Collaborative decision making: who decides, and how

The phrase sounds like a spectrum with autocracy at one end and everybody voting at the other, and it is usually taught that way. That framing is the source of the problem, because it implies that more participation is better participation, so a team that wants to be good at this drifts toward group decision by default and then discovers it cannot decide anything.

The useful framing is different: participation is a design variable to be set per decision, and the models below answer different questions. Some answer "how do we know we have agreement". Some answer "who has the final say". Some answer "how do we stop the loudest person setting the frame". A team can run all three kinds in a single week without contradiction, as long as it says which one is in play.

The practical failure is almost never that a team picked the wrong model. It is that half the room thought they were consulting and the other half thought they were deciding.

Victor Vroom and Philip Yetton's normative decision model, published in 1973 and revised with Arthur Jago in 1988, is the oldest usable answer and still the clearest. Rather than asking whether a leader should be participative, it treats the degree of participation as a choice to be made per decision, selected by diagnostic questions about how much the decision quality matters, who holds the relevant information, how much commitment is needed to carry it out, and how much time there is.

Five styles, and the point of the table is that the last one is one option out of five rather than the good end of a scale.

If you take one thing from this page, take the habit of saying which of those five is in play at the start of the meeting. It costs a sentence and it prevents the specific resentment that follows from someone believing they were in row five when they were in row four.

Consensus in its formal sense is a real process rather than a synonym for agreement: build a proposal collaboratively, surface unsatisfied concerns, amend, repeat until nobody is holding out. It is used seriously by cooperatives, activist organisations and some Quaker meetings, and it produces genuinely durable decisions when it works.

Its structural defect is easy to state. Because the process continues until nobody objects, it hands a veto to the most resistant person in the room. One holdout, whose objection may be a preference rather than a problem, stops the group indefinitely. In practice groups deal with this by exhaustion, which means the decision goes to whoever can outlast everyone else, and that is a worse selection mechanism than the one they were trying to avoid.

Most organisations that say they run consensus are running the informal version, where the facilitator declares agreement when the objections stop being voiced. That is not consensus. It is consultation with an ambiguous ending, and it is the single most common source of a decision being reopened later by someone who never agreed to it.

Consent is the fix for the veto problem and it is not the same thing as consensus, although the words are used interchangeably by people who have not met the distinction.

Gerard Endenburg, a Dutch electrical engineer, formalised the Sociocratic Circle-Organization Method at his family's company in the 1970s. Under consent, a proposal passes unless there is a paramount objection, and an objection is paramount only if it is reasoned and shows that the proposal would impede the group's aim. "I would prefer the other option" is not an objection. "This would break the thing we exist to do, and here is why" is.

The operative test is not "does everyone like this" but "is it good enough for now, and safe enough to try". That reframing does most of the work, because it converts an argument about the best possible answer into an argument about whether this particular answer is survivable, which is a much smaller question and one groups can actually settle.

The cost of consent is that it requires someone to hold the distinction firmly, in the room, against people who very much want their preference to count as an objection. Without that, consent degrades into consensus within about three meetings.

The advice process inverts the usual arrangement. One named person decides, and can decide anything within their remit, but must first seek advice from two groups: the people with relevant expertise, and the people who will be materially affected. They must genuinely consider what they hear. They do not need permission from any of them.

It was introduced at AES by Dennis Bakke and documented by Frederic Laloux in Reinventing Organizations. Buurtzorg, the Dutch nursing organisation, runs it across roughly nine thousand employees by posting proposals publicly and letting advice arrive from anywhere. What makes it work is that the obligation is real and visible: a decider who obviously did not consult loses standing quickly, which is a cheaper enforcement mechanism than a rule.

The honest caveat, which advocates usually omit: Bakke suspended the process at AES during a liquidity crisis. That is worth knowing, not because it invalidates the method, but because it tells you what the method assumes. The advice process trades speed for quality and buy-in, and there are conditions under which an organisation cannot afford that trade.

The models above describe how a group reaches a decision. These two describe who holds which part of it, which is a different question and often the more urgent one.

DACI, from Intuit and popularised by Atlassian, names four roles for a single decision. The driver runs the process and keeps it moving. The approver makes the call, and there is exactly one. Contributors have knowledge and are consulted. The informed are told afterwards, one way. Its virtue is that it is per decision rather than per project, so it costs a line rather than a document.

RAPID, from Bain, splits the same territory differently and more finely: who gives input, who recommends, who must agree, who performs, and who decides. The letters are deliberately not in process order, which trips people up constantly, and the role worth attention is agree, because it is a veto held by a named person for a named reason. If nobody holds it, say so.

Both are frequently confused with RACI, which is a responsibility assignment matrix: rows are tasks or deliverables, columns are roles, and each cell says who is responsible for doing the work, who is accountable for it being right, who is consulted beforehand and who is informed afterwards. The distinction worth holding is simple. RACI assigns work. DACI and RAPID assign a decision. Teams that try to run decisions through RACI end up arguing about whether the accountable role can overrule the responsible one, which the framework was never built to answer.

Two rules carry over from RACI to all three. Exactly one role is accountable, or approves, or decides: two owners is zero owners. And the columns are roles rather than named individuals, so the structure survives somebody leaving. Walking one of these through in a kickoff meeting rather than circulating it is what surfaces the objections while they are still cheap.

Jeff Bezos's 2016 letter to Amazon shareholders is the most-cited source in this area and it is worth reading as what it actually is: an argument against collaboration as a default.

Type 1 decisions are consequential and effectively irreversible, one-way doors, and they "must be made methodically, carefully, slowly, with great deliberation and consultation". Type 2 decisions "are changeable, reversible, they're two-way doors" and should be made quickly by high-judgement individuals or small groups. His claim is that large organisations apply the type 1 process to type 2 decisions, and that the result is slowness, risk aversion and a failure to experiment.

The same letter supplies the escape hatch from consensus, and it is more useful than most of the frameworks above. "Disagree and commit" is an explicit move: I do not agree with this, I am saying so on the record, and I am going to back it fully anyway. Bezos frames it as a question to ask directly, along the lines of asking whether the other person will gamble with you on it. It works because it separates the two things a holdout is usually conflating, which are being persuaded and being willing to proceed.

The reversibility test is also the fastest way to set participation. If the door swings both ways, the answer to "how much participation does this need" is usually less than the room assumes.

These are mechanics rather than models, and two of them exist specifically to stop the loudest voice setting the frame.

Nominal group technique. Everyone writes their ideas silently and alone first. Then a round-robin where each person contributes one item at a time with no discussion. Then clarification, then an anonymous ranked vote. The silent generation phase is the load-bearing part: it is the only reliable way to get the ideas of people who would otherwise anchor on whatever was said first.

Delphi. Rounds of anonymous input from experts, each round showing the previous round's aggregate, converging without the experts ever meeting. Slow and expensive, and the right tool when status differences in the group are large enough to distort a live discussion.

Dot voting. Each person gets a fixed number of marks to distribute across the options. Fast, visual, and good at surfacing the shape of preference. Bad at surfacing intensity, and easily gamed by whoever votes last, so vote simultaneously.

Fist to five. A five-point hand signal for how strongly you support a proposal, used as a temperature check rather than a vote. Its value is that it makes a weak yes visible, which is exactly what a nodding round of heads conceals.

Nearly every page on this topic asserts that groups decide better than individuals and cites nothing. The research is more interesting than that, and it says something specific: groups are good at pooling what they already share and bad at pooling what only one person knows.

Hidden profiles. Garold Stasser and William Titus demonstrated in 1985 that when information is distributed so that no individual has enough to identify the best option but the group collectively does, groups reliably fail to find it. They spend their discussion on the information members already had in common. Yiwen Lu, Y. Connie Yuan and Poppy McLeod's meta-analysis of 65 studies, 101 independent effects and 3,189 groups, found that common information was mentioned roughly two standard deviations more often than unique information, and that hidden-profile groups were around eight times less likely to reach the right answer than groups given the full picture. That is the single most useful finding in this literature: the meeting is structurally biased toward what everybody already knows.

Groupthink. Irving Janis's 1972 account describes the failure mode of a cohesive group under pressure: an illusion of invulnerability, self-censorship by people with doubts, direct pressure on dissenters, and self-appointed mindguards who shield the group from inconvenient information.

The Abilene paradox. Jerry Harvey's 1974 paper describes something that looks identical from outside and is not. In groupthink the conformity pressure is real. In the Abilene paradox it is imagined: everybody privately disagrees with the course of action, everybody believes everybody else supports it, and nobody says so. The group takes an action that not one member wanted. The remedy is different too. Groupthink needs a structural fix such as an assigned devil's advocate; the Abilene paradox needs one person to ask whether anyone actually wants this.

What the three have in common is that the corrective information existed in the room and was not said out loud, or was said and not heard.

A rough mapping. It is a starting point, not a rule, and the most important column is the last one.

Every model on this page shares one failure point, and it is not in the model. The group runs the process, reaches an outcome, and then a quarter later nobody can say what was decided, who decided it, or which objections were raised and answered.

Consent is the sharpest case, because a consent decision is defined by the objections that were raised and resolved. If that content exists only in what people said out loud in the room, then the decision as recorded is not the decision that was made; it is the conclusion with its reasoning removed. The advice process has the same shape: it obliges the decider to have genuinely considered the advice, and there is no way to demonstrate that from memory in March about a call in November.

What that argues for is a decision log, and specifically for the field most teams leave out, which is the options that were rejected and why. A log of conclusions is a list. A log that records what was considered and discarded is the thing that stops the same argument recurring.

The mechanical difficulty is that nobody wants to be the person typing during a decision meeting, and the person most able to capture the reasoning is usually the one arguing. This is the problem Earkeep was built for. It records your day continuously on your own device, transcribes it there, and writes plain files you own, so the material for the decision record is already there when you decide you need it. You mark the stretch of the day that was the meeting, and an agent working against it drafts the record, rather than somebody reconstructing it from notes nobody took.

Be clear about what that does and does not do. Earkeep does not decide anything, does not detect that a decision happened, and does not write the log for you. The record is what was said, in order, with timestamps: it does not identify speakers, so it cannot tell you who raised an objection. What it changes is the starting point, from recollection to transcript. Telling the room you are recording is yours to do, and there is a page on how to say it.

Decisions are made against something. If it is a set of objectives, see OKR and SMART; if it is a backlog, the prioritisation methods in MoSCoW and RICE are decision procedures in their own right and should be named as such. The ceremonies where most of these decisions actually happen are in standup and retrospective formats, and the case for having fewer decision meetings at all is meeting overload. Agreeing the decision path at the start of a project, rather than discovering you need one, belongs in the kickoff.