A useful corporate gifting experiment starts before anyone chooses a box. Decide who is eligible, what differs between the two groups, what outcome matters, how long it can be observed, and which costs count. This free, versioned planner turns those decisions into a shared record and calculates descriptive allocation, cost, and outcome figures. It does not decide whether the observed difference is statistically convincing or prove that a gift caused a business result.
Start with the decision, not the uplift you hope to see
The experiment question must identify a decision the organization could actually take. “Do gifts work?” combines different audiences, occasions, products, messages, and measures. “For eligible existing customers invited to an October renewal conversation, should we send the approved choice invitation or the current no-gift follow-up?” gives an operator two executable options. The control need not be “nothing”: it can be the current approved practice. Write the exact difference between arms and freeze it before the first allocation. If message copy, sender, timing, and follow-up also change, a later difference cannot be assigned to the gift alone.
Give the decision an owner who can act on the answer, an analyst who can preserve the allocation and check outcomes, a privacy owner who can approve the data path, and an operations owner who can deliver both arms as specified. These are distinct jobs even in a small team. The budget owner should approve the maximum exposure before the experiment starts. A marketing manager may own the commercial decision, but should not quietly redefine the outcome after seeing the data.
The National Institute of Standards and Technology's experimental-design handbook frames design around objectives, variables, and the choice of randomization or blocking. The UK government's comparative-study guide likewise describes a hypothesis, control and variation, sample, duration, randomized split, and analysis. Those are method references, not evidence that any particular gifting campaign produces a lift. Our planner applies the sequence to a gifting decision while leaving formal statistical design to a qualified analyst.
Freeze a one-page experiment contract
Open the Hypothesis sheet and state a single primary binary outcome. Examples include a qualifying meeting held within 30 days or an accepted renewal conversation within a fixed period. “Engagement” is not sufficient until someone defines the event, source system, and timestamp. Choose the unit that will be randomized: person, account, team, or another defensible cluster. When two contacts at one customer account can affect each other, account-level assignment may be safer than person-level assignment, even though fewer independent units can make a study harder to interpret. The analyst should record the reason rather than silently treating contacts as independent.
Record the observation start and end dates, the eligible population rule, every exclusion, the approved arms, the budget ceiling, and the decision rule. One practical decision rule is: do not scale treatment B unless delivery worked, the predefined primary measure has enough complete observations for a qualified review, total cost is acceptable, and no material privacy or recipient-experience issue emerged. This is a governance rule, not a statistical threshold. If sample size or minimum detectable effect is necessary to make an investment decision, have an analyst calculate them separately before launch; this file deliberately does not estimate power.
Write down the competing explanations too. A seasonal promotion, account-team intervention, different email sender, stockout, or delayed delivery could move the outcome without the gift design being the cause. The design should make those changes visible. Do not promise a “clean” trial by deleting inconvenient records afterward. If a record is excluded, the rule and timing of exclusion should already be specified. Preserve both the initial eligible list and a dated exception log so reviewers can reproduce who entered the experiment.
| Contract field | Decision to freeze | Owner | Acceptance evidence |
|---|---|---|---|
| Population | Eligibility, exclusions and unit | Program owner | Dated list of pseudonymous IDs and approved rule |
| Arms | Only the planned difference | Operations owner | Both arm specifications and change log |
| Outcome | One primary event and fixed window | Analyst | Event definition, source and timestamp rule |
| Money | Gift, shipping, tax allowance and handling | Finance owner | Approved ceiling and reconciled cost basis |
| Safeguards | Privacy, recipient policy and escalation | Privacy and policy owners | Written go/no-go decision |
The table is a preregistration checklist, not a promise that every organization has the same approval path. A small local test may use one person for several roles, but the evidence should still show which decision that person made and when.
Build a fair eligible population and stable assignment
The Eligible Population sheet holds pseudonymous IDs, strata, eligibility, exclusion reason, privacy notice or basis, and approver. Do not paste a customer relationship management export with names, addresses, email, personal notes, or sensitive categories into this workbook. Keep the linkage between an ID and a real person in the approved source system with access and retention controls. An apparently anonymous workbook can still identify people when its IDs are reversible, so the privacy owner must assess the actual workflow rather than assume a label makes it safe.
For a fuller handoff map across invitation, recipient choice, fulfillment and deletion, use the corporate gifting data-governance guide. This experiment planner intentionally narrows that larger journey to the fields needed for an approved comparison.
Choose eligibility before looking at outcomes. A customer pilot might include active accounts with a renewal meeting due in a specified month and exclude accounts with an open complaint or a policy ban on gifts. An employee pilot might include one job family across multiple offices but exclude managers who cannot receive the same intervention under employer policy. A decision about exclusions can protect a recipient; it can also bias the answer if made after allocation. Keep the rule consistent between arms and record unavoidable late exclusions as exceptions, not hidden deletions.
The workbook does not create random numbers. The analyst must generate a reproducible, frozen number for each eligible pseudonymous ID outside the file, using an approved method and a secured seed or allocation record. Enter a value from 0 inclusive to 1 exclusive in Assignment. Values below 0.5 map to A; values from 0.5 to below 1 map to B. Blank, negative, and 1-or-higher values do not allocate. The same person must not be rerandomized because a gift is unavailable. The sheet flags duplicate IDs, and the QA sheet exposes eligible records without an arm. The Diagnostics sheet displays group counts; an unequal split in a small sample is a reason to inspect the process, not automatically a defect or proof of a biased generator.
If market, account size, or employee site could strongly affect the result, consider strata or blocked randomization with a statistician. Record each stratum before assignment and allocate within it. The current workbook retains a stratum label but implements the simple 0.5 rule; it does not enforce exact balance within each stratum. If exact blocked allocation is required, use an approved external assignment procedure and import its frozen results under a new workbook version rather than representing the current formula as a complete block-randomization engine.
Walk through the nine sheets without treating a formula as a policy
Instructions names the decision sequence and reminds the team to keep identity and delivery data elsewhere. Hypothesis captures the owner, primary outcome, dates, unit, arms, and exclusion. Eligible Population stores the preapproved pseudonymous roster. Assignment links that roster and holds the analyst's frozen random values. Costs budgets both arms, while Outcomes links the assignments and records the binary result and observation date. Diagnostics counts allocation, in-window observations, positive outcomes, observed rates, their B-minus-A difference, and estimated arm costs. Methodology states what the workbook does not prove. QA exposes duplicates, unassigned eligible IDs, missing in-window outcomes, and an invalid observation window.
Work in that order. If you start on Diagnostics, the zeros can look like a legitimate null result when the real problem is an empty or incomplete roster. In the blank template, zero is a placeholder calculation, not a campaign finding. Enter both dates in Hypothesis, then inspect QA. The workbook's review state remains an instruction to review results or resolve inputs, not an approval button. An analyst should sign off on a versioned copy after reconciling the output to source events.
The CSV is intentionally narrower: it provides eight fields for ID, stratum, eligibility, random value, arm, ID check, outcome, and observation date. It is useful for exchange with an authorized analyst or approved tool. It cannot preserve the nine-sheet workbook's validation, cost assumptions, methodology, or quality checks. If a team exports it to another system, the data steward must confirm encoding, date interpretation, delimiter behavior, and access permission. Importing a file and seeing rows is not enough; match record counts, unique IDs, and a sample of dates back to the approved source.
When the roster exceeds 100 rows, this v1.0 workbook must be extended and retested. Its formulas and checks are populated for rows 5 through 104. Appending row 105 without extending all dependent ranges may silently omit someone from the counts. The analyst should version the expanded model, test first and last rows, and record the change. Do not disguise a capacity limit as a business limit on the number of recipients.
Put the full cost beside the observed result
In Costs, enter a budgeted recipient count and four unit components for each arm: gift, shipping, tax allowance, and handling. Total estimated cost is the recipient count multiplied by the sum of those units; cost per recipient divides the result by that count. This is an estimate, not a tax determination, booked expense, or statement of actual invoices. Finance should decide whether unused invitations, refunds, duties, platform fees, local currency conversion, and support costs require separate lines in its own ledger. Do not present a cheaper-looking arm by excluding a cost type from it alone.
For a fair operational comparison, separate three views: planned exposure, actual committed or delivered cost, and the outcome window. A gift option that looks effective but arrives late may be an execution failure. A no-gift control can still have outreach and staff costs. Whether those common costs are in both arms depends on the decision, but the rule must be written before results. If the campaign spans currencies, use a dated conversion policy and retain local amounts; a single unlabelled aggregate obscures exchange-rate effects.
The observed rate in Diagnostics uses only records that both have an assigned arm and a completed binary outcome dated within the declared window. It is not the same as an intention-to-treat estimate over every randomized unit. That distinction matters when delivery fails, recipients decline, or events arrive late. Review missingness by arm before comparing rates. An apparent advantage can arise because one arm has more incomplete event capture. A qualified analyst may choose a different estimand and missing-data treatment; the workbook provides a transparent descriptive starting point, not a substitute for that choice.
The cost and rate difference can be read together without turning them into a fabricated return on investment. For example, a hypothetical B arm that costs more per invite and has a higher observed rate still needs a decision about the value of an outcome, uncertainty, side effects, and budget. The workbook does not multiply a rate difference by pipeline value or claim incremental revenue. Connect it to a separate corporate gifting KPI and attribution framework only after those definitions and the downstream decision are clear.
Worked decision one: a customer renewal invitation
Hypothetical case: a marketing operations lead wants to choose between the current account-manager follow-up (A) and the same follow-up plus an approved recipient-choice invitation (B) for a defined set of renewing customer accounts. The team specifies one account as the allocation unit to avoid contacts at the same account receiving inconsistent treatment. The primary outcome is a renewal conversation held within 30 days of the scheduled invitation. The decision owner is the renewal lead; an analyst owns the frozen account-ID allocation; finance approves the per-account ceiling; privacy and the customer's gift-acceptance policy owners approve the data path and eligibility. The case is illustrative, not a Giftpack customer result or evidence of likely lift.
Before assignment, the lead excludes accounts with an open procurement restriction, a request not to receive gifts, or an unresolved service escalation, using the same rule for both arms. The operations owner prepares a fixed follow-up script and delivery date. Arm B adds only the approved invitation. The analyst keeps account-to-person mappings outside the workbook, creates stable IDs, generates frozen values, and checks uniqueness, assignment completeness, and basic balance. Both groups use the same meeting-event definition from the customer relationship system. A meeting booked but canceled does not count as held; a meeting after day 30 is outside the primary window even if the team may review it separately.
Suppose B has more observed meetings in the window but several invitations were delayed by stock constraints. The lead should not rewrite the outcome window to include those late events and call the trial successful. Log the delays, compare missingness and actual delivery, and ask whether the intended treatment was delivered. If execution failure prevents interpretation, the acceptance evidence is a reproducible roster, arm file, event extract, cost reconciliation, exception log, and an explicit “inconclusive—repair and rerun” decision. An alternative is a staged, nonrandom pilot to test logistics first; it is lower risk for operations but cannot answer the same effect question. A larger randomized test may be justified only after the design and needed sample are reviewed.
Worked decision two: an employee recognition format
Hypothetical case: a people-operations team is deciding whether a fixed approved recognition item (A) or an equally budgeted choice set (B) yields a better recipient experience across offices. The result to choose is a voluntary, short post-delivery response within 21 days, using the same question and invitation timing in both arms. A satisfaction score is not automatically a binary outcome; to use this workbook, the team must specify a defensible yes/no event in advance, or use another analysis tool designed for continuous or ordinal data. The team may choose response completion as its primary binary outcome and treat sentiment as a separately reviewed secondary measure.
Eligibility covers an approved recognition occasion and comparable employment groups. The employer-policy owner checks whether some recipients are barred from gifts, whether local value or tax handling differs, and whether a manager's choice could create fairness concerns. A privacy owner determines whether voluntary feedback can be linked to a pseudonymous ID and how long the link is retained. The analyst considers site-level allocation if coworkers' discussions could contaminate individual assignment; that choice changes the effective unit and may require a different statistical plan. Operations confirms that both formats can be delivered within the same period and records substitutions rather than quietly replacing B items with A.
Imagine the choice arm costs more in support time and has lower survey completion, while qualitative comments from respondents are positive. The narrow, predefined primary result would not justify a statement that “employees prefer choice.” Nonresponse may differ across arms, and the survey cannot speak for silent recipients. The team can decide to improve the invitation, test fulfillment first, or gather properly consented feedback before another controlled comparison. Acceptance evidence includes the policy sign-off, eligibility log, frozen assignment, delivery timestamps, approved survey instrument, response extract, support cost, and a memo that distinguishes observation from interpretation. The employee-specific recognition KPI framework can inform downstream monitoring, but should not be substituted for this preregistration.
Interpret the rate difference without overselling it
The workbook calculates each arm's positive in-window outcomes divided by that arm's observed in-window outcomes, then subtracts A's rate from B's. If either denominator is zero, the rate remains blank. A displayed difference is a description of records entered under these rules. It is not a p-value, confidence interval, power result, causal effect, or forecast for the next campaign. The workbook also does not adjust for clusters, stratification, repeated contacts, unequal exposure, multiple outcomes, or late data. The analyst must decide whether those issues require an expanded study design and formal analysis.
Review the denominator as closely as the percentage. Are assigned people missing because their invitation never reached them, because the event source failed, because they opted out, or because the window is incomplete? The answers have different implications. Count allocation by arm, observation completeness by arm, duplicate IDs, and budget variance. Investigate the exact records under access controls without putting identities in the shared workbook. If the intervention raises complaints, policy exceptions, or delivery burden, those are decision-critical even when the observed primary rate is favorable.
Use a decision memo with four headings: what was preregistered; what was actually delivered; what was observed; and what remains uncertain. State whether the recommended next action is scale, redesign, repeat, or stop, plus who can authorize it. Where statistical certainty matters, hand the frozen protocol and de-identified extracts to a qualified analyst for sample-size planning and inference. Do not derive a significance claim from a dashboard color or from a visually large percentage gap.
Why not use a before-and-after comparison instead?
A before-and-after comparison can be useful operationally when randomization is impossible, but seasonality, audience mix and other concurrent changes can explain the difference. Label it observational, document those changes, and avoid calling its result a randomized estimate. The governance decision may still be to run a smaller controlled test later.
Recovery rules when the experiment breaks
Treat deviations as evidence, not as rows to hide. Duplicate pseudonymous IDs require the analyst to reconcile the source roster and preserve the original extract. Eligible records with no arm mean allocation is incomplete; do not launch B while those records remain unresolved. Random values outside the permitted range must be regenerated under a documented allocation procedure, not manually clipped to fit the desired group. An observation end before the start is an invalid protocol. Outcomes outside the window remain in the exception log but do not enter the primary rate. A missing event can reflect failed instrumentation, not a negative outcome.
If gift inventory changes after assignment, the operations owner should pause the affected arm, record which units were exposed to which version, and take the issue to the decision owner. An approved substitution may preserve the practical purpose of a campaign while changing the treatment enough to undermine the test. The memo must distinguish operational completion from experimental validity. Likewise, if a privacy notice, employer policy, or recipient preference blocks a send, honor it; do not treat the experimental plan as permission to override the rule. Seek the relevant owner’s revised approval before continuing.
Use the QA sheet as an early-warning screen. It checks duplicate IDs, eligible records without an arm, assigned records without an in-window outcome, and observation dates. A zero in a check is not a complete audit. Inspect sample rows, event joins, date zones, actual spending, and exclusions; retain the input version and a signed review note. If formulas or row range change, create a new version and rerun blank, zero, duplicate, boundary, and out-of-window tests. V1.0 has 100 recipient rows, 5–104, and intentionally offers no automatic random-number generator. This limitation helps keep an external allocation auditable but requires a competent analyst to do that step.
Conclusion: use the template as a reproducible starting record
A good gifting trial is a decision with a frozen audience, treatment, measure, window, cost rule, safeguards, and recovery route. Download the workbook, fill Hypothesis before Eligible Population, have an analyst freeze Assignment, reconcile Outcomes and Costs, then ask the decision owner to sign a memo that states what the descriptive figures can and cannot support. Keep the template, source extracts and exception log together under approved access and retention rules. Refresh the method and file version when the study design, formulas, or local requirements change; do not silently edit an old decision record.
Where an organization has already approved the test design and recipient policy, Giftpack's corporate gifting team can discuss execution of an approved choice set, per-recipient limits and country-specific fulfillment. Giftpack does not replace the organization’s analyst, privacy counsel, employer-policy owner, tax adviser, or final business decision maker.