Is your workflow ready to become a ChatGPT app?

Look for friction, not novelty

Start with repeated moments where people lose time translating between unstructured requests and structured systems. A colleague describes a need in natural language, someone interprets it, gathers missing values and enters the result somewhere else. That translation layer is often a strong candidate.

Frequency alone is not enough. The task should have a meaningful outcome and enough consistency to describe what good looks like. A rare but expensive decision may be more valuable than a daily task that already takes ten seconds.

• Repeated interpretation of free-form requests.
• Information spread across several sources.
• Calculations or rules applied manually.
• A reviewable output or next action.

Score the workflow on five dimensions

We evaluate candidate workflows across value, repeatability, data readiness, risk and interaction advantage. Value asks what improves if the task works better. Repeatability asks whether cases share a recognizable structure. Data readiness checks whether authoritative inputs are accessible.

Risk covers the cost of a wrong result and the controls available. Interaction advantage asks whether conversation genuinely helps: perhaps users arrive with incomplete information, need explanations or revise assumptions several times.

• Value: time, quality, revenue or risk reduction.
• Repeatability: stable steps with variable inputs.
• Data: accessible, authorized and sufficiently clean.
• Risk: detectable errors and appropriate human review.
• Interaction: ambiguity that conversation resolves well.

Avoid the two common extremes

The first extreme is a generic company assistant expected to know everything and act everywhere. Its scope makes permissions, evaluation and user expectations difficult. The second is a conversational wrapper around a single button that was already easier to click.

A good first app sits between them. It owns one meaningful outcome, may use several focused tools and provides a structured artifact the user can inspect. It is narrow enough to evaluate but important enough to matter.

• Define one primary user and outcome.
• List the minimum tools and data sources.
• Specify where human confirmation is required.
• Choose examples that represent normal and difficult cases.

Prototype the decision, not only the interface

A visual prototype can validate comprehension and flow. It cannot validate whether the system makes the right decisions. Test the underlying rules and tool contracts with representative cases at the same time.

Before production, the team should be able to explain what the app does when information is missing, when sources disagree and when a user asks for something outside scope. Readiness is not the absence of uncertainty; it is having designed how uncertainty is handled.