How to make better decisions: Annie Duke on meetings, feedback loops, and quitting

Executive overview

Most decision-making fails not from lack of intelligence but from bad process: opinions formed in groups get contaminated before they can be compared, feedback loops are treated as fixed when they're actually a choice, and quitting is delayed long past the rational exit point.

Annie Duke's core fix is deceptively simple: make implicit judgments explicit so they can be examined, tracked, and improved. This applies equally to hiring rubrics, investment decisions, meeting structure, and kill criteria for projects.

The biggest lever in organisational decision-making is separating discovery from discussion — collecting opinions independently before anyone talks.

The one meeting change that improves decision quality immediately

• Most meetings conflate three distinct activities: discover (find out what people think), discuss (compare those views), decide (pick a direction).
• Only discussion should happen in the meeting. Discovery and deciding should happen outside it, asynchronously and independently.
• Collecting opinions together in real time causes cross-contamination: the most confident or senior voice anchors everyone else.
• Before the meeting, send prompts to each participant independently — brainstorm, force-rank, estimate timelines — and ask them not to reply-all.
• Google Forms dumping into a private spreadsheet works well for repeated decisions; so does any tool that hides others' inputs until all responses are collected.
• This approach is called a nominal group: people work as a group but not in the same room at the same moment.
• The same technique applies on the fly — if discussion starts to collapse estimates, stop and have everyone write down their number before hearing anyone else's.

Why alignment is the wrong goal

• "Alignment" implies that the meeting's job is to produce agreement. It isn't, and expecting it makes things worse.
• Seeking agreement turns discussion into coercion: people argue to win rather than to inform.
• The correct mode is conveying, not convincing — share your reasoning so others can use it, not to change their votes.
• Disagreement after good discussion is normal and healthy; it reflects the true difficulty of the problem.
• Hearing the full spread of opinions makes people feel genuinely heard and more willing to accept a decision that doesn't go their way.
• The word nevertheless is a practical tool: "I hear you; nevertheless, this is the decision." It signals that input was received without reopening the question.

Making implicit judgments explicit

• Humans apply implicit models constantly — in hiring, investing, product prioritisation — but rarely articulate what they're actually looking for.
• Kahneman's hiring research showed that structuring what "good" looks like and applying a rubric raised hit rates from ~50% to ~65%, a large gain in a noisy domain.
• The gap between knowing a framework and using it is where almost all value is lost. The people who actually implement structured processes have a large edge over those who merely understand them.
• At First Round Capital, Annie built a structured evaluation rubric: partners rate founder, product, and market on a 1–7 scale with shared definitions, plus explicit probability forecasts for Series A funding.
• After five years of data, the firm can now measure each partner's calibration — who is accurate on market quality, who is accurate on founder assessment — and feed that back to improve future decisions.
• Partners discovered that the things they were most emphatic about were sometimes highly predictive and sometimes not predictive at all, with equal confidence in both cases.

There is no such thing as a long feedback loop

• Venture investors often claim their feedback loop is a decade long. This is a choice, not a constraint.
• Every long-term outcome is preceded by necessary (if not sufficient) intermediate signals: Series A funding, product-market fit, net new ARR, talent retention, churn.
• If no company that exited above $1B had failed to raise a Series A, then Series A funding is a required checkpoint — and that result is known in ~16 months, not 10 years.
• The feedback loop is as long as you choose it to be. Shortening it means identifying what must be true on the path to the eventual outcome and tracking those things explicitly.
• There is also a psychological incentive to keep feedback loops long: if you were early into a famous success, a long loop lets you sustain a flattering self-narrative without finding out whether your process was actually good.
• Tighter feedback loops risk finding out you were wrong sooner, which is uncomfortable — but it's also the only way to improve.

Pre-mortems and kill criteria

• A pre-mortem asks: assume this project has failed six months from now — what signals would you have seen along the way?
• Pre-mortems alone rarely change plans; the key is to attach kill criteria to the signals they surface.
• Kill criteria are specific observable events that trigger a pre-committed action: investigate further, escalate, or stop.
• Example from a sales team: if a prospect only wants to discuss price and won't take a demo, kill the deal immediately. If the RFP appears written for a competitor, ask directly how far along that relationship is — then kill or pursue based on the answer.
• This structure prevents the common failure mode of seeing bad signals and continuing anyway because the exit feels socially or emotionally costly.
• Without pre-commitment, intuition about when to quit will almost always fire too late.

Why people quit too late — and what to do about it

• By the time most people seriously consider quitting, they are already past the rational exit point.
• Sunk cost: past investment feels like it will be wasted if you stop, even though waste is a prospective problem — everything you put in going forward on a failing path is the real waste.
• Endowment effect: we value things we've built more than identical things we don't own, making it hard to abandon products or projects.
• Identity: quitting feels like failure to yourself and others. Stopping only when you have "no choice" gives you a socially acceptable story.
• Uncertainty: we want to know for sure it won't work before we stop, but that certainty rarely arrives before the situation forces a decision anyway.
• The Glitch/Slack story illustrates the opportunity cost of staying too long: Stuart Butterfield shut down Glitch at growth of 6–7% per week because the unit economics couldn't reach venture scale. Two days later, he pivoted his internal communication tool into Slack. He could not see that opportunity while still running Glitch.
• Bezos's 70% rule addresses the mirror problem on the start side: be willing to decide with 70% confidence rather than waiting for certainty that never comes.

Mental time travel as a decision tool

• When a decision or situation feels overwhelming in the moment, project forward: how will I think about this in 10 years?
• Most things that feel enormous now — a breakup, a grounding, a setback — are barely recalled a decade later.
• The focusing illusion (Kahneman: "nothing is as important as it seems when you're thinking about it") makes present feelings appear more durable than they are.
• Applying this to near-term choices: "a week from now, when you get that test back, how do you think you'll feel about studying versus playing a game?" — trades the present feeling for a downstream one.
• This reframe works for both high-stakes personal decisions and everyday trade-offs.

Lessons from Daniel Kahneman

• Kahneman actively sought out people who disagreed with him and co-authored papers with them — adversarial collaboration — specifically to design studies that could resolve disagreements rather than talking past each other.
• He was among the first to publicly acknowledge when his own results didn't replicate, including priming research in Thinking, Fast and Slow.
• His curiosity about others' work, not just his own, was the practical expression of his intellectual humility.
• The willingness to ask "why am I wrong?" and be genuinely excited by the answer is rarer among experts than it should be, and is strongly correlated with producing genuine insight.