AI uses for records
Who ended the deal? Lost deal AI training labels by actor and timing
By SourceX Editorial · Updated
10 sections
Short answer
A lost deal becomes a useful AI training label only after two facts are recorded: who ended it and when. A timely walk-away is a correct decision, a rival's win is a positioning lesson, and only a seller miss is a failure. In our audit of four CRM defaults, none records who ended a deal, and six of Frappe CRM's eight described loss reasons cannot be classed without that fact.
Agent run
Lost deal AI training labels
Key takeaways
- A loss is an outcome, not a verdict; whether it was a failure depends on who ended the deal and when.
- None of four CRM defaults audited (ERPNext, Frappe CRM, EspoCRM, Dynamics 365) records which side ended a deal.
- Six of Frappe CRM's eight described default loss reasons fit two or more classes until timing, actor or the buyer's alternative is known.
- A seven-class taxonomy separates timely disqualification, competitive loss, no decision, offer gap, buyer-side event, seller miss and administrative close.
- The same lost deal can be a correct example for a qualification agent and a wrong-action example for a seller-assist agent, so labels are set per task.
What should a lost deal teach an AI system?
It depends on why the deal ended and who ended it. A loss is an outcome; calling it a failure is a separate judgment, and a system trained to treat every loss as a mistake learns to chase deals a good seller would drop. The question here is narrow: which lost deals belong in a training or evaluation set as examples of error, which as examples of sound judgment, and which should stay out.
Three kinds of revenue operations AI depend on the answer. A qualification agent decides which opportunities deserve time. A seller-assist agent recommends the next move inside a live deal. A forecasting model predicts which deals will close. The same lost deal can be a correct example for the first, a wrong-action example for the second and a plain negative outcome for the third, so one closed-lost flag cannot serve all three.
Useful evidence is specific: the furthest stage reached and the date the loss was recognized, the side that ended the deal, what the buyer did instead, the commercial terms in force, and who had authority over each decision. The sections below show how little of that default CRM schemas capture, then set out a taxonomy, label rules and a request specification that capture it.
What do four CRM defaults record when a deal ends without a win?
An outcome and, in places, a reason, but never the side that ended the deal. We read the shipped schemas of three open-source CRMs and Microsoft's reference documentation for Dynamics 365 on October 27, 2026. None has a default field recording whether the buyer or the seller ended a deal, and only two record a competitor separately from the reason.
ERPNext separates a declared loss from an abandoned one. Declaring an opportunity Lost requires at least one loss reason from a multi-select list, competitors go in their own multi-select field, and a free-text detail field sits beside them. A second ending, Closed, is applied automatically to opportunities left in Replied status without activity for a configured number of days, 15 by default, and it carries no reason at all.
Frappe CRM, from the same open-source family, ships nine default loss reasons and a single Lost status, logs status changes in a table on the deal, and makes notes mandatory only when the reason is Other. EspoCRM has a Closed Lost stage and no reason field, but it keeps lastStage, the furthest open stage a deal reached; a deal created directly as lost has its lastStage filled in automatically with the lowest-probability open stage. Dynamics 365 offers two lost status reasons, Canceled and Out-Sold, links a competitor on the close record, and lets administrators add custom status reasons.
The audit, product by product:
| CRM default (source) | Ways a deal ends without a win | Reason and competitor capture | How far it got, and who ended it |
|---|---|---|---|
| ERPNext (commit 821ca1d) | Lost, declared by a user; Closed, set automatically after inactivity in Replied status | Multi-select reasons, separate competitor field, free-text detail | No stage-reached field; no field for who ended it |
| Frappe CRM (commit bbca33d) | One Lost status | Nine shipped reasons; no competitor field; notes required only for Other | Status change log; no field for who ended it |
| EspoCRM (commit 9cf566c) | Closed Lost stage | No reason or competitor field | lastStage keeps the furthest open stage; no field for who ended it |
| Dynamics 365 (Microsoft reference, commit 4684789) | Canceled and Out-Sold status reasons | Competitor on the close record; status reasons can be extended | Documented pipeline phase field; no field for who ended it |
Can a loss reason alone tell a good no from a miss?
Rarely. Of the eight default loss reasons Frappe CRM ships with a description, two can be placed in a single class from the description alone, Competition and Missing Features. The other six cannot be classed without knowing when the loss was recognized, which side ended the deal or what the buyer did instead.
The descriptions share a point of view, too. Seven of the eight begin with the words "The prospect", which frames every ending as something the buyer did. A seller who rightly walked away from a poor-fit account and a seller who lost a winnable deal by ignoring fit until the proposal can both pick Poor Fit, described as "The prospect was not a good fit for the product or service."
Budget Constraints shows why timing decides the class. Found in the first call, a missing budget is a timely disqualification, the cheapest good outcome in sales. Found after a proposal, it is a qualification miss. Arriving after the buyer had confirmed funds, it is a buyer-side event. The reason text is identical in all three cases. The mapping below places each default by its own shipped description:
| Shipped reason | Description, abridged | Classes it could belong to | Fact that decides |
|---|---|---|---|
| Competition | The buyer chose a competitor's product or service | Competitive loss | Placeable as recorded |
| Missing Features | The buyer felt key features were missing | Offer gap | Placeable as recorded |
| Pricing | Pricing too high or not competitive | Competitive loss, offer gap or seller miss | What the buyer did instead |
| Budget Constraints | No budget to proceed | Timely disqualification, seller miss or buyer-side event | When the budget gap surfaced |
| Long Sales Cycle | The process ran long and interest faded | Seller miss or no decision | Which side stalled |
| No Decision-Maker | The contact lacked authority to proceed | Timely disqualification or seller miss | When authority was checked |
| Unresponsive Prospect | No reply to follow-ups | No decision, seller miss or administrative close | Whether follow-up met the standard |
| Poor Fit | Not a good fit for the product | Timely disqualification or seller miss | Who saw it, and how early |
Seven ways a deal ends lost, sorted by actor and timing
Classify every lost deal by two facts first, who ended it and when relative to a commitment point, then by what the buyer did instead. Seven classes cover the endings a revenue team sees, and only one of them, the seller miss, is a failure by definition.
The commitment point is the moment the seller starts spending serious effort for the buyer, usually a proposal, a scoped demo or a pricing request; set it once per company and keep it fixed across the dataset. Ending a deal before that point because the account cannot buy is a timely disqualification, a correct decision. Ending it after that point for a reason that was knowable before it is a late disqualification, coded as a seller miss. A deal may carry a secondary code: a competitive loss in which the seller also skipped a stakeholder keeps both.
The taxonomy, with the evidence each class needs before its label is accepted:
| Class | Who ended it, and when | Evidence required | Common mislabel |
|---|---|---|---|
| Timely disqualification | Seller, before the commitment point | Dated qualification finding: no budget, authority, need or fit | Recorded as a loss to price or to an unresponsive prospect |
| Competitive loss | Buyer, choosing a named alternative | Buyer statement, award notice or debrief naming the alternative | Logged as price when the rival won on terms or features |
| No decision | Buyer, keeping the status quo or going quiet | Last buyer action, follow-up log and a stated or observed status-quo choice | Coded as a competitive loss so it looks less avoidable |
| Offer gap | Buyer, because the offer could not meet a requirement | The requirement in the buyer's words and a product or pricing record confirming the gap | Treated as a sales failure, hiding a product signal |
| Buyer-side event | Buyer, after a qualified need changed | Dated evidence of a budget freeze, reorganization, acquisition or cancelled project | Coded as no decision, losing the outside cause |
| Seller miss | Either side, after an avoidable seller error | The error and its date: late disqualification, missed stakeholder, slow follow-up, wrong scope | Filed under an outside reason that protects the rep |
| Administrative close | No one: a timeout, cleanup or duplicate | System or bulk-update record with no buyer or seller decision | Counted as a loss in win rates and training sets |
Lost deal AI training labels: one deal, three tasks, three rules
Set the label per task, not per deal. A timely disqualification is a correct example for a qualification agent and a failure nowhere; a seller miss is the only class that belongs among wrong-action examples for a seller-assist agent; a forecasting model can treat every loss except an administrative close as a negative outcome.
Getting this wrong has a predictable cost. If every loss is a negative example for a qualification agent, the agent learns that stopping early is bad, because stopped deals all end lost, and it will keep weak opportunities alive. If administrative closes stay in, a forecasting model learns that inactivity causes loss when the timeout rule caused the label. The rules a data team can apply:
| Class | Qualification agent | Seller-assist agent | Close forecasting model |
|---|---|---|---|
| Timely disqualification | Correct stop: a positive example | Exclude: no seller error to learn from | Negative outcome, flagged as an early exit |
| Competitive loss | Correct pursuit if the fit was real | Context only, unless a seller miss is coded | Negative outcome |
| No decision | Review whether the need was ever confirmed | Wrong-action example only with a coded miss | Negative outcome |
| Offer gap | Correct stop if found early, a miss if late | Exclude and route to product feedback | Negative outcome |
| Buyer-side event | Correct pursuit | Exclude | Negative outcome with an outside cause |
| Seller miss | Wrong pursuit when the disqualification came late | Wrong-action example | Negative outcome |
| Administrative close | Exclude | Exclude | Exclude, or treat as unresolved |
Whose account of the loss should the label trust?
The buyer's, when you can get it, and the seller's only with corroboration. A rep choosing a loss reason is explaining an outcome they took part in, and people explain their own failures generously.
The research on that tendency is extensive. A 2004 meta-analysis in Psychological Bulletin by Mezulis and colleagues pooled 266 studies with 503 effect sizes and found the self-serving attributional bias, taking credit for success and blaming outside causes for failure, large on average (d = 0.96) and present in nearly all samples. A picklist completed by the person whose deal it was invites exactly that attribution, so reasons that point away from the seller need corroboration before they become labels.
Public procurement shows what a buyer-side record can contain. Under FAR 15.506, as published on acquisition.gov, a federal agency's debriefing of an unsuccessful offeror must cover the agency's evaluation of the significant weaknesses or deficiencies in that offeror's proposal, the overall evaluated cost or price and technical rating of the winner and the debriefed offeror, and the overall ranking of offerors when one was developed, while point-by-point comparisons with other proposals are barred. Section numbers may shift under the FAR overhaul.
Commercial buyers owe sellers no such debrief, but much of the same content can be gathered: a win-loss interview with the buyer, the buyer's award or rejection message, a scoring sheet shared after a request for proposals. Record the source of every reason with the label and rank it: buyer document, buyer interview, manager review, rep entry.
What to keep, and what to drop, in a lost-deal evidence set
Keep the commercial evidence, the decision authority and the dated outcome; drop the identities of the people involved unless a label cannot be checked without them. Roles carry the signal a model needs, while names add risk and little else. A lost-deal evidence set needs these elements:
At a glance
What to keep, and what to drop, in a lost-deal evidence set
- Stage history with dates, including the furthest stage reached and the date the loss was recognized, so the commitment point can be applied.
- The qualification record as it stood at each stage: budget, authority, need and timing findings with their dates.
- The ending event: who recorded it, the status set, reason codes, competitor fields and free-text detail, kept verbatim beside the class assigned later.
- Buyer-side evidence: rejection or award messages, scoring sheets, debrief or interview notes, summarized to their decision content.
- The commercial offer at the last proposal: price, discount and scope, needed to separate an offer gap from a competitive loss.
- Authority: the roles that approved pursuit, pricing and the decision to stop, not the people in them.
- Administrative markers: bulk updates, timeouts and merges, so these endings can be counted and excluded.
- Removed or replaced with role codes: contact names, email addresses, phone numbers and personal remarks in notes.
Could won deals teach the same lessons?
Only partly, because wins hide misses too. A deal can close despite a late qualification check or a skipped stakeholder, carried by a budget release or a rival's stumble, and a training set that studies only losses will mistake those habits for causes of losing.
Code wins with the same discipline. Apply the seller-miss check to won deals from the same period and segment, so that a miss common to both wins and losses is recognized as ordinary practice rather than a cause of loss. If a miss appears at the same rate among wins and losses, it did not decide those losses; if it is far more common among losses, it becomes a candidate cause worth testing.
That comparison is still observational. Deals with and without a given miss differ in size, segment and timing, so treat any gap as a hypothesis for a controlled change in sales practice, not as proof.
Measuring how often losses are misses: the design, results not yet published
Sample lost deals from one company's approved records, classify them without seeing the rep's chosen reason, and report the class mix with its uncertainty. SourceX has not run this study; the design is set out so that a revenue team or a data buyer can, and results are not yet published. The design:
At a glance
Measuring how often losses are misses: the design, results not yet published
- Frame: every opportunity closed without a win in a fixed 24-month window, with administrative closes counted and then set aside.
- Sample: stratified by furthest stage reached and drawn at random with a seed fixed and published before drawing, such as 20261027, with the sampling script kept beside the data.
- Coding: two sales operations reviewers assign a class from the evidence packet with the rep's reason hidden; a third reviewer settles disagreements.
- Agreement: Cohen's kappa for the class, plus the share of deals left undetermined for missing evidence.
- Outputs: class shares with 95 percent intervals, the share of rep-entered reasons that match the reviewed class, and median days to loss by class.
- Failure condition: if reviewed classes match rep reasons almost everywhere, the default picklist was adequate for that company and the extra labeling is not worth paying for.
Asking for lost-deal labels in a data request
Specify the classes, the commitment point and the evidence before asking for volume. A supplier can then say plainly whether its records carry the timing and buyer-side evidence the labels need, and a buyer can test that claim on a sample.
SourceX keeps no sales datasets waiting on a shelf: it sources each one from U.S. businesses to match a buyer's request, which states the scope, volume, language, format, timing and licensing needs. Whether one company's CRM and email history can support actor-and-timing classes is settled by inspecting a sample, not by assumption. Lines worth writing into the request:
- Unit: one row per opportunity closed without a win, with stage history and the date the loss was recognized.
- Commitment point: the company's definition, applied before any classification.
- Classes: the seven-class taxonomy, with secondary codes allowed.
- Evidence: buyer-side documents where they exist, and the source rank of every reason.
- Exclusions: administrative closes counted and removed; contact identities replaced with roles.
- Acceptance test: blind re-coding of a random sample, with agreement above a level set in advance.
Frequently asked questions
Should lost deals be negative examples in sales AI training?
Only for the classes and tasks where a loss reflects a wrong action. A timely disqualification is a correct decision and belongs among positive examples for a qualification agent; a seller miss is a wrong-action example; administrative closes such as timeouts should be excluded from every label set.
What is the difference between disqualifying a deal and losing it?
Disqualification is the seller's decision to stop pursuing an account that cannot or should not buy, while a lost deal is any ending without a win. Disqualifying before serious effort is spent is a good outcome; disqualifying late, for a reason that was knowable earlier, is a seller miss.
How should a deal closed automatically for inactivity be labeled?
As an administrative close, not as a loss. ERPNext, for example, can move inactive opportunities in Replied status to Closed after a configured number of days, 15 by default, with no reason attached. Count these endings, then keep them out of win rates and training labels.
Are closed-lost reasons entered by sales reps reliable?
Not on their own. Reps choose reasons for outcomes they were part of, and research on the self-serving attributional bias shows people tend to credit outside causes for their failures. Corroborate rep-entered reasons with buyer documents, buyer interviews or manager review, and record which source each label came from.
Which CRM fields are needed to classify a lost deal?
Stage history with dates, the date the loss was recognized, the loss reason and any competitor, the qualification findings in force at each stage, the last commercial offer and the roles that approved pursuit and stopping. Most CRMs lack a field for which side ended the deal, so it usually has to be added or reconstructed.
Sources
- ERPNext's Opportunity requires at least one loss reason to declare a deal Lost, stores competitors and a detailed reason separately, and automatically moves inactive opportunities in Replied status to Closed without a reason. Source
- ERPNext's CRM settings set Close Replied Opportunity After Days to 15 by default. Source
- Frappe CRM's install code creates nine default loss reasons, eight with descriptions, seven of which begin with the words The prospect, and one default status of type Lost. Source
- Frappe CRM's deal record links a loss reason, makes lost notes mandatory only when the reason is Other, keeps a status change log and has no competitor field. Source
- EspoCRM's Opportunity defines a Closed Lost stage at probability 0, a lastStage field and no loss reason or competitor field. Source
- EspoCRM's LastStage hook records the last open stage reached and assigns the lowest-probability stage to an opportunity created as lost. Source
- Microsoft's Dynamics 365 Opportunity reference lists Canceled and Out-Sold as the status reasons of the lost state and documents a pipeline phase field. Source
- Microsoft's Opportunity Close reference documents a Competitor field on the close activity. Source
- Microsoft's documentation says administrators can add multiple status reasons for the Won and Lost states of a Dynamics 365 opportunity. Source
- Mezulis and colleagues' 2004 meta-analysis of 266 studies and 503 effect sizes found a large self-serving attributional bias (average d = 0.96) present in nearly all samples. Source
- FAR 15.506 requires a debriefing to cover the agency's evaluation of significant weaknesses or deficiencies, the overall evaluated cost or price and technical rating of the awardee and the offeror, and the overall ranking when developed, and bars point-by-point comparisons. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.