Skip to content

Privacy and preparation

Expert decision counterfactuals: how to capture what would have changed the call

By SourceX Editorial · Updated

10 sections

Short answer

To capture evidence that would have changed an expert's decision, pick the factors the expert says mattered, build variants that change one factor to a value your own records actually contain, and ask the expert to decide each variant blind. Record where the decision flips, where either path is defensible and how confident the expert is, then check each stated boundary against how real past cases were decided.

Call to transcript

Expert decision

This page covers expert decision counterfactuals: how to capture what would have changed the call.Illustrative

Key takeaways

  • A counterfactual is the smallest real change to the facts that would have produced a different decision; asking for it directly gives vague answers.
  • Ground every variant in values that already occur in the company's records, and change one factor at a time.
  • Have the expert decide variants blind and out of order, so the boundary comes from decisions rather than a story about them.
  • Record a zone where either path is defensible instead of forcing a single threshold.
  • A stated boundary is only kept once it agrees with how comparable past cases were actually decided.

How do you capture evidence that would have changed an expert's decision?

You capture it by testing the decision, not by asking about it. Take one real decision, list the factors the expert relied on, and present variants in which a single factor changes to a value the business has really seen. The expert decides each variant without knowing which one is the original, and you record where the answer flips.

The output is a set of contrast pairs: two cases that differ in one documented factor and received different decisions. Each pair carries the expert's confidence, a note on whether the other path was also defensible, and a pointer to historical records that confirm the boundary. That is far more useful than the sentence an expert produces when asked, weeks later, what might have changed their mind.

This page sets out the protocol, the record each pair should carry, a worked example and the situations where the method should not be used. It assumes the company has a decision type with enough history to supply real values, which is the condition that makes the counterfactuals grounded rather than imagined.

Why a what-if question must be anchored in records the business already holds

An unanchored what-if invites invention. Ask a claims supervisor what would have changed a call and you will hear about factors that sound important, values that never occur in practice, and thresholds that are rounder than the ones the supervisor actually applies.

The concept itself is well defined. Wachter, Mittelstadt and Russell described counterfactual explanations as statements of how the facts would need to differ for a different decision to follow, and argued they can be offered without opening the system that made the decision. The same idea applies to a person: the useful answer is a specific, minimal change, not a general account of what matters.

Outcome knowledge is the second reason to anchor. Hindsight bias research going back to Fischhoff in 1975 found that people judge a known outcome as more predictable than it was, even when told to ignore it. An expert who knows the order was later disputed will describe a stricter boundary than the one they used at the time. Grounding variants in real values and checking them against real past decisions keeps the boundary tied to practice.

Which factors should vary, and which must stay fixed?

Vary only the factors that plausibly decided the outcome, and hold everything else constant. Start from the expert's own list of what they looked at, add any factor the policy names, and drop factors that only describe the case without bearing on it, such as a customer's region when the policy is national.

Kaushik, Hovy and Lipton used the same discipline when they asked people to rewrite documents so a different label would apply: revisions had to stay internally consistent and avoid gratuitous changes to facts unrelated to the label. A variant that changes two factors at once cannot tell you which one moved the decision, and a variant that also edits incidental details teaches a model to watch the wrong signals.

Some factors cannot move alone. Order value and quantity are linked; tenure and lifetime order count are linked. Treat such pairs as one compound factor and say so in the record, rather than creating a variant the business could never produce.

  • Decisive: the expert names it unprompted and the policy references it. Always vary.
  • Supporting: the expert mentions it when asked. Vary once, at the most common alternative value.
  • Incidental: present in the record but not used. Hold fixed; never edit.
  • Linked: cannot change without another factor changing. Vary together as one compound factor.

The grounded counterfactual elicitation protocol

Protocol version 1.0, dated 2026-10-18, runs per decision type, not per case, and assumes an interviewer who knows the records but did not make the decisions. The opening step borrows from the critical decision method, an interview technique that extends the critical incident technique with probes for the cues and distinctions experts use. The later steps are specific to counterfactual work. The table sets out the nine steps, and the list gives the fields each verified contrast pair carries once they are done.

At a glance

The grounded counterfactual elicitation protocol

  • Pair ID and version of the protocol used
  • Base case reference and variant reference, both pointing to record IDs or a variant spec
  • Factor changed, original value and variant value, with the source query for the value
  • Decision on each side, confidence on each side, and whether the other path was defensible
  • Indifference zone, where one exists, as a lower and upper value
  • History check result: count of past cases agreeing and disagreeing, with case references
  • Expert role, tenure band and authority level; no names
  • Reviewer notes and any disagreement between experts
The grounded counterfactual elicitation protocol
StepWhat you doWhat you recordGrounding check
1. Choose anchor casesPick five to ten recent decisions of one type, including some near the policy lineCase IDs, decision, decision date, decider roleEach case exists in the system of record with its original fields
2. Walk one caseAsk the expert to narrate what they saw and noticed, with the outcome hiddenFactors named, in the order namedEvery named factor maps to a field, document or message
3. Classify factorsSort factors into decisive, supporting, incidental and linkedFactor list with class and field referencePolicy text cited where a factor is policy-driven
4. Find real valuesQuery the records for the observed range of each decisive factorMinimum, typical and edge values with countsValues come from records, never from the interviewer
5. Build variantsChange one factor per variant to an observed valueVariant ID, base case, factor changed, new valueA second reviewer confirms the variant is internally consistent
6. Decide blindPresent variants and the original shuffled, with no outcome shownDecision, confidence from one to five, time takenOriginal case included so its re-decision can be compared
7. Probe the flipFor each flip, narrow the value until the decision changesFlip point, indifference zone, stated reasonRepeat one variant later in the session to test consistency
8. Check historyRetrieve past cases on both sides of each stated boundaryCounts agreeing and disagreeing with the boundaryDisagreements reviewed with the expert, not deleted
9. Second expertRun steps 6 and 7 with another qualified decider on a sampleAgreement on flip points and zonesDisagreement kept as data with a short note

How do you build a variant the expert has never seen without inventing a business?

Start from a real case and change one field to a value that appears elsewhere in the same records. If the original return arrived twelve days after the window closed, and the records show late returns between one and sixty days, the variants might use three, twenty-five and fifty. Every value then has a source query behind it, and a reviewer can reproduce it.

Keep the rest of the case identical, including the notes and messages, unless a field changes as a direct consequence of the edit. If you raise the order quantity, the line total changes too; if you change the customer's tenure, a note that says new account must change with it. A second reviewer checks each variant for internal consistency before the session.

Mix in two kinds of foils. Include the original case, unlabelled, so you can see whether the expert re-decides it the same way. Include one variant that changes only an incidental factor, which should never flip the decision. If it does, the expert is reading something other than the factors you listed, and the factor classification needs another pass.

Present variants the way the expert normally sees cases. If decisions are made from a screen in the order management system, show a mock of that screen with the same fields in the same places, not a paragraph summary. A summary already selects and orders the facts, which quietly tells the expert what the interviewer thinks matters.

When is another path equally valid?

Another path is equally valid when two qualified experts, or the same expert on different days, would defend either decision for the same facts. Treat that range as an indifference zone with a lower and upper value, not as noise around a single threshold.

Experts often feel pressure to give a crisp number. Resist it. EFSA's guidance on expert knowledge elicitation notes that unaided expert judgement, particularly about uncertainty, is often biased, which is why it recommends structured methods. A forced threshold reported to two significant figures overstates what the expert knows and teaches a model a false certainty.

Ask two questions at each flip point. First, at what value would you be uncomfortable with either answer? Second, would a colleague you respect have decided the other way here? The first marks the edges of the zone; the second tells you whether the zone reflects personal style or genuine ambiguity in the policy. Record both answers, and record disagreements between experts with a short note on what each one weighed differently.

Checking the stated boundary against how past cases were actually decided

A boundary is kept only if the history broadly agrees with it. For each flip point, retrieve past cases on both sides that match the base case on the factors held fixed, and count how many were decided the way the boundary predicts.

Gardner and colleagues used small, label-changing edits to build contrast sets that test a model's local decision boundary, and reported that model performance on them fell well below performance on the original test sets. Your contrast pairs play the same role for an evaluation set, which is why their boundary must reflect practice and not the expert's tidier recollection.

Disagreement is information. If the expert says returns more than thirty days late are refused, but a third of such past returns were accepted, review those cases with the expert. Often a factor was missed, for example a standing agreement with a key account. Add it to the factor list, rerun the affected variants, and keep the original disagreement in the record so a later reviewer can see how the boundary was refined.

Match past cases carefully. A case only counts as evidence for a boundary if it shares the base case's values on every decisive factor except the one being tested, within a tolerance written down before the count. Loose matching inflates agreement, because cases that differ on several factors at once can land on the expected side for the wrong reason.

Worked example (illustrative): an out-of-window return at a parts distributor

This case is invented to show the protocol end to end. A distributor's policy accepts returns of unopened stock items within thirty days. A customer service manager accepted a return that arrived twelve days after the window closed. Walking the case, the manager named days past the window, item condition, whether the item is a stock or special-order part, and the account's return history.

The records showed late returns ranging from one to sixty days, so variants used values from that range. The manager decided the shuffled set blind, including the original and one incidental variant that changed only the shipping carrier.

  • Stored boundary: days past window flips between roughly 20 and 35 days, recorded as an indifference zone rather than one number.
  • Stored pairs: Original with V2, V3 and V4 as clean contrast pairs; V1 kept as an ambiguous case, not a pair.
  • Open item: V5 sent to a second manager, because the history did not support a single answer.
Worked example (illustrative): an out-of-window return at a parts distributor
VariantFactor changedValue (illustrative)Decision and confidenceHistory check (illustrative)
OriginalNone12 days late, unopened, stock itemAccept, 4 of 5Re-decided the same way
V1Days past window25 days lateAccept, 2 of 5; either path defensiblePast cases split roughly evenly
V2Days past window50 days lateRefuse, 5 of 5Most past cases refused
V3Item conditionOpened packagingRefuse, 4 of 5Most past cases refused
V4Part typeSpecial-order partRefuse, 5 of 5Nearly all past cases refused
V5Return historyThree late returns in the past yearRefuse, 3 of 5Mixed; flagged for second expert
V6 (foil)Shipping carrierDifferent carrierAccept, 4 of 5Not applicable; no flip expected

Where counterfactual elicitation gives the wrong answer

The method fits decisions with a handful of decisive factors and enough history to supply real values. Narrow the scope or use a different method when those conditions fail.

Decisions that turn on many interacting factors, such as a full underwriting review, produce too many variants to decide in one sitting and too few comparable past cases to check them. Decide a smaller sub-question instead, such as whether to refer the file, and vary only its factors. Decisions dominated by information outside the record, such as a phone call, will produce variants the expert cannot judge; label them as low-grounding and keep them out of evaluation sets.

Regulated decisions, including consumer credit, employment and anything touching health information, need counsel and the compliance owner involved before variants are built, because a variant set can reveal how protected or sensitive attributes affect outcomes. Very small decision histories cannot support the history check at all; record the boundary as the expert's stated view and mark it unverified. Finally, stop if the expert re-decides the original case differently from the recorded decision more than occasionally. That suggests the factor list or the case selection needs another pass, not more variants.

Contrast pairs as a record a company could later license

Verified contrast pairs are a compact form of judgment that rarely exists anywhere else: real cases, a single documented change and the decision on each side. Teams can use them internally to test policies and train staff, whether or not they ever leave the company.

Should a business later want to license decision records of this kind, SourceX takes care of finding and qualifying them, reviewing rights, contracting, preparing and delivering. The conversation opens with a few questions about the company rather than any upload, and no record is delivered until the company has signed an agreement approving the scope and terms, so running this protocol for internal use commits it to nothing. Variants built from customer records still carry the privacy and contract obligations of those records.

Frequently asked questions

What is an expert decision counterfactual?

It is the smallest realistic change to the facts of a case that would have led the expert to a different decision. A good one names one factor and one value, such as the same return arriving fifty days late instead of twelve. Paired with the original, it shows where the expert's decision boundary sits.

Why not just ask the expert what would have changed their mind?

Direct questions produce general answers and round thresholds, and they are shaped by knowing how the case turned out. Asking the expert to decide concrete variants blind gives decisions you can observe and check. The direct question is still useful as a starting list of factors to vary.

How many variants should one session include?

Enough to cover each decisive factor at two or three observed values, plus the original and one foil. For a decision with three or four decisive factors that usually means ten to fifteen variants, which is a planning figure. More than that tends to tire the expert and lower consistency.

Can synthetic variants be used as training data?

They can, if they are labelled as constructed and each one carries its base case, the single change made and the source of the new value. Keep them separate from unedited historical records so a buyer or model team can decide how to weight them. Variants that failed the history check should be excluded or marked.

What if two experts place the boundary in different places?

Keep both boundaries and a short note on what each expert weighed differently. Where the disagreement reflects genuine ambiguity in the policy, record an indifference zone that covers both. Forcing one answer hides the cases a model most needs to handle carefully.

Sources

  • Wachter, Mittelstadt and Russell describe counterfactual explanations of automated decisions and argue they can be offered without opening the decision system. Source
  • In Kaushik, Hovy and Lipton's counterfactually augmented data, people revised documents to a counterfactual target label while keeping them internally consistent and avoiding gratuitous changes unrelated to the label. Source
  • Gardner and colleagues build contrast sets from small, meaningful edits that typically change the gold label, to give a local view of a model's decision boundary, and report much lower model performance on them. Source
  • The critical decision method extends the critical incident technique with probes for the cues, discriminations and typicality judgments experts use. Source
  • EFSA's guidance on expert knowledge elicitation states that unaided expert judgement, particularly of uncertainty, is often biased and sets out structured elicitation methods. Source
  • Peer-reviewed research describes hindsight bias, first shown by Fischhoff in 1975: people overestimate how predictable a known outcome was, even when trying to ignore it. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies