# Delphi Method: Consult Experts Without Manufacturing Consensus

Delphi method explained: panel selection, questionnaire rounds, anonymous feedback, stopping rules, and disagreement without false consensus.

- Canonical URL: https://www.harmate.com/en/blog/delphi-method-consult-experts-without-manufacturing-consensus
- Author: Harmate Team
- Published: 2026-09-10
- Updated: 2026-09-10T14:11:22.509598+00:00
- Language: en

## Content

# Delphi Method: Consult Experts Without Manufacturing Consensus

The Delphi method helps a group consult experts about a question that cannot be measured directly. It uses several questionnaire rounds and controlled feedback between rounds to show where views converge and where disagreement remains. It is not a quick vote, and it is not a meeting in which the most confident participant wins.

The term covers different uses, from forecasting to priority setting. The [Office québécois de la langue française](https://vitrinelinguistique.oqlf.gouv.qc.ca/fiche-gdt/fiche/17012115/methode-delphi) describes a separate consultation of experts, while the [French Biodiversity Research Foundation](https://www.fondationbiodiversite.fr/ressource/la-methode-delphi/) highlights anonymity, iteration, and the possibility that the process reveals a diversity of views rather than consensus. This article presents a human, traceable protocol. It does not turn statistical agreement into truth.

## Start with the decision, not the questionnaire

Before recruiting anyone, write the decision the panel must inform. A question such as “what does the group think of our programme?” is too broad. A useful formulation names the context, the time horizon, and the decision that is still open:

> Which capabilities should the next training cycle for team leads prioritise, given the situations they face today?

The question must not force participants into a solution chosen in advance. Check three conditions:

- more than one option is genuinely possible;
- participants have relevant experience of the topic;
- the sponsor agrees to publish limits and minority views.

The [actionable questionnaire guide](https://harmate.com/en/blog/actionable-questionnaires-start-with-the-decision-not-the-questions) expresses a simple rule: collection has value only when it clarifies a decision. Delphi adds a particular requirement: each expert must be able to answer independently from the others.

## Build a defensible panel

A panel is not representative merely because it is large. Define selection criteria before sending invitations: direct experience, role in the decision, variation in context, and ability to complete several rounds. Record who is absent and why.

Anonymity protects freedom to answer, but it does not repair a biased recruitment process. It does not make experts interchangeable, and it does not guarantee the accuracy of their judgement. The professional Delphi overview from [EM-Consulte](https://www.em-consulte.com/article/1480317/a-la-recherche-d-un-consensus-professionnel-la-met) includes problem definition, questionnaire construction and testing, panel selection, data handling, and a pre-defined consensus rule.

State from the beginning:

- the question and the intended decision;
- the planned number of rounds or stopping rule;
- what information will be returned to the panel;
- what data will be retained and what anonymity means;
- what happens if someone skips a round;
- how withdrawals and non-responses will be reported.

A participant may leave because of time, disagreement, or a change of role. Attrition is not a detail to hide: it changes what the final round can support.

## Keep the first round open

An open first round prevents the sponsor from asking experts to choose too early from a pre-written list. Ask them to describe priorities, risks, success conditions, and examples that matter to them. For instance:

> Which capabilities should be developed first in this programme, and in which concrete situations would they make a difference?

Use neutral probes: “Can you give an example?”, “What would change your view?”, “Which condition is missing from this proposal?” Avoid wording that assumes the programme is successful or that a priority is obvious. [Framing bias](https://harmate.com/en/blog/framing-bias-why-your-questionnaires-already-contain-the-answers) can enter before the first round is complete.

After this round, group proposals into understandable units. This requires documented interpretation: preserve original wording, variants, and proposals that cannot be safely merged. A category created by the facilitator is not an automatic discovery made by the panel.

## Structure the later rounds

The second round can present revised proposals and ask participants to rate them. Explain the response scale, but leave room for reservations and conditions. A closed question makes comparison easier; it does not replace the reason behind an expert’s answer.

Between rounds, return aggregated, non-identifying feedback: response distributions, recurring arguments, objections, and new proposals. Do not reveal who wrote what. The [CASRAI Delphi guide](https://casrai.org/guides/delphi-method) distinguishes anonymity, controlled feedback, and iteration. A survey sent twice is not a substitute for all three.

Feedback should remain descriptive. Write “8 of 12 participants selected this priority, three added a condition, and one view did not return in the next round” rather than “the group agrees.” If a minority raises a material constraint, keep it visible even when it does not change the ranking.

## Set a stopping rule without a magic threshold

A Delphi study should state what makes it possible to conclude. The rule may combine:

- a maximum number of rounds;
- stability between two rounds;
- an agreement threshold defined before collection;
- no relevant new proposal;
- a sponsor decision when disagreement remains.

A protocol may choose 75% as its internal threshold. That is not a universal rule, and it does not mean the remaining 25% are wrong. If answers are still changing, if the panel loses many participants, or if several formulations remain ambiguous, stopping because a percentage was reached would be misleading.

The final output should therefore separate convergent items, disputed topics, conditions raised by a minority, and information that is still missing. The decision then belongs to the named owner, who explains what was accepted and what was not.

## A fictional example: choosing a training format

A learning manager must choose among three formats for the next cycle: case-solving workshops, peer practice sessions, and short asynchronous modules. She invites team leads from different departments.

In round one, participants describe recent situations: starting a new role, balancing urgent requests, and communicating a difficult decision. Several mention the need to practise, while two explain that distributed teams cannot meet every week. The facilitator keeps these distinctions instead of creating a category called “preference for in-person sessions.”

In round two, each expert rates the expected usefulness of each format and names the condition that would make the choice fail. Aggregated feedback shows a preference for case workshops, with a recurring concern about available time. In round three, the majority remains stable; three participants insist that an asynchronous pathway should accompany the workshops for remote teams. The final decision is a mixed pilot with one case session and asynchronous preparation. The report says clearly that this is a design decision informed by a selected panel, not proof that the format is best for every team.

## Keep a revision log

A short log makes the process inspectable. For each round, keep:

| Field | What it documents |
|---|---|
| Version and date | The questionnaire actually sent |
| Invited and responding participants | The real scope of the round |
| Changes | Wording added, removed, or merged |
| Shared feedback | Non-identifying aggregates and arguments |
| Disagreements | Minority positions and their conditions |
| Attrition | Withdrawals, non-responses, and known reasons |
| Round decision | What remains open before the next step |

Do not rewrite the history to make progress look smoother. A discarded proposal may explain an important divergence.

## What the method cannot establish

Delphi does not turn expert opinion into an objective measurement. Panel selection, wording, attrition, and consensus rules affect the result. Anonymity reduces some social pressure, but it guarantees neither competence nor good faith. Agreement may reflect shared information, a vague definition, or a desire to finish.

Do not confuse this protocol with an in-person stakeholder consultation, a World Café, or a standardised survey. [Stakeholder consultation without false consensus](https://harmate.com/en/blog/stakeholder-consultation-deciding-without-false-consensus) distributes voice and decision power differently; a World Café uses rotating conversations and a shared harvest. The right choice depends on the question, participants, and trace needed.

## Where Harmate may help, late in the process

Harmate can collect or import questionnaires from several rounds, preserve open contributions, connect themes and contradictions to verbatim evidence, and prepare a report or decision dossier that people can inspect. That can make the evidence log easier to follow. Harmate does not select experts, guarantee anonymity, run Delphi rounds, or automatically declare a consensus valid. Those responsibilities remain human.

The Delphi method is useful when uncertainty must be organised without being disguised as certainty. Consult people separately, return the arguments, preserve disagreement, and explain the final decision: that complete chain is what makes the result defensible.