Use ChatGPT to Choose Which Business Process to Improve First

When a team says it wants to use AI everywhere, the sensible next question is usually less exciting: which process should we examine first?

This documentation-based example uses ChatGPT to compare three fictional processes. The output is a prioritisation table and a decision record. It is not a claim that ChatGPT can discover the best automation target on its own.

Evidence label: Documentation-based and illustrative. The workflow follows OpenAI’s current prompting guidance. The process data and sample scores are fictional. No model run, time saving, or business result is claimed.

This is also the office-friendly substitute for a builder lesson in this batch. The current main already contains several Codex, MCP, and Claude Code integration guides. A new integration variant would repeat an existing task or output, so this piece teaches a distinct business decision instead.

What you need

ChatGPT on the web or an approved workplace surface
a short list of recurring processes
someone who can check the scores and own the next experiment
permission to share the process description with the service

Start with processes where a person can review the result and reverse the experiment. Keep regulated, safety-critical, and high-impact decisions out of this practice exercise.

Sample input

Paste this fictional table into a new chat:

Process | Frequency | Current pain | Data sensitivity | Human review possible | Reversible pilot possible
Weekly sales report | Weekly | Manual copy and unclear missing-data notes | Internal | Yes | Yes
Customer refund approval | Daily | Slow handoff between support and finance | Personal and financial | Yes, but authority matters | Limited
Supplier comparison | Monthly | Terms are spread across three proposals | Confidential commercial | Yes | Yes

The table is deliberately incomplete. It does not include hours saved, error rates, tool prices, or a guaranteed return. ChatGPT should flag those gaps instead of filling them.

Steps

1. Define the decision before asking for a score

The decision is: Which one process should receive a small, reviewable AI experiment next?

The decision is not: Which process can we automate completely?

That distinction keeps the first experiment bounded.

2. Paste the process table and this prompt

You are helping a manager choose one bounded AI workflow experiment from the process table below.

Return:
1. A comparison table with the original process name, a 1 to 5 score for pain, repetition, reviewability, and reversibility, plus a short reason for each score.
2. A list of missing facts that prevent a confident choice.
3. One recommended pilot, but only as a conditional recommendation.
4. A pilot brief with: input, output, human reviewer, approval stop, data boundary, and stop condition.

Scoring rules:
- Pain: 1 means little visible friction; 5 means recurring work blocks or delays a team has already described.
- Repetition: 1 means rare; 5 means frequent and consistent.
- Reviewability: 1 means a person cannot reasonably check the output; 5 means the source and output can be compared.
- Reversibility: 1 means a mistake is hard to undo; 5 means the experiment can run without changing the source system or sending an external action.

Rules:
- Use only the facts in the table.
- Do not invent hours, savings, error rates, prices, legal conclusions, or customer impact.
- Treat personal, financial, and confidential data as a reason to narrow the pilot, not as a reason to rank a process higher.
- Show the arithmetic for any total score.
- If a score cannot be supported, write "UNKNOWN" instead of guessing.
- Do not recommend automatic approval, sending, payment, deletion, or production changes.

<process_table>
[PASTE THE APPROVED PROCESS TABLE HERE]
</process_table>

3. Turn the answer into a human decision record

Copy the proposed pilot into a short note and correct anything that does not match the input. For the sample, a reasonable conditional choice might be the weekly sales report because the output can be checked against the source and the pilot does not need to change the source system.

That is not a measured result. It is a decision based on the stated rules and fictional input.

Write the actual decision in this format:

Decision: [process selected, or no process selected]
Reason: [two source-backed reasons]
Pilot input: [approved sample or file]
Pilot output: [review table or draft artifact]
Human reviewer: [named person]
Approval stop: [what must be checked before use]
Stop condition: [what ends the pilot]
Unknowns: [facts still needed]

Checkable artifact

You should have two artifacts:

a comparison table with visible scores and reasons
a pilot brief with a named reviewer, data boundary, approval stop, and stop condition

If either is missing, the process choice is not ready.

Acceptance checks

The work passes when:

every score can be traced to the scoring rules or is marked UNKNOWN
the sensitive refund process is not treated as a safe automatic workflow
the selected pilot has a human review point
the pilot does not modify a source system or send an external action
missing facts remain visible
the decision record states who can stop the experiment

Troubleshooting

Every process receives a high score. Ask ChatGPT to display the score evidence in a separate column and mark unsupported scores UNKNOWN. A table with more numbers is not necessarily more useful.

The recommendation sounds like an ROI claim. Remove words such as saves, efficient, or high return unless you have evidence. Replace them with the actual process condition, such as repeated weekly task.

The prompt ignores data sensitivity. Put the data boundary in the pilot brief and use fictional or approved samples first. Do not paste customer or financial records into a practice chat.

Limits and privacy

This method helps a team select a small experiment. It does not prove demand, calculate a business case, or decide whether an AI tool is safe for a sensitive process. Confirm access, retention, and workplace policy before using real data.

Avoid automating approvals, payments, hiring decisions, customer eligibility, or other consequential actions from this exercise. Keep the first run read-only or draft-only where possible.

Next step

After choosing a process, write its input and acceptance checks before opening the tool. You can also review How to Evaluate AI Tools for Real Work before comparing products.

Want one practical AI-for-work lesson each week? Join The Weekly Verdict.