Dots, automation, or a custom agent: which fits the job?
Choose the operating model before the tool. Predictability, judgment, integration, and ownership point to different implementation paths.
Describe the job without naming a product
Write down what starts the work, which information it uses, what it must produce, and what can go wrong. Include the required timing, the systems it must change, and who accepts the result. This description is more useful for architecture selection than a request to “install agents.” Several different approaches may satisfy the same objective, with very different maintenance and governance requirements.
Consider an illustrative invoice intake process. Extracting a supplier name from varied documents may benefit from model interpretation. Checking that a required identifier exists is a deterministic validation. Approving payment is a business decision with authorization requirements. A useful design can combine these steps instead of assigning every responsibility to one open-ended agent. Separate the parts that need judgment from those that should behave predictably.
Use a dot for a supervised personal responsibility
A dot fits work organized around an individual who can supply context, inspect progress, and make decisions. Research preparation, internal draft updates, and coordination around changing information are useful candidates for evaluation. The product provides its own cloud computer and can use supported connections. That can reduce initial setup compared with building a complete application, provided the required accounts and capabilities are available.
The tradeoff is that the workflow lives within the product's account, workspace, permissions, and operating model. Do not assume an undocumented fleet-management interface or the ability to package an individual's dot as a custom multi-customer application. Confirm the actual product boundary. A personal responsibility with occasional review is a different requirement from a centrally operated service with defined tenancy, audit coverage, and recovery behavior.
Use deterministic automation for stable rules
If the trigger, inputs, decision rule, and output are well defined, ordinary automation may be the clearest foundation. Examples include validating a form, routing an approved record, or copying a known field between systems. The key engineering work is handling duplicates, retries, source changes, and partial failure. Introducing a language model does not remove those responsibilities and may add variability where none is needed.
A workflow can still use AI for one bounded step. For example, a model may propose a document category while deterministic code validates allowed categories and routes uncertain cases to a reviewer. Keep authoritative writes behind explicit checks. Evaluate whether the variable step improves the process enough to justify its operating cost and review burden. A smaller implementation can be easier for a business to understand and maintain.
Build a custom agent when the product is the workflow
A custom agent becomes relevant when you need your own user experience, controlled tools, application-specific state, or integration behavior beyond a personal assistant. OpenAI's Agents API provides a managed harness for application developers, while the Agents SDK supports building agent applications. These are development building blocks with their own contracts and constraints. They are distinct from configuring a user's dot.
A custom application makes the delivery team responsible for requirements such as identity, permissions, observability, failure recovery, evaluation, and ongoing operation. Review current API availability and data-processing constraints before selection. The September 29 Agents API documentation, for example, describes US-only data residency and no Zero Data Retention support, including with a self-hosted sandbox. A different execution location does not automatically change every service-level policy.
Compare total operating responsibility
Evaluate more than build effort. Ask who will notice a failed run, update an integration, review uncertain outputs, and respond when the source data changes. Estimate the workload using actual pilot observations rather than a promised percentage reduction. A rapid first demonstration can conceal substantial recurring work if every output requires correction or the system has no visible failure state.
Write a decision record covering fit, constraints, ownership, and the simplest rejected alternative. For each candidate, identify one requirement that could rule it out. A cloud-accessible source requirement may affect a dot workflow. Strictly deterministic outcomes may favor conventional automation. A customer-facing interface or specialized permission model may justify custom development. Record the evidence needed to settle uncertain requirements before purchasing or committing to a build.
Prove the selected approach on a bounded slice
Run the same representative cases through the proposed workflow. Include a normal input, a missing field, a conflicting source, and an unauthorized action request. Compare the accepted output and human effort. The purpose is to discover whether the operating model fits your team, not to win a comparison through a carefully selected demonstration.
A reasonable starting recommendation is the least complex approach that meets the verified requirement. You can add an agent where interpretation creates value and keep stable controls in deterministic systems. Revisit the decision when the business scope changes. The right architecture for one employee preparing a brief may be different from the architecture needed to serve several departments or external customers.
The source record
Sources & editorial notes
Primary documentation checked September 29, 2026. Our implementation recommendations are editorial analysis. Illustrative workflows are proposed examples, not completed client case studies or performance claims.
- Meet dots OpenAI
- Control your dot OpenAI
- Agents API overview OpenAI
- Agents SDK OpenAI
- Connect computers and apps to your dot OpenAI
Found something that needs updating? Contact the editorial team with the passage and a supporting source.