In short: Agents raise the volume of released change, while reading what each change actually did stays slow and costly. Outcome capacity is how many outcomes get read after release and fed back to Direction. It is bounded by three things: a declared outcome, attribution, and a reading loop with the authority to change a decision. Not every bet needs an experiment, so reading should be tiered by the size and reversibility of the bet, and the number of bets in flight sized to the readings an organisation can complete.
Why isn't faster shipping reaching customers sooner?
Atlassian's State of Product 2027, based on a survey of 1,000 product professionals, puts the tension plainly. Product managers say AI helps their teams ship faster, yet customers are not seeing value any sooner. In the same survey, 86% agree that deciding what to build matters more than ever, and the report lists gut instinct still overriding customer data among the shifts product teams face.
DORA's 2025 research points in a similar direction from the engineering side. AI adoption now shows a positive relationship with both delivery throughput and product performance, and a user centric focus amplifies AI's positive influence on team performance. Speed helps when something is pointing it at a problem and checking whether the problem moved.
Read together, the two studies suggest that delivery got faster while the evidence of value did not keep pace. A plausible interpretation, which the data supports without proving, is that much of the work now waits after release, in a step most organisations never sized: reading what happened.
What is outcome capacity?
Every step of the AI Engineering Production System has one capacity, the limit on how much work it can absorb at its quality bar. At the last step, Observe & Improve, that limit is outcome capacity: how many outcomes get read after release and fed back to Direction.
It closes the loop the rest of the system depends on. In What Is Your AI Engineering System Optimising For?, a change counts as validated business value only after two windows. The health window asks whether the change holds in production. The outcome window asks whether it delivered the value declared for it at intent. Absorption rate counts the changes that hold through the first window. Outcome capacity governs the second.
A change can therefore be absorbed and still unread. It reached production, it did not break anything, and nobody ever checked whether it moved the measure it was built to move. The capacity map names the sign that this capacity is binding: released work rarely has its outcome read, so Direction steers on output.
Why do agents make the loop harder to close?
Three effects stack up as agents join the workforce. The first is volume. Agents raise the number of changes and, through cheaper starts, the number of bets in flight. Reading an outcome is mostly human time: someone has to compare what was declared with what was observed, interpret the difference and decide. That time does not grow with a token budget.
The second is granularity. DORA's AI capabilities model counts working in small batches among the practices that amplify AI's benefits, which is good news for verification and integration. A small change, however, rarely moves a business measure on its own, so attribution has to happen at the level of the bet the change belongs to. When nothing links changes to bets, the evidence dissolves into the general movement of the product.
The third is the base rate. Kohavi and his colleagues at Microsoft reported that only about one third of the ideas they tested in controlled experiments improved the metric they were designed to improve. Large online companies are not typical enterprises, and the ratio would differ elsewhere. If anything close to it holds for an enterprise roadmap, though, most of an unread portfolio is neutral or harmful, and nobody can say which part. Faster building without reading widens that unknown.
The strongest objection is that most organisations already have product analytics and dashboards, often good ones. That is true, and it matters. A dashboard shows the state of the product. Outcome capacity concerns a narrower matter: which bet caused the movement, and whether anyone decided anything because of it. Without attribution and a read date, a dashboard describes the weather without showing whether the umbrella worked.
What bounds outcome capacity?
Outcome capacity is bounded by three supplies, and the weakest of them sets the limit, much as intent capacity is bounded by the weakest of intent, context and policies.
A declared outcome. Each bet states, before it is built, the measure it should move, the expected direction and a date when it will be read. Without that declaration there is nothing to read, only things to notice. This is where outcome capacity depends on intent capacity: an intent with no declared outcome cannot be validated, however well it is specified.
Attribution. The bet can be traced to the measure. In practice that means a bet identifier carried from work management into changes, feature flags and product events, records of who was exposed to what and when, and metric definitions that someone owns. None of this is exotic. It is engineering work, and its natural home is the production system, where it can be maintained with every release instead of rebuilt each quarter by an analytics project.
A reading loop with decision rights. Someone reads the outcome on the declared date, compares it with the declaration, and the reading can stop, extend or double the bet. The result is recorded where Direction can see it. This is the input to the stop loop described in Designing Focus Capacity: a stop loop with nothing read has nothing to act on.
In one organisation where I contribute, delivery is well measured, from pull requests and cycle time to the DORA metrics. The step after release is the hard one: tracking which feature a change belongs to and validating whether it delivered, because the work management system does not yet model the value stream from bet to change. The effect is quiet. Direction receives plenty of evidence about output and little about outcomes, so it steers on what it can see. The pattern is probably common, and it reflects missing attribution more than anyone's intent.

Figure 1. The outcome loop. Readings return to Direction only when an outcome was declared, the change can be attributed to its bet, and someone has the authority to act on the reading.
How much reading does a bet need?
Reading has a cost. Instrumentation takes effort, experiments need enough traffic to detect an effect, analysts and product managers need time, and every reading delays a decision. Demanding a controlled experiment for every change would make outcome capacity bind immediately. The more useful question is how much reading each bet deserves, and the answer depends mostly on its size and how hard it is to reverse, which echoes the risk tiers of the Verification System.
Large or hard to reverse bets, such as pricing changes, a new customer journey or a migration customers will feel, warrant a staged rollout or a controlled experiment. The metric, the expected effect and the read date are declared before the build starts, and the readout goes to Direction.
Medium bets, such as a feature within an existing journey, can usually be read through a flag or cohort comparison at the level of the bet, with a declared read date and some guard against seasonal effects.
Small, reversible changes rely on the health window only. Their outcome is read in aggregate, through the bet they belong to. This is where most agent output is likely to sit, which is why attribution to the bet matters more than measuring each change.
Maintenance, security and compliance work is declared as carrying no outcome reading and is judged on health and risk. Declaring it honestly keeps it out of the outcome count, so the share of work that looks unvalidated reflects the bets that should have been read.

Figure 2. How much reading a bet receives depends on its size and how hard it is to reverse.
Tiering turns outcome capacity into something an organisation can size. In practice, it is roughly the number of readings for large and medium bets that the organisation can complete in a period at its quality bar. If agents push the number of such bets in flight beyond that number, the excess runs unread. Focus capacity suggests capping bets in flight at the binding downstream capacity; when outcome capacity is the one binding, the cap is the number of readings you can actually complete.
Agents can raise this capacity at the margins. They can draft readouts, join event data to bet identifiers, flag bets that have passed their read date and summarise qualitative feedback. The judgement about what a reading means, and the decision it should trigger, stays with people. Agents do the work; people decide.
Where does shared meaning get in the way?
Outcome readings often fail for a reason that has nothing to do with experiments: the measure itself is contested. "Active customer", "conversion" or "order" can carry different definitions in different teams, so two readings of the same bet disagree before anyone interprets them. On the capacity map this is semantic coupling, and it is the usual reason outcome capacity is lower than it looks.
At CTO altitude, the response can stay modest. Every metric used to read a large or medium bet has a named owner and a versioned definition, and changes to it are announced the way an interface change would be. Enterprise semantics beyond that belong with data leadership. Building a Governed Context Supply covers how owned definitions reach the agents that act on them.
What does the CTO need to build?
Most of what outcome capacity requires is infrastructure engineering already knows how to build. The gap is that it is rarely built as one system.
The first piece is bet identity from end to end, from the work management item through branches, pull requests and flags to the events the product emits. The second is flag and exposure tracking offered as a platform capability, so that teams do not each build their own. The third is a read date held as a field in work management, with overdue readings visible to Direction in the same place as the work in flight. The fourth is a short readout format: what was declared, what was observed, and what was decided.
Ownership can follow the same split as the workforce band. Product owns the reading and the decision. Engineering owns attribution. The CTO, or a CPTO where the roles are joined, owns the fact that the loop closes, which is less glamorous than any of its parts and probably the reason it so often stays open.
How do you know outcome capacity is binding?
Three shares give a first signal without any new tooling: the share of bets released last quarter that declared an outcome before they were built, the share of those read by their declared date, and the share of completed readings that changed a decision to stop, extend or double a bet.
If the first share is high and the second low, attribution or the reading loop is the likely constraint. If the third is close to zero, reading has turned into a ritual that Direction never uses. If the first is low, the constraint sits upstream, at Intent.
The question for your organisation
The Atlassian finding describes organisations that ship faster without customers noticing sooner. One plausible explanation is that the system was tuned for supply while the step that turns release into evidence stayed at its old size. Any capacity on the map can become the limit once the ones around it are relieved. Outcome capacity sits at the last step of the spine, and it is the one that feeds the first.
If your organisation doubled what it releases next year, how many of those outcomes would anyone read?
Read further
The AI Engineering Capacity Map: one capacity for each step of delivery, and where the work waits.
What Is Your AI Engineering System Optimising For?: the signature metric and its two windows.
Instrumenting the Metric: Building an Absorption Dashboard: absorption rate and how to compute it.
Designing Focus Capacity: How Many Bets Can You Actually Steer?: the stop loop that outcome readings feed.
Designing Intent Capacity: What Belongs in a Spec, and What Never Should: where the declared outcome is written.
The Absorption Gap: why AI supply outruns an organisation's ability to absorb it.
FAQ
What is outcome capacity?
Outcome capacity is how many outcomes an organisation reads after release and feeds back to Direction. It sits at the Observe & Improve step of the AI Engineering Production System and governs the outcome window of validated business value.
How is outcome capacity different from product analytics?
Product analytics describes the state of a product. Outcome capacity measures whether released bets are read against a declared outcome, attributed to that bet, and acted on. An organisation can have rich analytics and low outcome capacity.
Does every change need an A/B test?
No. Controlled experiments suit large or hard to reverse bets. Medium bets can be read through flags or cohorts, small reversible changes through the health window and the bet they belong to, and maintenance work can be declared as carrying no outcome reading.
How does outcome capacity relate to absorption rate?
Absorption rate counts the changes that reach production and hold through their health window. Outcome capacity covers the next window: whether the bet those changes belong to delivered the value declared for it.
Who owns outcome capacity, product or engineering?
Both, in different parts. Product owns the reading and the decision, engineering owns attribution, and the CTO or CPTO owns the fact that the loop closes.
