Professional

#NoEstimates: Can We Really, and Should We, Stop Estimating?

Automatic translation · View original

Topic illustration: #NoEstimates: Can We Really, and Should We, Stop Estimating?

Few labels have sparked as much debate in the agile community as #NoEstimates. For some, it is a long-overdue liberation from the theater of estimation. For others, it is an irresponsibility no serious business can afford. As is often the case, the provocative label hides a far more nuanced conversation than its name suggests. Nobody serious actually proposes working blindly or promising things with no sense of timing at all. The underlying question is a different one: what does estimating really give us, how much does it cost us, and are there ways to forecast the future without falling into the ritual of story points? We are going to try to answer it without fanaticism for either side, which is exactly what this debate usually needs.

Where the label comes from

The movement that became popular under the hashtag #NoEstimates was driven, among others, by Woody Zuill and Vasco Duarte, starting from a widespread and very real discomfort. Teams spent a remarkable amount of time estimating, the estimates were rarely accurate, and yet they were used to commit to deadlines and evaluate performance as if they were reliable. The question they raised was deliberately provocative: if estimating costs so much and is so often wrong, why do we do it?

It is worth clearing up a misunderstanding right away, because it poisons half of the discussions on the topic. #NoEstimates does not mean, in its serious version, promising that something will be ready without any notion of when, nor giving up all forecasting. It means questioning estimation of effort task by task as the default practice, and exploring whether there are better ways to answer the legitimate questions that estimation is meant to address. It is more an invitation to rethink a habit than a dogmatic prohibition. The name, deliberately radical, has worked against it, because it invites caricature as a defense of improvisation, when its most serious proponents defend exactly the opposite: more rigorous forecasting based on data.

What questions does an estimate try to answer?

To judge whether we can do without estimating, we first need to understand what we estimate for, because it is not always what we believe, and often very different purposes coexist under the same word.

Sometimes we estimate to decide whether something is worth doing, comparing expected cost with expected value. That is a legitimate question and hard to answer without some notion of size, even a rough one. If something will cost a fortune and deliver little, it is better to know before starting. Other times we estimate to coordinate, because other teams, other areas, or the clients themselves need to know when something will be ready in order to organize their own work. Also legitimate, because a team's work rarely exists in isolation. And many times we estimate for less admirable reasons: because someone wants a date to put pressure on, because the organization measures teams by their compliance, or simply because it has always been done that way and nobody remembers to question it.

The criticism behind #NoEstimates aims mostly at this last group. When estimation serves neither to decide nor to coordinate, but to control or out of sheer inertia, it is pure waste disguised as rigor. The problem is that in many teams these three motivations are mixed together indistinguishably, and beneath the innocent appearance of "we need to estimate to plan" sometimes hides "we want a date so we can demand accountability later". Separating these purposes is the first step toward an honest conversation about the topic, because each one admits different answers.

The hidden cost of estimating

Estimating is not free, although its cost is rarely recorded anywhere. There is the direct time of estimation meetings, which in a team can add up to many hours a month between refinement sessions, planning poker, and discussions about whether something is a five or an eight. But the greater cost is usually indirect and harder to see.

When an estimate becomes a commitment, it triggers the well-known Goodhart's law, which the anthropologist Marilyn Strathern formulated by saying that when a measure becomes a target, it ceases to be a good measure. The figure stops being an honest prediction and becomes a target that must be met, with all the distortions that entails. Teams learn to inflate estimates to protect themselves from pressure, or to cut quality to meet what was estimated, or to argue about the definition of points instead of about the work. Estimation, which was born to inform, ends up distorting the behavior it was meant to measure.

In addition, estimating effort feeds a dangerous illusion, that of precision. A number, however invented its origin, conveys a false sense of control that can lead to worse decisions than honestly acknowledging uncertainty. Saying "this is thirteen points" sounds more serious and more reliable than saying "we are not sure, it could take quite a while", even though the second sentence is more truthful. The organization clings to the number because it gives it comfort, not because it is accurate.

Daniel Kahneman documented in his work on human judgment what is called the planning fallacy, our systematic and persistent tendency to underestimate how long it will take us to complete tasks, even when we have abundant prior experience that should warn us otherwise. It is not a flaw that can be corrected with more discipline or more experience. It is a structural feature of how human beings think, and it shows up again and again in all kinds of projects. Estimating better, in many cases, is a battle lost in advance against our own minds, and acknowledging this is more honest than promising that this time we will get it right.

The alternative: predicting with data instead of guessing with opinions

Here is the constructive part of the debate, which is often lost between the provocation of the label and discussions about whether it is heresy. Those who question traditional estimation do not propose blindness or improvisation, but replacing guesswork with forecasting based on real data.

The idea is simple and surprisingly powerful. Instead of estimating each task, a team can measure how many work items it completes per unit of time, that is, its throughput, and use that history to forecast when a set of work will be finished. If a team completes on average a certain number of items per week, and we know how many items remain, we can project a probable range of dates without having estimated a single task individually. The team's real past predicts its future better than its own opinions about how long things will take.

This approach, closely tied to Kanban thinking and developed rigorously by authors such as Daniel Vacanti, has notable advantages over traditional estimation. It is based on what the team actually does, not on what it believes it will do, which makes it immune to the planning fallacy. And it expresses the result as a range of probability rather than as a single, falsely precise date. Saying "there is an 85% probability of finishing between these two dates" is both more honest and more useful than saying "it will be ready on such a day", because it communicates uncertainty instead of hiding it, and allows whoever receives the forecast to decide how much risk they are willing to take.

There are even more sophisticated techniques, such as Monte Carlo simulations applied to the team's history, which Troy Magennis and others have popularized. Instead of a single projection, thousands of possible futures are simulated from real performance data, and a probability distribution of delivery dates is obtained. It sounds complex, but the underlying idea is accessible: use what we know about the real behavior of the system to predict its future behavior, instead of asking people to guess.

For all of this to work, one condition is needed that is often overlooked and worth emphasizing: slicing the work into items of reasonably similar and small size. If the units are comparable to one another, counting how many get completed is sufficient and the individual size of each one matters little, because the variations offset each other in the aggregate. A good part of the craft thus shifts from estimating to dividing well, which is a more useful and more learnable skill. A team that knows how to split work into small, homogeneous pieces needs to estimate much less, because the uniform size itself does the work that estimation used to do.

When to stop estimating and when not to

The mature answer is not an absolute yes or no, but an "it depends" that should be spelled out honestly, because both the dogmatism of estimating everything and that of estimating nothing are ways of ceasing to think.

Doing without effort estimation makes most sense when the team works on a continuous flow of similar items, has enough history to forecast with data, and operates in a context where a range of probability is an acceptable answer for whoever receives it. In that case, estimating task by task adds little that data-based forecasting cannot provide better, faster, and with less wear. The team can spend on building the time it used to spend arguing about points.

Estimation still makes sense, on the other hand, in several specific situations. When an important investment decision must be made before starting, and a notion of cost is needed to compare it with the expected value. When there is no history because the work is genuinely new and there is no data to start from. And when the contractual or coordination context requires a specific commitment that does not accept a simple range of probability. Denying these situations would be as dogmatic as estimating everything out of habit, and would do a disservice to anyone seeking an honest answer.

There is also a political dimension that should not be ignored, because it is often the true root of the conflict. Sometimes the problem is not the estimation technique, but what it is used for. If an organization uses estimates to pressure people, to compare teams, or to assign blame when something is late, changing technique will solve nothing, because the underlying conflict is one of trust, not of method. In such environments, the honest conversation is not about estimating or not estimating, but about why that level of control is needed and what it says about the relationship between the organization and its teams. Moving from story points to data-based forecasting in an environment of distrust will only change the tool through which pressure is exerted.

The AI context: estimating the work of a machine

The arrival of artificial intelligence adds a new and little-explored wrinkle to this debate, one worth examining closely because it changes some of the underlying premises.

Traditional estimation of human effort loses much of its meaning when a portion of the work is carried out by a machine that can generate in minutes what would take a person days, but that may also require several unpredictable rounds of review and correction. Variability skyrockets. A task can be solved almost on its own or turn into a pit of successive corrections, and anticipating which of the two will happen is even harder than estimating traditional human effort. Story points, which were already a rough approximation, become almost irrelevant when the actual effort depends on how an agent whose behavior we do not fully control responds.

This reinforces, rather than weakens, the argument in favor of forecasting with data. In an environment where effort per task is hard to anticipate, measuring the real throughput of the whole system, humans and agents working together, and projecting from that history is more reliable than estimating each piece separately. The system as a whole still has a measurable rhythm even when each individual task is unpredictable. AI does not resolve the #NoEstimates debate, but it does tip the balance toward empirical forecasting based on the behavior of the system and against guessing task by task, which was precisely what its proponents had been defending since before AI existed.

There is also a new risk worth pointing out. If we measure productivity by the quantity of work completed and AI inflates that quantity, throughput-based forecasts can also become distorted. A team that closes many more items thanks to AI does not necessarily deliver more value, so it is worth remembering that forecasting when we will finish is not the same as ensuring that what we finish is worthwhile. Forecasting answers the when, never the what or the why.


The value of #NoEstimates lies not in its answer, which its most radical defenders sometimes exaggerate to the point of caricature, but in its question. It forces us to examine a practice that many teams carry out out of pure habit, without asking what it gives them or what it costs them. And that question is healthy regardless of the answer each team arrives at, whether it ultimately decides to keep estimating, to replace estimation with data-based forecasting, or to combine both depending on the context.

Perhaps the most useful conclusion is to stop treating estimation as a moral obligation of professionalism, something done because serious teams do it, and to start treating it for what it really is: one tool among several, with a cost and a benefit that should be weighed in each specific situation. Sometimes the most professional answer is not to estimate better, but to stop estimating what does not need to be estimated and to measure instead what does matter. And sometimes, when a large investment decision is on the table, the professional answer is indeed to devote time to estimating carefully. Dogma, in either direction, is the enemy.

And in your team? Do you estimate out of conviction that it gives you something real, or out of inertia and because it has always been done that way? Have you ever tried forecasting with data instead of estimating by effort, and what happened?

Bibliography

  • Anderson, D. J. (2010). Kanban: Successful Evolutionary Change for Your Technology Business. Blue Hole Press.
  • Duarte, V. (2016). No Estimates: How to Measure Project Progress Without Estimating. Oikosofy.
  • Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
  • Magennis, T. (2011). Forecasting and Simulating Software Development Projects: Effective Modeling of Kanban & Scrum Projects using Monte-Carlo Simulation.
  • Reinertsen, D. G. (2009). The Principles of Product Development Flow: Second Generation Lean Product Development. Celeritas Publishing.
  • Strathern, M. (1997). 'Improving ratings': audit in the British University system. European Review, 5(3), 305-321.
  • Vacanti, D. S. (2015). Actionable Agile Metrics for Predictability: An Introduction. ActionableAgile Press.
  • Vacanti, D. S. (2019). When Will It Be Done? Lean-Agile Forecasting to Answer Your Customers' Most Important Question. ActionableAgile Press.

🔒 Restricted Poll

Please log in to your Scrum Manager account to do this.

Log in
Share:

Comments (0)