Apply Now!
← Back to Blog

ChatGPT Work Changed the MVP: Pilot the Workflow Before You Build the App

Dhaval Bhatt
A domain-expert founder shaping business documents and workflow steps into a glowing AI product prototype

The first version of your AI product may not need to be an app.

It may need to be a workflow you can run for a real customer this week.

On July 9, OpenAI introduced ChatGPT Work, an agent that can work across connected apps and files, stay with a project for hours, create materials such as spreadsheets, presentations, documents, and web apps, and pause for human input or approval. OpenAI also put Sites in public beta, letting people turn work into an interactive site or lightweight web app that can be shared by URL.

That changes the MVP question.

Instead of asking, “Can I build the software?” ask, “Can I deliver the outcome repeatedly, with AI doing most of the execution and my expertise controlling the important decisions?”

The workflow can come before the product

A traditional MVP requires you to decide the interface, features, data model, and automation before you have much evidence about how customers will actually use it.

That is backwards for many AI products.

A better first step is to run the job manually with an agent. Use the customer’s real inputs. Follow the real process. Deliver the real output. Watch where the agent gets confused, where the customer changes direction, and where your judgment becomes essential.

Imagine you spent 15 years in commercial insurance. You know how to review renewal documents, identify coverage gaps, prepare questions, and organize the next conversation. Your first MVP does not need to be a complete insurance platform.

It can be a controlled workflow:

  • collect the customer’s policy documents;
  • extract the important terms and changes;
  • compare them against your review checklist;
  • flag missing information and uncertain cases;
  • prepare a structured brief for the customer;
  • pause before any recommendation that needs expert review.

The customer experiences a useful outcome. You learn what the product must do. You have not spent months guessing at features.

Horizontal AI makes your domain knowledge more valuable

OpenAI says more than five million people use Codex each week, including more than one million people using it for work outside software development. In June, the company also introduced role-specific plugins and previewed Sites for creating dashboards, planners, review workspaces, project boards, and other lightweight tools.

The direction is clear: general AI tools are moving deeper into everyday work.

That does not eliminate the opportunity for a focused founder. It changes where the opportunity lives.

A horizontal tool can create a dashboard. It does not automatically know:

  • which inputs are reliable in your industry;
  • which exception creates real risk;
  • which step requires approval;
  • what a good final result looks like;
  • how the buyer wants the work documented;
  • what must be remembered for the next case.

That operating knowledge is the product blueprint.

Your advantage is not access to a model. Everyone can rent the model. Your advantage is knowing the sequence, standards, exceptions, and decisions that turn a plausible output into work someone can trust.

Run a seven-day workflow pilot

Pick one job you understand well and can finish in a week. Not “AI for logistics.” Not “an agent for healthcare.” One narrow result for one specific user.

Then run this pilot:

  1. Define the finished output. Name the artifact or decision the customer receives: a risk brief, account plan, compliance review, candidate shortlist, maintenance schedule, or forecasting memo.
  2. Collect real inputs. Use the documents, messages, spreadsheets, and forms the customer already has. Synthetic examples hide the mess your product must handle.
  3. Write the workflow as steps. Capture how you actually do the job, including the checks you perform automatically in your head.
  4. Mark the judgment points. Identify where the AI should stop and ask you or the customer for a decision.
  5. Deliver the result. Do not call it a platform. Sell or pilot the outcome.
  6. Record every correction. Each correction reveals a missing rule, input, permission, or interface requirement.

By the end of the week, you should know more than whether the model can produce a good-looking answer. You should know what the customer provides, what the system must remember, where trust breaks, and which part of the workflow creates the most value.

Those are product requirements earned from reality.

Build only after repetition appears

Do not turn every manual step into software after one run.

Run the workflow for several customers or several cases. Look for repetition:

  • The same input arrives in different formats.
  • The same three questions are always missing.
  • The same exception requires expert review.
  • The same output gets copied into another system.
  • The same customer asks for status at the same point.

Now you know what to productize.

The intake form becomes a feature because it prevents missing information. The approval screen becomes a feature because customers need control. The audit trail becomes a feature because buyers need evidence. The integration becomes a feature because copying the final result wastes time.

This is a stronger path than building a polished app around an imagined workflow. You are turning observed friction into software, one repeatable piece at a time.

Your first moat is the method

ChatGPT Work, Codex, and the next generation of AI builders will keep making prototypes easier to create. That is good news for experienced professionals who have been waiting for permission to build.

But the tool is not the business.

The business begins with a method that produces a dependable result for a specific person. Start by delivering that method with AI. Learn from real work. Then turn the repeated parts into a focused product customers can use without you in the room.

That is how domain expertise becomes an AI product: not through a giant feature list, but through one valuable workflow made visible, testable, and repeatable.

If you want help choosing the right workflow and turning it into a focused 12-week launch plan, book a strategy call. That is the work we do inside the AI Product Accelerator.

Sources: OpenAI, “ChatGPT is now a partner for your most ambitious work” (July 9, 2026); OpenAI, “Codex for every role, tool, and workflow” (June 2, 2026); OpenAI, “Introducing workspace agents in ChatGPT” (April 22, 2026).