complete a workflow while the person remains responsible for choosing and executing each step.
The agent guide draws the boundary more precisely. An agent uses a large language model to manage workflow execution and make decisions. It can recognize completion, correct an action when needed, halt when it cannot proceed, and return control to the user. It also has tools for retrieving context or taking actions in external systems, selecting among them according to the workflow’s current state and within defined guardrails. A simple chatbot, a single-turn LLM feature, or a sentiment classifier does not meet that definition when the model is not controlling the workflow.
That level of control is not appropriate merely because it is technically possible. OpenAI recommends looking for workflows with complex judgment and exceptions, rules that have become difficult to maintain, or heavy reliance on natural language and unstructured information. In a payment-fraud example, a rules engine checks preset conditions, while an agent can examine context and subtle patterns. But the same guide says a deterministic solution may be sufficient when those conditions are absent. Rules are often preferable when the state space is known, the action is easy to verify, and consistency matters more than flexible interpretation.
The practical boundary is therefore conditional. Start by identifying a real source of friction. Use the primitives to describe the kind of assistance involved. Prioritize the opportunity against impact, effort, scope, and existing alternatives. Map the workflow and its dependencies. Only then ask whether adaptive workflow control is worth the additional evaluation, security, and oversight burden of an agent.
This is the core relationship between OpenAI’s two guides. The first prevents architecture from outrunning the problem; the second prevents a promising workflow from being mistaken for a reliable system. The most defensible first move is not to grant an AI system more autonomy. It is to look closely at the work people already do, and to describe the friction precisely enough that the next design choice can be tested.