Gather anonymous expert views in rounds, sharing the results each time, until the group converges.
Put a group of experts in one room and the result often reflects rank, confidence and who speaks first. The Delphi method avoids the room altogether. A facilitator sends the same questions to each expert, who answers alone and anonymously. The answers are combined into a summary and sent back to everyone, and the experts answer again, now knowing what the group thinks.
Round by round, people reconsider. Some move toward the group view; others hold firm and explain why, which is valuable too. After two to four rounds you usually have either a clear consensus or a well-understood disagreement. Because no one knows who said what, people can change their minds without losing face, and a quiet specialist counts as much as a famous name.
Use it when you need informed judgment where hard data is thin, when experts are spread across places or schedules, or when status and strong personalities would distort a face-to-face meeting.
Each round, experts answer alone, see the group's anonymous summary and reconsider. Two to four rounds is typical.
Choose a part of the diagram or a tab to explore each stage.
What exactly do you want the panel to judge, estimate or prioritize this round?
What does each expert think before seeing anyone else's view?
What does the group think, and how far apart are people?
Seeing the group's view, would anyone change their answer, and has the panel converged?
Delphi was designed to capture a group's knowledge without the usual costs of group meetings.
In a meeting, the most senior or confident person often sets the direction. With anonymous answers, an idea is judged on its reasoning, not on who said it. Junior experts and quiet specialists count as much as anyone.
Seeing the group's view, and the reasons behind it, gives each expert new information. Some update their answer; others hold firm and explain why. Either way, each round is better informed than the last.
Backing down in public is hard. Because no one knows who gave which answer, experts can quietly revise their view without the social cost of admitting they were wrong.
Experts never need to be in the same place at the same time. A panel can include people from different cities, organizations and time zones, bringing in a wider range of knowledge than one meeting room could.
Write the problem clearly: a forecast, a set of priorities or a guideline to agree on. Decide in advance what counts as agreement, such as three quarters of the panel rating an item 4 or 5 out of 5.
Invite people with relevant and varied expertise, often 10 to 30 of them. Explain the time commitment, the number of rounds and how anonymity will be protected. Expect some to drop out.
Send an open question or a first list of items. Ask each expert to answer alone, giving a reason for each point. A sentence or two per point keeps the next round manageable.
Combine the answers into a clear, anonymous summary: themes for open answers, the median and spread for ratings, and the strongest reasons on each side. Take care not to steer the panel with your wording or grouping.
Send the summary back and ask everyone to rate or re-estimate, explaining any answer far from the group. Repeat until answers stop changing much.
Share where the panel agreed, where it didn't, and why. A disagreement with clear reasons on both sides is a useful finding, not a failure.
A clear question, a panel of experts, a simple survey tool or email, a spreadsheet for the summaries and a few weeks. With AI, use the prompt below to plan and summarize.
Who: a regional health network and a panel of 18 nurses, pharmacists, doctors and patient advocates from four hospitals. Goal: agree on the top priorities for safer discharge. Agreement: 75% rating an item 4 or 5.
Round 1 asked one open question: “What changes would most reduce medication errors when patients leave the hospital?”
Experts replied alone, in writing. Suggestions ranged from pharmacist reviews to large-print medicine charts and follow-up phone calls.
The facilitator merged the 64 suggestions into 22 distinct items, removed every name and turned them into a rating list from 1 to 5.
Experts rated the items, saw the group's median and spread, then re-rated. One anonymous comment about patients' own medicine lists shifted many ratings.
After round 3 the ratings barely moved, so the panel stopped. Five items passed the 75% bar; two stayed contested, with clear reasons on both sides.
A five-point discharge checklist that all four hospitals adopted, plus two open questions for a pilot study.
AI fits Delphi in two ways. It can play a mock panel, so you can test your questions and catch confusing items before real experts see them. And it can do the heavy lifting between rounds: merging answers, stripping out names and summarizing neutrally. What it can't do is replace the panel. Real expertise is the whole point.
Test your round 1 question on simulated experts before sending it out.
Answer my Delphi question as 5 different experts: a nurse, a pharmacist, a family doctor, a patient and a manager.
Turn a pile of open answers into a clean list of distinct items.
Merge these 64 anonymous suggestions into distinct items. Keep the panel's own words where you can.
Summarize a round without steering the panel toward any view.
Summarize these round 2 ratings: the median and spread for each item, plus the strongest reason on each side.
Catch leading, vague or double-barreled items before the next round.
Check my rating list for items that are leading, vague or ask two things at once, and suggest fixes.
Describe your question and panel, pick a mode, then copy the prompt into Claude, ChatGPT, Gemini or Copilot. In coach mode the AI helps you design each round, waits while your real panel responds, then helps you summarize their answers.
The “Open in” links put the prompt in the web address. If your challenge includes anything sensitive, copy and paste instead.
After the session, click one to copy it, then paste it into the same chat.