Counting cancelled and returned work in an outsourcing pilot

Headline statistic

One declared intake cohort, one observation unit, and zero universal performance promises.

Methodology: Research question: Which requested items belong in the denominator when a delegated research pilot is evaluated? This article addresses whether the apparent completion rate reflects usable delegated capacity or only the easiest surviving assignments. The observation unit is one request that entered the agreed pilot intake, including requests later cancelled, merged, returned, paused, or ruled out of scope. It is a prospective local evaluation design for a buyer of research-assistant support, not a claim that OutsourcingAssistant.com has measured client outcomes. No productivity rate, cost saving, staffing ratio, hiring guarantee, or location-based advantage is asserted. Method: define eligibility and fields before collection; retain every disposition; compare intake cohorts tracked to a named disposition rather than a published-only or completed-only list; have the accountable owner review exceptions and interpretation. Evidence basis: five primary or official methodology sources, checked September 18, 2026. Inference boundary: A complete disposition table can reveal selection and workflow loss. It does not make every cancellation a defect or show what would have happened if a cancelled request continued.

Key stats

• Observation unit: one request that entered the agreed pilot intake, including requests later cancelled, merged, returned, paused, or ruled out of scope.
• Evidence base: five named primary or official methodology sources.
• Publication standard: report counts, exclusions, missing fields, changes, and uncertainty with the result.

Key takeaways

• A complete disposition table can reveal selection and workflow loss. It does not make every cancellation a defect or show what would have happened if a cancelled request continued.
• Report the full cohort and reasons beside the usable-output count, then change scope only through a dated owner decision.
• The minimum case record is: intake identifier, eligibility decision, scope class, disposition, disposition reason, decision owner, dates, evidence state, and whether the item consumed preparation or review time.

The buyer decision and the claim boundary

Which requested items belong in the denominator when a delegated research pilot is evaluated?

This article addresses whether the apparent completion rate reflects usable delegated capacity or only the easiest surviving assignments. The observation unit is one request that entered the agreed pilot intake, including requests later cancelled, merged, returned, paused, or ruled out of scope. It is a prospective local evaluation design for a buyer of research-assistant support, not a claim that OutsourcingAssistant.com has measured client outcomes. No productivity rate, cost saving, staffing ratio, hiring guarantee, or location-based advantage is asserted.

The UK Government Service Manual begins measurement with the service outcome and the information needed to improve it. Applied here, that means the buyer should write the decision before selecting a metric. A result that will determine whether the apparent completion rate reflects usable delegated capacity or only the easiest surviving assignments needs a population, observation unit, comparison, owner, and decision rule that are visible before anyone sees a favourable or unfavourable number.

A Philippines-based research assistant may prepare records, apply a declared codebook, and assemble an exception packet. The accountable buyer still decides the scope, judges consequential exceptions, and approves any staffing or process change. Geography is part of the operating context; it is not evidence of quality, speed, or causation.

Define the eligible population before observing results

Eligibility starts when work enters the agreed intake, not when a polished output appears. Define which requests qualify, the start and end of the review period, how reopened or merged work is treated, and which exclusions are permitted. Retain an identifier and final disposition for every eligible item. This prevents the denominator from silently improving as difficult cases disappear.

The intended comparison is intake cohorts tracked to a named disposition rather than a published-only or completed-only list. Like-for-like does not mean pretending all briefs are identical. It means retaining the factors that could reasonably alter the result, then reporting where comparison is weak. If a class has only a few observations, publish the count and individual pattern instead of a confident percentage.

GAO data-reliability guidance asks whether information is sufficiently reliable for its intended purpose. That is a better test than asking whether the records look complete. Missing owner timestamps may be acceptable for a topic inventory but fatal to a review-delay estimate. Reliability must be decided field by field against the buyer decision.

Build a case record that survives handoff

For each unit, retain intake identifier, eligibility decision, scope class, disposition, disposition reason, decision owner, dates, evidence state, and whether the item consumed preparation or review time. Use system events where they exist and label self-reported times. Preserve the original value when a correction is made, record who made the correction, and explain why. An empty field means unknown; it must not be converted into zero, success, or “not applicable” without evidence.

A research assistant can prepare this record without deciding its meaning. The assistant should link the supporting event, flag a conflict, and stop when a field requires an owner judgment. That separation makes the packet reviewable and reduces the chance that an operational guess becomes a public claim.

Use a short codebook. Define status, readiness, return, cancellation, active work, waiting, and approval in observable terms. Include one positive and one negative example for fields likely to be confused. Date each codebook version so a later definition change can be separated from a real workflow change.

Run the comparison without erasing variation

The NIST statistical handbook distinguishes process inputs from outputs and treats designed collection as a way to learn about their relationship. Here, the output should match the decision: readiness at first handoff, owner rework, elapsed waiting, usable disposition, or another explicitly defined state. Do not combine them into a single quality score merely because one number is easier to present.

Known nuisance factors should be retained or blocked where practical. Topic consequence, requested depth, source volatility, reviewer availability, new templates, and tool outages can all move the result. NIST's discussion of randomized block designs provides the transferable principle: compare within meaningful groups when a known source of variation would otherwise obscure the question. This article does not claim that a formal experiment is always feasible.

The most serious distortion for this question is dropping difficult requests after intake, treating owner cancellation as assistant failure, or counting a merged request twice while reporting only successful outputs. The repair is to show the full flow from eligibility to disposition, preserve the relevant context, and state where records are not comparable. A transparent “cannot determine” is decision-grade when the alternative is false precision.

Separate fact, analysis, inference, and uncertainty

Facts are retained events and field values: a request arrived, a source was attached, a reviewer returned a brief, or a decision occurred at a recorded time. Analysis applies the declared definitions to those records. Inference is the explanation proposed for a pattern. Uncertainty includes missing events, ambiguous states, reviewer disagreement, small counts, and unmeasured changes. Label all four layers.

A complete disposition table can reveal selection and workflow loss. It does not make every cancellation a defect or show what would have happened if a cancelled request continued. This boundary should appear beside the result, not in a detached disclaimer. If the finding changes when one unusual case is removed, show both views and explain why that case belongs or does not belong. If open work has no final time, retain it as open at cutoff rather than treating it as fast, slow, or successful.

Do not turn association into individual evaluation. A delay can arise from missing source access, owner availability, scope change, or a responsible escalation. A returned brief can reveal a weak intake rule rather than weak preparation. Case evidence determines the operating response; a headline metric does not assign fault.

Decision rule, escalation, and a bounded next test

Report the full cohort and reasons beside the usable-output count, then change scope only through a dated owner decision. Write the rule in advance: who reviews the result, what evidence is sufficient, which exceptions require inspection, and what change is allowed. Avoid a rule that automatically expands access or publication authority. Tool permission and repeated task completion do not transfer accountability.

An exception packet should contain the case identifier, blocked decision, relevant records, conflicting interpretations, consequence if wrong, proposed options, and the exact owner response needed. The assistant can continue non-consequential preparation while the decision waits if the written boundary permits it. Silence is not approval.

After the decision, preserve the old method and effective date of the new one. Change one major rule where feasible, collect a fresh eligible cohort, and repeat the same definitions. If several changes are unavoidable, record all of them and narrow the conclusion. This creates a learning loop without pretending a local operational test is a controlled market study.

Limitations and evidence-led conclusion

The cited sources provide general measurement, data-reliability, experimental-design, and governance principles. None evaluates this exact OutsourcingAssistant.com workflow, a specific client, or Philippines-based assistants as a population. Applying those principles to delegated research is analysis, and the transfer may omit factors unique to a buyer's systems, people, or regulated obligations.

Small operational samples are vulnerable to unstable percentages, incomplete event histories, learning effects, seasonal demand, reviewer adaptation, and changes in task mix. Consequential work may require complete review regardless of the routine sampling plan. Legal, employment, privacy, security, and financial decisions require qualified advice outside this operational research design.

The evidence-led conclusion is modest: Report the full cohort and reasons beside the usable-output count, then change scope only through a dated owner decision. The design makes the buyer's reasoning inspectable. It does not guarantee an outcome, establish a price or staffing ratio, or replace direct review of the work.

Sources and checked dates

The five sources below were checked on September 18, 2026. They were selected because they are primary or official publications and because each supports a specific part of the method. GAO supports intended-use reliability checks; the UK Service Manual supports decision-linked measures; the NIST handbook supports process modeling and blocking known variation; and CSF 2.0 supports explicit governance.

Source authority does not make every inference automatic. Readers should open the linked publication, confirm that the relevant guidance remains current, and distinguish the source's own claims from this article's application to outsourcing decisions. If a source changes materially, record the checked date and reassess the affected conclusion before reusing it.

Related Research

Questions people ask