This appendix supports the working note above. It is a proposal, not a report of a test that has been run.
Conditions held equal
- Work class: one bounded, recurring class of routine asset, for example internal event announcements, with the qualifying template versions and editable fields listed before assignment.
- Materials: the same templates, locks and Brand Kit in both groups.
- Checks: the same factual checks. Pre-approving a format does not authorise every message entered into it.
- Exceptions: the same written rule for what counts as an exception, and the same route for it. Anything outside the eligible class keeps the existing review.
- People and support: the same owner and permissions, and the same training time and support. Whether creators can keep going without the champion's help is assessed after facilitation stops, not held equal.
- The one difference: per-asset brand approval for eligible work (control) against no brand-approval requirement for it (treatment).
Before the treatment starts, confirm in a walkthrough that the designated-template approval setting does what is intended on the team's plan, and that templates have been republished after any change. If the setting cannot implement the treatment, redesign the implementation before running the test.
Measures
| Measure | Purpose | Qualification |
|---|---|---|
| Delivery time and completion by deadline | Test the practical effect on the requesting team | Start the clock at a complete brief and log intake delays separately. Keep unfinished, cancelled and escalated requests in the denominator |
| Defects and their severity | Check acceptable communication, not just matching colours | Use one rubric (identity rules, factual accuracy, readability, required content, correct destination) and hide group assignment from assessors where feasible. Not seeing a rare serious error in a small pilot is inconclusive |
| Creator, reviewer, support and correction minutes | Detect work moving between people | Include setup and template upkeep. These are estimates, not audited savings |
| Exceptions and work outside the agreed process | See whether the delegated class fits real demand | Use a proportionate task log. Do not infer every bypass from Canva activity data |
| Repeated eligible use after assistance ends | Assess whether the workflow becomes dependable | Count real demand, not logins. Early success is not long-term retention |
| Paid scope and retained use | Test the commercial consequence separately | Do not require an upgrade for the test to succeed, and do not read a trial rollout as a causal revenue result |
The quality audit must not quietly become a replacement pre-release approval in the treatment group. It reviews delivered assets, including corrections made afterwards.
Design notes
- Randomise where the policy lives. Assign at team or workspace level, with separate clusters where reviewers are shared. Do not assign two policies inside one approval configuration that cannot be split.
- Match or stratify on baseline volume, review load and asset complexity.
- Qualitative first. Observe a few teams to check the configuration and the eligibility rule, and record setup failures and support needs, not only the teams that finish onboarding.
- Predeclare the observation period, using the real work cadence, the quality limit the organisation will accept and the effect size worth acting on. Do not invent a sample size or an uplift target. Use baseline rates to size any larger test.
- Safeguard. A serious defect, defined in advance, returns that team to per-asset review. Record the defect, the stopping time and the return to review. Retain the team in its originally assigned group for the main analysis, and report treatment exposure separately. Otherwise, removing unsuccessful teams could make delegation appear more effective than it was.
- Success rule. Recommend expansion only if delivery improves, quality remains within the predeclared acceptable limit, and correction, support and maintenance effort do not erase the benefit. Insufficient evidence on quality means an inconclusive result.
- Maintenance check, held equal. Observe one ordinary approved template update in both groups. Can the owner publish it, and can creators use the current source? This is supporting evidence about upkeep. It does not show that every earlier output updates.
- Allow an inconclusive result. Small teams give wide uncertainty. Say so if there are too few completed requests to compare.
Measurement traps
Canva's reports use specific definitions. A template counts as used when it is applied to a design, and its "published" measure includes exporting, downloading, printing or sharing. Applying Brand Kit assets is recorded as use. None of these alone shows that an asset reached its intended audience correctly.
- Fewer review requests could reflect appropriate delegation, lower demand or work outside the process. Check output and quality.
- More designs could reflect useful production, duplicates or drafts. Use the complete-request denominator.
- Brand Kit applied does not show the whole design or message is acceptable.
- Seats purchased measures a commercial decision, not operational dependence.
- Templates created measures setup. The cost of keeping them current still counts.
- Saved hours are not cash. Say whose time it was, how it was valued and whether the capacity was used elsewhere or spending was actually avoided.