A follow-up of sorts to Thinking, Fast and Slow, this book focuses on 'noise' — unwanted, often invisible variability in judgements that should be consistent, distinct from bias (which is a systematic, predictable error). The same underwriter, doctor or hiring manager often makes meaningfully different judgements on similar cases depending on factors as trivial as time of day or mood.
Noise is not bias
Kahneman, Sibony and Sunstein open with a distinction most organisations have never explicitly drawn. Bias is systematic error, always pushing the same way — an underwriter who consistently prices risk too high. Noise is unwanted variability: the same underwriter, given the same case on a different day, produces a meaningfully different answer for no principled reason. Their target analogy does the work. A team whose shots cluster tightly but off-centre is biased; a team whose shots scatter all over the target is noisy. Look at the back of the target, where you can see the scatter but not the bullseye, and bias becomes invisible while noise remains perfectly obvious — which is exactly the wrong way round from how organisations actually inspect their own decisions.
That asymmetry is the book's central practical point. Bias shows up in the pattern of outcomes and has a whole industry devoted to it. Noise is invisible unless you deliberately give the same case to several people, or the same person on several occasions, and measure the scatter — which almost nobody does. And because noise has no direction, it cannot be argued about politically, which the authors clearly regard as an advantage: nobody has to be accused of anything for a company to accept that its quotes vary more than they should.
The book's most quoted evidence is a noise audit the authors ran inside a large insurer, giving realistic case files to dozens of underwriters and comparing what they charged. Executives, asked in advance how much two qualified underwriters pricing the same risk would differ, guessed around ten per cent. The measured median difference was around fifty-five per cent, with claims adjusters not far behind — one underwriter pricing a risk at roughly $9,500 where a colleague priced the same file at $16,700. The organisation was not merely noisy; it had no idea it was noisy, and had been running for years on the assumption that professional judgement was broadly interchangeable. The authors report a senior executive's own estimate of what that variability cost the company annually, in lost business from overpriced quotes and losses on underpriced ones, running to hundreds of millions of dollars.
Three kinds of noise, and an equation
The authors break system noise into components, which is what makes the concept operational rather than merely alarming. Level noise is variation in how harsh or lenient different judges are on average — one underwriter simply prices higher than another across the board. Pattern noise is more subtle and, they find, usually larger: a judge who is tougher than average on one type of case and softer on another, an idiosyncratic personal signature invisible without case-by-case analysis. Occasion noise is the least intuitive and most unsettling — the same person, on the same case, differing by time of day, by mood, by whether they have just eaten, by whether the last case they saw was worse.
Behind it sits a piece of arithmetic that carries more weight than its simplicity suggests. In squared-error terms, overall error decomposes into bias squared plus noise squared. The consequence is that reducing noise improves accuracy regardless of which direction the bias runs, and that at typical magnitudes noise contributes as much to error as bias does. This is the argument that should get a manager's attention: you can improve decisions measurably without first agreeing on what the right answer is.
The historical case that opens the legal material is Judge Marvin Frankel's work in the early 1970s, which documented wildly different sentences handed down for near-identical cases and led directly to sentencing guidelines in the United States. The authors use it as their model of the whole problem — a system where everyone involved believed they were exercising careful judgement, where the variation was enormous, and where nobody had looked until someone thought to compare.
Why expert intuition fails silently
The uncomfortable finding, replicated across fields, is that experts are far more confident in their own consistency than the evidence supports, and organisations rarely check. Worse, professionals experience their judgement as coherent — the sense of a case 'clicking into place' — and the authors argue that this feeling of coherence is produced by the mind's tendency to suppress conflicting evidence, not by the evidence lining up. They call the resulting satisfaction the internal signal of judgement completion, and their view of it is unsentimental: it tells you that you have reached a conclusion, and nothing whatever about whether it is right.
From here the book revisits an old and awkward literature. Paul Meehl's work in the 1950s found simple statistical rules outperforming trained clinicians, and the finding has held up across decades of replication. The authors push it a step further with a result that is genuinely startling: a simple model built from an individual expert's own past judgements typically outperforms that same expert going forward. The model is not smarter. It just has no bad mornings — it strips out occasion noise while keeping the expert's actual policy intact.
They are careful about the limits. Rules and algorithms are not immune to bias, they can bake in historic discrimination, and the authors do not pretend otherwise. Their argument is narrower and harder to dismiss: where judgement is repetitive and comparable, human variability is a cost paid for no benefit, and the alternative to a rule is not perfect judgement — it is noisy judgement that no one has measured.
Decision hygiene
The proposed remedy is deliberately unglamorous, and named on purpose after handwashing: you do not know which specific error you prevented, only that the rate goes down. Because noise, unlike bias, cannot be corrected after the fact by trying harder to be fair, everything has to be structural and in advance.
The practical list is short enough to act on. Sequence information so that judges see facts in a controlled order and one vivid detail doesn't anchor everything that follows — and withhold information that is genuinely irrelevant to the judgement, even when people insist they want it. Break a complex decision into several independently assessed components before combining them, which is their mediating assessments protocol. Gather independent judgements before any discussion, because the first confident voice in a meeting collapses the group's variance for reasons that have nothing to do with being right. Prefer relative judgements — ranking or comparison against concrete cases — over absolute scores on a scale, since people are far more consistent at comparing than at rating. Use shared scales anchored in real examples rather than adjectives. And resist premature intuition: the authors do not ask you to abandon your gut, only to delay it until the structured work is done, which they call the principle of disciplined intuition.
Finally they take the objections seriously, which is one of the book's better sections. Noise reduction can feel dehumanising, it can be gamed, rigid rules produce absurd results at the edges, and people would rather be judged by a person than a formula even when the formula is more accurate. Their answer is that these are real costs to be weighed against a real and usually much larger cost — the arbitrariness of being one of two people who got different answers to the same question, and never knowing it.
Key lessons
- Noise — unwanted variability in judgements that should be consistent — is a distinct problem from bias, and often invisible without deliberate measurement.
- The same person can make measurably different judgements on similar cases depending on irrelevant factors like time of day.
- Structured decision processes and checklists reduce noise more reliably than trusting expert intuition alone.
- Organisations rarely measure their own noise, because doing so requires deliberately testing consistency, not just checking individual decisions.
Unwanted inconsistency in repeated judgement calls (pricing, hiring, reviews) is usually invisible and rarely measured — structuring the decision process reduces it far more reliably than trusting expert judgement alone.
What this means for a UK small business
Any UK business making repeated similar judgement calls — pricing quotes, screening CVs, performance reviews, disciplinary decisions — almost certainly has real noise in the system and has never measured it. A noise audit is genuinely available at small scale and costs an afternoon: take five past jobs, have two people price them independently without conferring, and compare. The spread is usually wider than anyone predicted, and it is money.
The structural fixes translate cheaply. A scoring sheet applied consistently to every CV, with the criteria agreed before anyone reads a name, removes far more unwanted variability than good intentions do — and in the UK it also builds the documented, consistent record that matters if a hiring or dismissal decision is ever tested at an employment tribunal, where inconsistent treatment of comparable people is precisely what gets picked apart. Gathering interview scores independently before the panel talks is free and does more than any amount of unconscious-bias training.
Pricing is where most owners find the money, and it is worth putting real numbers on it. Take an illustrative fit-out firm turning over £600,000 on 80 quoted jobs a year. Two estimators price the same five past jobs blind, and the spread comes back at roughly 20% — one prices a job at £6,000 where the other says £7,200. Assume half the gap is genuine underpricing on work won: on an average job value of £7,500 and a 40% conversion rate, that is 32 jobs a year carrying around £600 of missing margin, or a shade over £19,000 of profit a year — for a firm whose net profit might be £60,000. The other half of the spread is the overpriced quotes that quietly lost, which never show up in any report at all.
That is why the audit is worth an afternoon. A shared rate card anchored to five real past jobs with their actual costs, rather than a per-quote feel, is the small-business version of decision hygiene, and it is the cheapest margin any owner will find this quarter.
What’s aged well
Recent and grounded in rigorous research; likely to remain a key reference on decision quality.
What feels outdated
Nothing significant given recent publication.
Where it falls short
It is a demanding, research-dense book, considerably harder going than Thinking, Fast and Slow, and the volume of studies and sub-categories can bury a core insight that is genuinely simple. It is also longer than it needs to be: the argument is made convincingly by about a third of the way in, and much of what follows is elaboration.
The decision-hygiene chapters read as more aspirational than practical for a small organisation without the resources of the courts, insurers and hospitals where most of the research was done. And it lands in an awkward moment for behavioural science — several findings in the wider field have failed to replicate, and while the noise research itself is largely sound, some of the supporting material inherits that uncertainty.
The Business Stuff verdict
Dense and demanding, but genuinely important for any business making repeated, high-stakes similar judgement calls.
Three things to actually do after reading it
- Identify one repeated judgement call in your business (pricing, hiring) and test how consistent it actually is across similar cases.
- Introduce a simple structured checklist for one high-stakes recurring decision to reduce unwanted variability.
- Separate 'is this decision biased' from 'is this decision noisy' the next time a judgement call goes wrong.
If you liked this, read next
Five similar books
- Thinking, Fast and Slow (Daniel Kahneman)
- Superforecasting (Tetlock & Gardner)
- Thinking in Bets (Annie Duke)
- The Checklist Manifesto (Atul Gawande)
- Predictably Irrational (Dan Ariely)
Common questions
What's the difference between noise and bias, in one sentence?
Bias is being wrong in the same direction every time; noise is being inconsistent, in random directions, on decisions that should match. If two of your staff price the same job and both come in fifteen per cent too low, that's bias. If one comes in high and the other low and next week they swap, that's noise. The practical difference is that bias shows up in your results eventually — you notice the margin is thin — whereas noise cancels out in the average and stays invisible forever unless you deliberately give the same case to two people and compare.
Do I need to have read Thinking, Fast and Slow first?
No, they're independent, and Noise is arguably about the subject Thinking, Fast and Slow left out. Reading the earlier book helps with the vocabulary — System 1 and 2, anchoring, substitution — but Noise defines what it needs as it goes. Be warned that it is the harder read of the two: less narrative, more measurement, and considerably more statistics. If you want one book on judgement and you're not going to finish a dense one, read Thinking, Fast and Slow. If you run a process where several people make the same kind of call, Noise is the more directly useful of the pair.
Can a five-person business actually do a noise audit?
Yes, and it's cheaper at five people than at five hundred. Pick five completed jobs from the last year, strip out the price and the outcome, and give them to two people to quote independently with no discussion. Compare the spread. Do the same with three CVs from your last hire, scored blind. That is the whole method — the corporate version just has more cases and a statistician. What you cannot easily replicate is the follow-up: large organisations can build a model of their own past decisions, and a small firm generally can't, so your fix is a shared rate card and a scoring sheet rather than an algorithm.
Isn't replacing judgement with checklists just bureaucracy?
It's a fair worry and the authors address it directly. Rules do produce absurd results at the edges, they can be gamed, and people genuinely prefer being judged by a human. The counter is that the alternative isn't good judgement — it's judgement that varies for reasons nobody would defend out loud, like the time of day or who spoke first in the meeting. The version that works in a small business is light: agree the criteria before you look at the case, score independently before you discuss, and keep the right to override the score as long as you write down why. That preserves judgement while removing the arbitrary part.
