Engineering outcomes

What it is

Engineering outcomes are changes in what people can do or what a
system can reliably deliver after an engineering intervention. An
operator can recover an interrupted import without asking a
developer to repair data by hand. A dispatcher can tell whether an
order is on its way before promising a delivery date. A maintainer
can safely update a rule without changing three independent
implementations. The work may finish with a release, a migration, a
written procedure or a decision to remove a failing dependency. The
outcome is the difference the work makes in use. A harmful change is
also an outcome and belongs in the review.

It helps to distinguish the deliverable, the outcome and the later
impact. A searchable inventory is a deliverable. Finding a valid
stock record on the first attempt is an outcome a team can observe.
Booking more orders with fewer returns is a possible business
impact. Changes in supply, pricing or the mix of customers also
affect that last measure. Treating the three as interchangeable
encourages a project to claim effects it cannot attribute or to
postpone useful learning until a business metric finally moves.

An outcome statement can name the person or process, what changes,
under which conditions and how someone will see the difference. It
can concern accessibility, support effort, correctness, security or
recovery as much as sales. When you need several measures, give them
separate jobs. A primary observation tells you whether the intended
behavior improved. A guardrail says which quality you must protect
while trying to improve it. The two may have different units and
different owners.

Why it matters

Teams rarely agree on what good means just by agreeing on a feature.
A warehouse manager wants a trustworthy number before loading
starts. The on-call engineer wants an alarm early enough to
intervene. Finance wants to know which records were included. A
faster synchronization job can leave all three worse off if it
quietly drops late events. Naming the result beforehand lets them
ask for useful evidence while a design can still change.

Give an agreed number a denominator and a clock. A completion rate
needs a specified population of attempts. A response time needs a
start event, an end event and a view of the slow tail. A recovery
measure needs a definition of restored service. A weekly total
without the volume of traffic can hide a deteriorating rate.
Recording definitions and a short baseline makes the next reading
comparable and leaves the team a way to explain a change in the
result.

Delivery metrics still provide valuable evidence. DORA describes
ways to measure delivery throughput and instability, including the
time from committing a code change to deploying it in production and
the time to recover from a failed deployment. Those measures help a
team see how safely it can get a technical change into use. A
separate observation can show whether that change made the intended
user task easier. Both questions belong in a serious review, and
answering one leaves the other open.

What it looks like in practice

Engineer ApS describes its commitments about documentation, support
and handover on its service levels page. Its engagement guide
explains that the deliverable, responsibilities and billing basis
are agreed in writing. A buyer can use those commitments to ask for
a practical handover outcome. Give another authorised maintainer a
representative recovery exercise with the documentation and access
they will receive. Record whether they can find the right procedure,
identify the failure and restore a known test case. Agree in advance
which steps require someone else's approval. The exercise is a
suggested way to assess a handover.

Consider a fictional nightly import that feeds Monday purchasing.
The buyer's issue is having to wait for a specialist to interpret an
incomplete report. A proposed outcome is that the duty operator can
recognise a stale report and retrieve a reliable one before
purchasing starts. Start by observing recent runs with permission
and writing down what a stale result looks like. Then choose a small
sample of clean imports and recovery scenarios that a reviewer can
reproduce. Include delayed source files and duplicate deliveries,
since the team needs to know which day is represented and whether it
counted a row twice.

Record how many scenarios the operator completed unassisted, the
elapsed time from alert to a verified report and any unresolved
discrepancies. Confirm that repeated recovery still preserves totals
and avoids damaging current data. Store an example observation next
to the definition so the next reviewer can follow the same
procedure. A rehearsal can establish that the new path works under
the conditions tested. Live observations after a normal operating
cycle add evidence about its performance in ordinary use.

Where it goes wrong

An easily collected count can become a substitute for the result.
Closing many alerts may make a dashboard look healthier even while
important failures spend longer unnoticed. Improving average
response time may leave the slowest requests unusable. Start with
the task that is meant to get better and then choose what to count.
Keep failure cases and their context available so that a promising
average cannot erase them.

Before-and-after data can also be misleading. A new import path
might be launched in a quiet week, or its early traffic might come
from a single supplier with cleaner files. Keep comparable
observations by workload and retain the denominator. In a pilot,
describe the cohort, dates and important differences. A small review
can still justify extending a reversible trial or repairing a clear
fault. It supports a claim at the scale observed.

Finally, give someone responsibility for checking the result after
release. Name the owner, the first review date and the evidence that
would justify keeping, adapting or rolling back the change. If
normal use is rare, agree on a rehearsal and a way to review an
actual occurrence when it happens. A delivery review should leave
the next person enough information to make that decision without
reconstructing the project from commit history.

Further reading

  * GOV.UK Service Standard, define success and publish performance
data

  * DORA, software delivery performance metrics

  * Google SRE Workbook, implementing service level objectives

hThis page on the web	URL:https://engineer.company/glossary/engineering-outcomes/	engineer.company	70
