Six Primitives, One Shared Vocabulary for AI Work
OpenAI’s use-case guide turns a large and uneven collection of AI experiments into six recurring patterns: content creation, research, coding, data analysis, ideation and strategy, and automation. The important question is not whether these labels are the final taxonomy of AI. It is whether they give people from different departments a common way to describe the help a task needs. Used that way, the primitives make a candidate easier to discuss without pretending that it is already valuable, safe, or ready for deployment.
A taxonomy for translating work into capability
The guide says OpenAI analyzed more than 600 customer-sourced use cases and found that most fell into the six primitives. Each primitive is presented as a fundamental type that can recur across roles and industries. That makes the framework useful at the moment when a team is moving from “this task is frustrating” to “this is the kind of assistance that might change it.” It is not a score for priority, a system architecture, or a guarantee that a use case belongs to only one category.
The distinction matters because a label answers one question: what kind of work is AI being asked to help with? It does not answer other questions: how important is the work, what data can it access, who reviews the result, what happens when it fails, or should an agent control it? A content-creation use case might remain a person-directed drafting aid. An automation might be a repeatable instruction applied by an employee, not an autonomous agent. The same primitive can lead to very different designs.
Here is the compact model:
- Content creation transforms context into drafts, summaries, edits, translations, or material adapted for another audience or channel.
- Research gathers and organizes information so a person can learn, compare, investigate, or make sense of a body of material.
- Coding uses natural language and model assistance to create, debug, translate, or explain software and queries.
- Data analysis works across spreadsheets, dashboards, and other structured or semi-structured inputs to surface patterns and present them clearly.
- Ideation and strategy helps generate options, structure thinking, test a plan, or consider goals, constraints, risks, and dependencies.
- Automation makes a repeatable task more consistent by applying stable instructions, inputs, and output expectations over time.
The six are best understood as verbs describing a relationship between a worker and a task. They are not six departments. A product manager can use research, analysis, coding, and content creation in one morning. A finance team can combine data analysis with automation and a policy draft. A marketer may move from ideation to research to content production. The framework gains power when it helps people see those connections rather than when it becomes a six-column inventory.
OpenAI’s reported customer analysis is useful evidence that these patterns recur, but its sampling method and the exact distribution are not supplied in the guide. “Most” also leaves room for use cases that do not fit neatly. The taxonomy should therefore be treated as a practical organizing device, not as an independently established map of all organizational AI work.
How the primitives compose in real workflows
Content creation and research depend on context
Content creation is more than asking a model to write. The guide describes summarizing sales calls, drafting strategy documents and product requirements, creating blog posts and web pages, editing and proofreading, translating, and repurposing material for different audiences, channels, or programs. It says output can be shaped by a complete conversation, uploaded documents, a writing guide, or examples of a company’s best work. Those inputs give the model constraints about subject, tone, structure, and audience that a short request would leave implicit.
The mechanism is context transformation. A person supplies information and a desired purpose; AI turns that material into a draft or a new form. The more important the source context, the less reasonable it is to treat fluent output as evidence that the transformation is faithful. A draft can sound aligned while omitting an exception or changing the emphasis of the source. The use-case guide describes the speed and range of these applications, but it does not provide accuracy rates or a universal review procedure. A responsible workflow still gives someone authority to check meaning before publication or downstream use.
Research uses a similar context mechanism for a different purpose. OpenAI describes quick learning about a concept, searching for articles or competitive data, analyzing long internal documents, and specifying the form of the result—tables, bullets, sections, or cross-references. Output structure is not cosmetic. A comparison table can make differences visible; a cross-reference can connect a conclusion to the material that supports it; a defined set of sections can make a long document easier to inspect.
The guide also introduces deep research as an agentic ChatGPT capability that conducts multi-step internet research and synthesizes hundreds of online sources into a report. That description belongs to the product capability presented in the guide, not to every interaction with an AI system. Even when a system gathers many sources, a reader still needs to ask which sources were used, whether they are relevant, and whether the synthesis preserves disagreement. Research assistance improves the route to an answer; it does not remove the need to judge the answer’s basis.
Content creation and research often form a pair. Research supplies the facts, perspectives, or internal context; content creation turns them into a brief, report, message, or explanation. The handoff is a quality boundary. If research is weak, a polished draft can make weak evidence more persuasive. If the content format hides uncertainty, later readers may not know which parts were observed, inferred, or proposed. The primitive vocabulary helps identify both tasks, but workflow design must preserve the relationship between their inputs and outputs.
Coding and data analysis broaden access—and spread the checking obligation
The coding primitive includes debugging, generating first-draft syntax in unfamiliar languages, porting code, researching API endpoints, and using AI as a conversational partner while reasoning through a problem. The guide also says non-coders are using natural language to build Python scripts, SQL queries, and visualizations. That can shorten the distance between a business question and a first technical artifact. Someone who previously waited for a specialist may be able to express the question, inspect a result, and collaborate more effectively with an engineer.
The gain is not the same as production readiness. Generated code can use the wrong assumptions, mishandle permissions, or fail on edge cases that the initial example does not expose. A query can return a plausible table while joining the wrong records. A visualization can make a trend look important because its scale or comparison is misleading. The guide’s examples support capability broadening and faster first drafts; they do not independently verify correctness or establish that expert review is unnecessary.
Data analysis extends the same idea to spreadsheets, screenshots of dashboards, and information assembled from different sources. OpenAI describes AI helping users harmonize data, identify trends, interpret charts, and format a report according to a requested chart type or summary structure. This is valuable when the barrier is not a lack of data but the time or skill required to turn it into a usable view.
The guide reports that Tinder uses ChatGPT for first-draft syntax in specialized languages such as Bash, for API documentation, and for troubleshooting architectural and design decisions. It also describes Poshmark using AI-generated Python to reconcile millions of spreadsheet rows and then generate weekly performance reports and accounting memos. These are concrete illustrations of how the primitive can enter existing work. They are OpenAI-presented customer accounts, not independent audits of code quality, reconciliation accuracy, or the reported time savings. The safe general lesson is that AI can move analysis and coding earlier in a workflow, while the checking obligation moves with the artifact.
This is why the primitive should be connected to a reviewer and a consequence. For a low-stakes exploratory chart, a quick human plausibility check may be enough. For a financial reconciliation or a change to a live system, the reviewer must examine inputs, assumptions, permissions, and the result’s effect on later decisions. The label “data analysis” says what the system is doing; it does not determine the acceptable error or the required control.
Ideation can become routine automation, but the boundary is real
Ideation and strategy occupy the most visibly open-ended part of the framework. The guide describes brainstorming campaign ideas, structuring a document, troubleshooting a strategy, reviewing a product requirements document against goals or stakeholder preferences, and using voice or vision to interact with AI more naturally. It also says teams are asking models to account for data, goals, constraints, risks, and dependencies when building strategic plans.
In this primitive, the output is often a set of options or a clearer problem statement rather than a final decision. That makes the human role different from proofreading a draft. The person needs to assess whether the proposed direction fits the organization’s goals and constraints, whether relevant evidence is missing, and which trade-offs are acceptable. An idea generator can widen the search space, but it can also create noise or make a weak premise feel sophisticated. The useful test is whether the interaction improves the quality of the next decision, not whether it produces a large number of suggestions.
Automation begins when a useful pattern is repeated. OpenAI gives examples such as weekly competitive updates, executive reports, launch summaries, customer-insight sharing, and localization. It attributes repeatability to memory, custom instructions, and Custom GPTs that apply the same document, directions, and output expectations each time. This makes a routine more consistent and reduces the need to restate its operating context.
Repeatability is not the same as independence. A reusable routine may still be initiated by a person, checked by a person, and limited to a low-risk output. It becomes more agent-like only when a system manages workflow execution, selects tools, and takes steps on the user’s behalf. The use-case guide points toward deep research and Operator as examples of a future in which AI can perform multi-step work independently and on a schedule, but it does not define a reliability threshold or say that every Custom GPT is an agent.
The difference between ideation and automation can be seen in the trigger and the acceptance test. Ideation begins with an uncertain question and succeeds when it helps a person clarify options. Automation begins with a sufficiently stable routine and succeeds when it produces the expected result consistently enough to review efficiently. If the inputs, policy, or desired output change often, automation may simply repeat uncertainty at speed. If the routine is stable but its action has serious consequences, repeatability may make a human approval gate more important, not less.
Overlap is the point, not a classification failure
Consider a launch-planning workflow. Research might gather market and customer information. Data analysis might estimate the size of an opportunity. Ideation and strategy might turn those inputs into a brief. Content creation might produce messages for different audiences. Automation might prepare recurring updates or translations. Coding could support an interactive visualization that lets other stakeholders inspect the analysis. None of these steps is “the AI use case” by itself. The value and risk lie in the sequence, the handoffs, and the decisions that each output enables.
That sequence is why forcing one label onto the whole effort is misleading. A workflow can use several primitives while assigning different levels of human control to each. A person might approve strategic options, ask AI to draft copy, and allow a standard report to be generated automatically. The same organization might keep a data reconciliation under specialist review even if the preceding summary is self-service.
The relationships in the diagram add the missing operational questions. What context does a task receive? What output does it produce? Which later task consumes that output? Who checks it, and what permission or consequence surrounds the handoff? A primitive becomes useful when it is attached to those relationships. Without them, a team can accurately classify a task and still misunderstand the workflow.
A practical use of the vocabulary is therefore to annotate, not to conclude. For each task, name the dominant primitive, note any secondary primitive, describe the input and expected output, and identify the reviewer. Then ask whether the task is exploratory, repeatable, or action-producing. This preserves the speed of a shared language while keeping value, risk, and architecture as separate judgments.
The six primitives help people see where AI may fit. They do not tell an organization what to scale. That next decision requires comparing impact and effort, including the cost of review and the value of the workflow that follows. A richer vocabulary expands the set of possibilities; disciplined selection decides which possibilities deserve to become work.