From individual productivity to delivery flow

12.1. Local acceleration needs to cross the whole system

Consider a fictitious numerical example: a delivery takes 100 hours of elapsed time, of which 40 correspond to implementation and 60 to decisions, waiting, integration, and validation. If implementation drops to four hours and everything else stays the same, the delivery will take 64 hours. A tenfold speedup in that stage produces roughly a 1.56× speedup overall.

The calculation is purely illustrative and assumes sequential stages. It shows why the engineer needs to see where time is being consumed. The next improvement may lie in test environments, contracts between teams, CI capacity, or product decisions.

DORA's AI Capabilities Model identifies seven capabilities associated with better outcomes with AI: a clear organizational stance, a healthy data ecosystem, AI access to internal data, version control practices, small batches, user focus, and quality internal platforms. The model reinforces that the collective environment shapes the results. F04

12.2. A small dashboard with useful questions

The dashboard proposed below combines outcome and process measures. There is no need to instrument everything before starting; select what answers the current bottleneck.

For delivery indicators, DORA's current guidance presents five metrics: time from commit to production (change lead time), deployment frequency, failed deployment recovery time, the proportion of deployments requiring immediate intervention, and the proportion of unplanned deployments resulting from incidents (deployment rework rate). They should be contextualized by application or service, avoiding artificial comparisons between different kinds of work. F38

The SPACE framework also guides measuring productivity across several dimensions, including satisfaction, performance, activity, collaboration, and efficiency. This approach helps prevent token consumption, commits, or code volume from becoming the purpose of the work. F11

12.3. Measuring cost without hiding human attention

An illustrative operational formula is:

cost per accepted change =
  (cost of models and tools
   + cost of execution and validation infrastructure
   + cost of human time for preparation, review, and correction)
  / number of accepted changes in a comparable class

It does not replace measures of value, quality, and consequence. It serves to reveal that a configuration that is cheap per token may require more rework, and that a more expensive configuration may free up enough human time to make up for it.

The classes need to be comparable: fixing a piece of text and migrating a database have different profiles. It is also advisable to track the period after delivery to capture failures and fixes that did not appear at the moment of acceptance.

Record calendar time, agent time, and human effort separately. Concurrency changes the relationship between them. An agent can work for hours while the person handles another request; adding the times together as if they were a sequential activity produces a misleading measure.

12.4. Turning recurring failure into improvement

A simple procedure for the team:

1. Record the failure with the task, configuration, and evidence.
2. Distinguish ambiguous requirement, inadequate context, environment, tool, verification, and model limitation.
3. Choose the simplest intervention that addresses the cause.
4. Re-run a case that reproduced the failure.
5. Share the improvement and watch for recurrence.

Possible outcomes include a regression test, an API constraint, a usage example, a documentation update, or a clearer tool. The growth of instructions should respond to real need; obsolete guidance also needs to be removed.

This practice extends an engineering tradition already expressed by SRE: reducing repetitive operational work by eliminating its cause. AI makes it more accessible to build automations, but it is still necessary to choose where they produce benefit. F34

12.5. Organizational responsibilities

An engineer can improve their local context, create checks, and share procedures. Some conditions require institutional support: suitable environments, authorized access, budget, CI capacity, integration with internal data, and criteria for professional evaluation.

Leadership should recognize work that increases collective capacity, even when it reduces the visible volume of intervention by the person who did it. This includes useful documentation, infrastructure, training colleagues, and eliminating recurring tasks.

The 2025 study The SPACE of AI gathered responses from more than 500 developers along with qualitative observations, highlighting the role of organizational support and peer learning. These results are consistent with the collective dimension of the manifesto. F13

An adoption initiative should have an owner for the flow, prioritized problems, and tracked outcomes. Distributing tools without solving access, integration, and verification leaves an important part of the potential with no path to reach the user.