Skip to content

Systems and records

Sales workflow dataset buyer acceptance: look past closed-won rows

By SourceX Editorial · Updated

10 sections

Short answer

Evaluate revenue workflow data against every opportunity in the period, not only the wins. A sales workflow dataset passes buyer acceptance when won, lost, stalled and disqualified deals appear in the source CRM's proportions, each deal can be rebuilt as it looked on the decision date from history tables, and the CRM's story is reconciled with signed contracts, approvals and post-sale outcomes. Then judge fit against one named task.

Checklist

Sales workflow dataset buyer

This page covers sales workflow dataset buyer acceptance: look past closed-won rows.Illustrative

Key takeaways

  • Closed-won rows are a filtered survivor set, so a dataset made mostly of them teaches what a finished win looks like rather than how revenue decisions are made.
  • CRMs rewrite end states: HubSpot can reset close dates when a deal closes, and Salesforce keeps stage history in a separate object from current values.
  • The acceptance guide tests outcome mix, period and segment coverage, as-of reconstruction, contract and approval trails, post-sale outcomes, leakage and a documented sampling rule.
  • Utility is task-specific: the same export can fail for deal risk scoring and pass for a renewal model once it is joined to later outcomes.
  • When a check fails for one period or segment, narrow the offer and document why instead of padding it.

How do you evaluate a sales workflow dataset without overvaluing closed-won rows?

Evaluate it against the full population of revenue decisions the data is supposed to teach, not against the deals that happened to close. A sales workflow dataset passes buyer acceptance when won, lost, stalled and disqualified opportunities appear in proportions you can explain, when each deal can be replayed as it looked on the day a decision was made, and when the CRM's story is checked against signed contracts, approvals and what the customer did afterward.

Closed-won rows are the most complete, most celebrated and most often exported part of a pipeline, which is exactly why they mislead. They carry the final amount, the final stage and a tidy close date, but they hide the decisions that went the other way and the edits made after the fact. A model trained or tested only on them learns what a finished win looks like, not how a revenue team decides under uncertainty.

This guide gives a representativeness and task utility acceptance guide a revenue operations lead can run before a dataset is offered for licensing, plus the contradiction checks, a task fit matrix and an illustrative review with one exception that changes the verdict.

Why a pipeline export of finished wins overstates what the data can teach

A won deal is the end of a filtered path. Every opportunity that was lost, went quiet, was merged into another record or was disqualified at discovery left the population before the win was recorded. If the export keeps only closed-won rows, any pattern a model finds, such as larger deals having more stakeholder meetings, may simply describe the survivors.

The CRM also rewrites the end state. HubSpot's documentation lists Closed won with a fixed 100 percent deal probability and Closed lost at 0 percent, and its close date automation sets a deal's close date to the current date when the deal moves into a closed stage, unless a team turns that off. A close date in the export can therefore be the day someone tidied the pipeline, not the day the customer signed.

Salesforce separates the current record from its past. The Opportunity object holds today's values, while the OpportunityHistory object represents the stage history of an opportunity and OpportunityFieldHistory holds changes to tracked field values. A dataset built from the Opportunity table alone shows each deal frozen at its last edit, which is the least useful moment for teaching forecasting, risk or next-step decisions.

Which records rebuild a revenue decision as it was made?

A revenue decision can be rebuilt only when the stage and amount at that moment, the authority behind it and the outcome that followed are all linked to one account and one opportunity. The table lists the evidence a reviewer should expect and the system where it usually lives.

Which records rebuild a revenue decision as it was made?
EvidenceTypical systemWhat it provesCommon gap
Account and legal entity IDsCRM account, ERP customer masterWhich customer a deal, contract and invoice belong toParent and subsidiary accounts merged or duplicated
Stage, amount and close date historySalesforce OpportunityHistory and field history; HubSpot deal property historyWhat the team believed on each dateHistory tracking switched on late or for few fields
Signed order form, contract and amendmentsContract repository or e-signature platformThe terms the customer actually acceptedAmendments stored outside the CRM and never linked
Discount and deal desk approvalsCRM approval records, deal desk queue, emailWho had authority to accept the termsApprovals given in chat with no record ID
Forecast category and commit notesCRM forecast fields, weekly forecast snapshotsWhy the deal was called commit, best case or pipelineOnly the final forecast snapshot kept
Delivery, invoice and renewal outcomeERP, billing system, customer success platformWhether the booked deal turned into revenue and stayedBilling joined on a name instead of an ID
Source and attribution fieldsMarketing automation, CRM campaign influenceWhere the deal came from, with stated limitsAttribution model changed mid-period without a note

The representativeness and task utility acceptance guide

Run each check on the candidate dataset and record the evidence file next to the result. A check either passes, narrows the scope to the part that passes, or blocks the dataset until it is fixed. Thresholds are set per deal with the buyer; the guide fixes what to measure and what each outcome means.

At a glance

The representativeness and task utility acceptance guide

The representativeness and task utility acceptance guide
CheckHow to test itAccept whenNarrow or block when
Outcome mixCount opportunities by final status: won, lost, open, disqualified, mergedEvery status the task needs is present and the ratios match the source CRM for the same periodWon rows are overrepresented against the source, or lost and disqualified rows were filtered out
Period coveragePlot created and closed counts by month across the whole rangeNo unexplained gaps; process or pricing changes are dated in the notesA migration, reorganization or field change splits the history and is not documented
Segment coverageCross-tab outcome by segment, region, product line and deal size bandEach segment the buyer cares about has wins and lossesA segment exists only as wins, so loss patterns cannot be learned there
As-of reconstructionRebuild a random sample of deals at a past date from history tablesStage, amount and close date at that date match the history logOnly current values exist, or history starts after most deals closed
Contract reconciliationMatch a sample of won deals to signed documents by IDAmount, term and products agree, or differences are explainedMany booked amounts differ from signed amounts with no amendment
Authority trailCheck discounts and nonstandard terms against approval recordsApprovals carry an approver, date and record linkApprovals are missing for terms outside policy
Post-sale outcomeJoin won deals to invoices, delivery and renewal by customer IDMost won deals link to billing and a later statusThe join fails or relies on fuzzy name matches
Leakage screenList every field and the date it becomes knownNo feature used for prediction is set after the decision dateFields like final amount or closed reason are present at prediction time
Documented samplingWrite the selection rule and the counts it removedThe rule can be rerun and explains every exclusionRows were hand-picked or the rule cannot be stated

Which signed term or later outcome contradicts what the CRM says?

The CRM is the team's account of a deal, and it is wrong in predictable places. Check these contradictions on a sample before trusting any aggregate, because each one changes what the row teaches a model.

  • Booked amount versus signed amount: the opportunity says one figure, the order form or a later amendment says another. Label which value a task should learn from.
  • Close date versus signature date: the stage moved to Closed won during a cleanup weeks after signature, or before the counterparty signed. Use the signature event as the decision time when the task depends on timing.
  • Won, then reversed: the deal closed, then the customer cancelled inside a trial period, never paid, or churned at first renewal. Mark these so a model does not treat them as clean wins.
  • Approved outside the system: a discount or payment term that the deal desk approved by email appears in the CRM as standard. The approval record, not the field value, shows who decided.
  • Lost that was really merged: a closed-lost row exists because a rep created a duplicate and closed one copy. Treat merges as their own status, not as losses.
  • Attribution rewritten: the source field changed after a reporting model update. Keep the value as of the decision date and record the model in use.

Does the dataset fit the task a buyer will train or test?

Utility depends on the task, so state the target task before judging the data. A dataset that is weak for win prediction can be strong for a deal desk approval assistant, and the reverse. The matrix below shows what each common revenue task needs and what goes wrong if the export is mostly closed-won rows.

Does the dataset fit the task a buyer will train or test?
Target taskRows it needsHistory it needsWhat a closed-won export breaks
Win likelihood or deal risk scoringWon, lost and stalled deals in realistic proportionsStage and amount as of each scoring dateEvery example is a success, so the model cannot separate risk from noise
Forecast commit reviewDeals in every forecast category, including slipped onesWeekly forecast snapshots and close date changesSlips and pushes disappear, so forecast error cannot be measured
Next-step or follow-up agentOpen and closed deals with activity and outcomeActivity timeline, emails and meetings in orderThe agent sees only paths that worked and copies them
Discount and term approval assistantDeals with standard and nonstandard terms, approved and rejectedApproval requests, decisions and approver roleRejected requests are missing, so the assistant approves everything
Renewal and churn predictionWon deals joined to renewal and cancellation outcomesUsage or billing status after signatureWon rows without later outcomes teach nothing about retention

Illustrative: a software company's pipeline export fails one check

The pipeline below is a made-up case, and every count and share in it is illustrative. A mid-market software company prepares four years of opportunities from its CRM for a deal risk task. The first export contains 9,000 opportunities, 61 percent of them closed-won, because the sales operations team started from its commissions review report, which keeps every closed-won deal but only the lost deals that reps logged with a loss reason.

The source CRM for the same period shows about 40,000 opportunities, with roughly 18 percent won. The outcome mix check fails. The team reruns the export with a written rule (every opportunity created in the period, excluding test records and merged duplicates) and documents that 1,100 merged duplicates were removed. The new mix matches the source within a percentage point.

The as-of reconstruction check then passes for the last two years and fails for the first two, because field history tracking on amount and close date was switched on midway. The team narrows the offer: the full outcome mix for all four years, but as-of features only for the later two, stated in the data dictionary.

The exception: one product line was sold almost entirely through a reseller, and its opportunities exist only as won rows entered after the reseller reported a sale. For that line, losses were never recorded. The guide's segment check flags it, and the team excludes the line from the risk task rather than letting it inflate the win rate.

Keeping evaluation honest: time splits, history windows and leakage

Split any revenue dataset by time, not at random. The scikit-learn documentation notes that standard shuffled cross-validation assumes independent samples and gives poor estimates on time series data, and recommends evaluating on future observations, which its TimeSeriesSplit class does. A deal risk model tested on a random split can learn quarter-end patterns from the same quarter it is graded on.

Leakage is the other common failure. Kapoor and Narayanan's survey of machine-learning-based science describes leakage as a widespread source of overoptimistic results and includes temporal leakage and illegitimate features in its taxonomy. In revenue data, the usual culprits are the closed reason, the final amount and any field filled at or after close.

Check how far back history actually goes before promising as-of features. Salesforce help states that without Field Audit Trail, field history is retained for up to 18 months in the application and up to 24 months through the API. Teams that never archived it may have current values for old deals but no record of how they got there.

Which failed checks narrow a revenue dataset, and which need a contracts reviewer?

Narrow the dataset when a check fails for a definable part of it: a period before history tracking, a segment without losses, a product sold through partners. Narrowing keeps the rows that can be trusted and records why the rest were excluded, which the Datasheets for Datasets questions ask a dataset's authors to explain when a sample is not representative of the larger set.

Stop and request specialist review when the evidence touches contracts, personal data or regulated customers. Contract terms, approval emails and call notes carry names, prices and sometimes confidential customer information. A revenue operations lead can judge whether the rows are complete; a contracts or privacy reviewer must decide what can be shared and in what form.

Bring in a statistician or the buyer's evaluation lead when the task depends on calibrated probabilities. The acceptance guide shows whether the data is representative and usable; it does not set the sample size needed to detect a given improvement.

Offering a revenue history for AI licensing

Companies that pass these checks often hold the kind of multi-year pipeline history that AI developers ask for. A revenue history like this can be licensed through SourceX, which takes it through qualification, rights review, contracting, preparation and delivery. As a sizing rule of thumb, buyers want two or more years of consistent pipeline history and tens of thousands of opportunities in one CRM, although rarer work can qualify with less.

A revenue team can find out whether its history qualifies through a brief company assessment, without uploading any records. Records are delivered only after the company has approved the scope and terms and signed.

Frequently asked questions

Is closed-won data worthless for AI?

No. Won deals are useful for tasks like renewal prediction or contract term extraction when they are joined to later outcomes. The problem is using them alone for any task that needs to tell good decisions from bad ones, where the missing losses make the data one-sided.

What win rate should a representative pipeline dataset show?

It should show the win rate of the source CRM for the same period and selection rule, not an industry benchmark. Record the source count, the export count and every exclusion so a buyer can see the ratio was not changed by filtering.

Do we need field history to license CRM data?

Not for every task, but as-of reconstruction depends on it. Without stage and amount history you can still offer outcome-level data, such as which deals were won or lost, while stating that features reflect final values rather than values at decision time.

How do I handle deals that were won and then cancelled?

Keep them and add an explicit post-sale status such as cancelled in trial, unpaid or churned at first renewal. Removing them makes the won population look healthier than it was, and relabelling them as lost hides the fact that the sale closed.

Who should sign off on a revenue dataset before it is offered?

A revenue operations lead should confirm completeness and the acceptance results, and a contracts or privacy reviewer should confirm what customer and personal information can be shared. Where the data includes regulated customers, counsel reviews the specific facts.

Sources

  • Salesforce's OpportunityHistory object represents the stage history of an opportunity. Source
  • Salesforce's OpportunityFieldHistory object records the history of changes to the values in an opportunity's tracked fields. Source
  • Without Field Audit Trail, Salesforce retains field history for up to 18 months, and up to 24 months through the API. Source
  • HubSpot's default sales pipeline sets Closed won at 100 percent and Closed lost at 0 percent deal probability. Source
  • HubSpot can set a deal's close date to the current date when it moves into a closed stage; teams can turn the automation off. Source
  • Shuffled cross-validation assumes independent samples and gives poor estimates on time series data; TimeSeriesSplit evaluates on future observations. Source
  • Kapoor and Narayanan document data leakage as a cause of overoptimistic results in ML-based science, with a taxonomy that includes temporal leakage and illegitimate features. Source
  • Datasheets for Datasets asks whether a sampled dataset is representative of the larger set and, if not, why not. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies