Measurement
Part of Accounting services measurement: a practical guide
Accounting measurement mistakes, from invisible clocks to averages without distributions
Avoid nine accounting service measurement mistakes involving vague metrics, averages, clocks, corrections, attribution, targets, privacy and inaction.
This list was developed on 5 September 2026 from recurring design failures in service measurement. It is a practical diagnostic, not a claim about the prevalence of any error across England. Check each issue against the practice's own data and decisions.
What to take away
- Available data fields are not automatically useful indicators, so start with the service objective and the decision to be made.
- Every metric needs a numerator, denominator, population and clock, plus exclusions, source and owner.
- Averages hide small groups waiting weeks, so report medians, ranges, counts and serious outliers by service and risk.
- Submission does not prove acceptance, timeliness or correctness, so retain timestamps, responses, acknowledgements, rejections and corrections.
- Targets can distort work, so pair speed with usefulness, completion with rework and volume with quality.
1. Measuring what the system already exposes
Available fields are not automatically useful indicators. Begin with the service objective and the decision a person must make. Add collection only when the expected use justifies it.
2. Leaving definitions in people's heads
Words such as complete, error, urgent and resolved can vary between teams. Give every metric a numerator, denominator, population, clock, exclusion, source and owner. Version changes and do not splice unlike definitions into one trend.
Metric definition checklist
- Numerator
- Denominator
- Population
- Clock
- Exclusion
- Source
- Owner
3. Reporting averages without distributions
An average response time can conceal a small group waiting weeks. Show median or suitable ranges, counts and serious outliers. Segment by service and risk when sample sizes allow fair interpretation.
Average hides the tail
Average only
- Typical wait
- Looks fine
- Slowest clients
- Invisible
- Counts
- Omitted
- Segmentation
- None
Median and ranges
- Typical wait
- Shown honestly
- Slowest clients
- Serious outliers shown
- Counts
- Reported
- Segmentation
- By service and risk
4. Stopping the clock invisibly
Removing time while waiting for a client may support internal diagnosis, but it can misrepresent the client's total experience. Report practice-controlled and end-to-end time together. Define valid pause events and retain evidence.
5. Treating submission as successful completion
A send event does not prove acceptance, timeliness or correctness. Retain the submitted version, timestamp, response and acknowledgement. Record rejections and post-submission corrections separately.
6. Combining unlike corrections
A spelling amendment and a material tax error should not have the same weight. Classify consequence, detection point, cause and affected population. Show high-severity incidents directly rather than hiding them in a broad correction rate.
7. Claiming causation from sequence
If performance improved after a new accountant or tool arrived, staffing, demand, training or customer mix may also have changed. HM Treasury's impact-evaluation quality guidance is for government evaluation, but clearly explains why observed change and causal attribution are different questions.
Use a suitable comparison, investigate alternatives and qualify confidence. Do not turn a monitoring chart into a return-on-investment claim by assertion.
8. Allowing targets to distort work
A response target can encourage empty acknowledgements. A utilisation target can reward starting more work while queues age. Pair speed with usefulness, completion with rework and volume with quality. Audit behaviours around thresholds.
9. Collecting data without decision or restraint
A dashboard can accumulate client and staff details that nobody needs. The ICO's data minimisation guidance explains the requirement for personal data to be adequate, relevant and limited to what is necessary.
Assign access and retention, suppress identifying small groups where appropriate and remove unused fields. More data can increase risk without improving the decision.
Repair the system, not the colour
For every indicator, state the action threshold, decision owner and time to respond. Track whether the action changed the process and created side effects. Government Digital Service performance guidance, although aimed at government services, offers the helpful principle that metrics should reveal what works and inform improvement.
Review measures after changes to staff, software, scope or deadlines. Retire charts that produce no decision. Honest measurement preserves bad results, definition changes and uncertainty, because those are often the signals that make the service better.
Run a quarterly challenge in which someone outside the reporting process reproduces a sample metric and asks how it could be manipulated. Check missing work, reopened tasks, denominator exclusions and manual overrides. Record corrections openly. A dashboard should become more trustworthy after finding an error, not less transparent in order to preserve a smooth trend.
Before you act
- Define the decision each metric supports before collecting data.
- Specify numerator, denominator, population, clock, exclusion, source and owner.
- Report practice-controlled and end-to-end time together.
- Classify corrections by consequence, detection point, cause and population.
- State the action threshold, decision owner and response time.
- Retire charts that produce no decision.
Common questions
Why can an average response time be misleading?
An average response time can conceal a small group waiting weeks. The article says to show median or suitable ranges, counts and serious outliers, and to segment by service and risk when sample sizes allow fair interpretation.
What should be recorded when a submission is sent?
A send event does not prove acceptance, timeliness or correctness. Retain the submitted version, timestamp, response and acknowledgement, and record rejections and post-submission corrections separately rather than treating submission as successful completion.
How should a practice test whether its measures can be manipulated?
Run a quarterly check where an independent person recalculates a sample metric and probes how it might be gamed, looking at gaps, reopened jobs, exclusions and overrides.
