Skip to content

AI uses for records

Seven causes, one baseline: building a sales scope correction dataset

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

A change order reveals a bad initial scope only when the work it adds or reprices was implied by records dated before signature, or the seller's own sizing inputs showed the sold quantity was too small. Build a sales scope correction dataset by joining each amendment line to the signed baseline version, the first dated mention of the requirement and the delivery signal behind it, then label one of seven causes.

Value drivers

Seven causes, one baseline

This page covers seven causes, one baseline: building a sales scope correction dataset.Illustrative

Key takeaways

  • Only three of seven change-order causes, omitted requirements, under-sized lines and failed assumptions, say the first scope was wrong.
  • The deciding fact is the date of the first record naming the requirement, compared with the signature date of the baseline version.
  • IFRS 15 calls change orders contract modifications and separates new distinct work at stand-alone prices from changes to the existing promise, a useful cross-check.
  • ERP and billing systems record what changed and leave the reason to free text; federal procurement data shows a coded reason can be required at entry.
  • Projects without change orders are not proof of good scoping, and change orders after an overrun need dated evidence before they count as scope corrections.

When does a change order prove the first scope was wrong?

A change order proves the first scope was wrong only when the work it adds or reprices was already implied by records dated before signature, or when the sold quantity was too small by the seller's own sizing inputs. Most change orders prove something else: the buyer wanted more, a planned phase began, an outside rule changed, or someone fixed a date or a price.

That distinction is what a sales scope correction dataset has to preserve. A model trained to review quotes, draft statements of work or flag risky deals needs examples of scopes that failed and of scopes that held while the deal grew. If every amendment is treated as a scoping failure, the model learns to pad quotes; if none is, it never learns which missing sentence cost the seller a renegotiation.

The decision for an AI team is therefore narrow and testable: for each amendment line, which record fixes the date the requirement first appeared, and does that date fall before or after the signed baseline? The rest of this article builds the evidence map, label set and rules that answer it from records a seller already keeps.

What revenue accounting already decides about each change order

IFRS 15 already treats a change order as a defined event and sorts it in a way a dataset can borrow. Paragraph 18 defines a contract modification as a change in the scope or price, or both, approved by the parties, and notes that it may be described as a change order, a variation or an amendment. Approval can be written, oral or implied by customary business practice, which is why a dataset cannot rely on signed documents alone to date a change.

Paragraph 19 says a modification can exist even when the parties have approved a change in scope but not yet agreed the price, in which case the seller estimates the price effect. That unpriced state deserves its own field, because scope corrections are often agreed in principle long before anyone prices them.

Paragraphs 20 and 21 then split modifications by substance. A modification is a separate contract only when it adds distinct goods or services at a price reflecting their stand-alone selling prices. Otherwise the seller treats it as ending the old contract and starting a new one, or adjusts the existing contract on a cumulative catch-up basis. Finance's choice is a free, independently made signal: new distinct work sold at list price looks like growth, while a change folded into the existing promise looks like a correction or a concession. Treat it as a feature to check against, not as the label.

What ERP, billing and federal records keep, and the field they skip

Commercial systems record what changed in detail and record why only as free text, so the cause label has to come from review rather than export. Federal contracting shows that a coded reason can be captured at entry when a system requires it.

Tryton's sale amendment module, read at the current GitHub mirror commit, stores each amendment with a date, a free-text description, a state of draft or validated, and action lines that can recompute taxes, change the payment term, change parties and addresses, change the warehouse, or change an existing line's product, quantity, unit, unit price or description. Validating an amendment raises the sale's revision number and replaces draft shipments and invoices. The module depends on Tryton's sale history module, which keeps historical versions of sales and sale lines and appends the revision to the sale number, so the signed baseline survives. No action adds a line, so new items cannot enter a sale through an amendment.

Zuora's order actions describe changes mechanically: add, update or remove a product, change a plan, change terms and conditions, renew, suspend, resume, transfer ownership or cancel, working at the rate plan or subscription level. None of the listed types names a cause. The U.S. Federal Procurement Data System, by contrast, requires a Reason for Modification on every modification, with codes that include additional work under a new agreement, a supplemental agreement for work within scope, a funding-only action, a change order and exercise of an option. The comparison:

What ERP, billing and federal records keep, and the field they skip
Record sourceKeeps the signed version?Records what changedRecords whyWhat a correction label still needs
Tryton sale amendment with sale historyYes: historized sale and lines, revision numberAction lines with new valuesFree-text description onlyFirst-mention date and a reviewed cause
Zuora order actionsDepends on how orders are retainedAction type at rate plan or subscription levelNot among the action typesLink to the requirement record and a cause
Federal procurement data systemBase award and numbered modificationsModification transactionCoded reason, required on modificationsEvidence for whether a within-scope label was right
Finance modification memo under IFRS 15Refers to the existing contractScope and price changeSeparate contract or modification of the existing oneConfirmation that the treatment matches the facts

Seven causes behind a change order

Seven causes cover the amendments a seller usually sees, and only the first three say the first scope was wrong. Each label carries a test that a reviewer applies to dated records, so two people working from the same evidence reach the same answer.

The labels describe the scope, not the commercial result. A change order that adds revenue can still be an omitted requirement, and a concession that costs revenue can have nothing to do with scoping. The taxonomy and its tests:

Seven causes behind a change order
LabelTest applied to the recordsEvidence that settles itWhat it says about the first scope
1. Omitted requirementRequirement appears in a record dated before signature and is absent from the signed baselineRequest for proposal, discovery notes or buyer message with its dateWrong: the seller had the requirement and left it out
2. Under-sized lineBaseline names the work, but the corrected quantity exceeds the sold quantity and the seller's sizing inputs at signing already implied moreEstimate worksheet, sizing questionnaire, delivery records for the lineWrong: the work was seen but under-estimated
3. Failed assumptionA stated baseline assumption is contradicted by a dated delivery recordAssumptions list in the statement of work; the record that broke itExposed: the risk was named; the owner of the assumption decides who carries it
4. New buyer requirementFirst dated mention comes after signature, raised by the buyer, not implied by the baselineBuyer request after signing; separate-contract treatment where presentHeld: demand grew
5. Planned option or phaseThe baseline already contains the option or phase being exercisedOption clause or phase plan in the signed versionHeld: the expansion was anticipated
6. Outside changeA regulation, third-party product or index changed after signatureDated external notice or vendor change recordNeither side's scoping
7. Administrative, concession or remedyNo change in scope: dates, parties, billing terms, a price-only reduction or a credit for a delivery shortfallAmendment lines with no scope delta; delivery defect recordsNot a scope event; a remedy points at delivery
Evidence span chart for seven change-order causes: omitted requirement and under-sized line start before the signed baseline, failed assumption and planned option start at it, new buyer requirement and outside change start after it, and administrative changes appear only at the change order.
Where the first dated record sits relative to signature decides the label; positions are schematic.

The evidence map: five records joined on one signed baseline

Every labeled change order needs five records joined on the identifier of the signed baseline version: the baseline itself, the first dated mention of the requirement, the delivery signal that raised it, the amendment, and the outcome afterwards. Without the baseline version, a reviewer is comparing the change against the current contract, which already contains it.

Commercial evidence and decision authority must survive the export, but people's identities rarely need to. Replace names with roles, such as buyer budget holder, seller account executive or delivery lead, and keep each role's approval date. The records, what to keep and what to drop:

  • Signed baseline: version identifier or revision number, lines with quantities and prices, stated assumptions, options and phases, signature date. Drop signatory names; keep signatory roles.
  • First-mention record: the date, channel and author role of the earliest record naming the requirement, with the requirement sentence quoted. Drop the rest of the message body, addresses and signatures.
  • Delivery signal: the dated event that surfaced the gap, such as an acceptance test failure, an issue in the project tracker or a support ticket, keyed to the baseline line it concerns.
  • Amendment: each changed line with its old and new values, the approval date and approving role on each side, whether the price was agreed or still open, and finance's treatment where recorded.
  • Outcome: the invoice or credit that followed and whether the changed scope was later accepted. Keep amounts if the supplier permits, or normalize them to a share of the baseline value.

Labeling rules a second reviewer can reproduce

Apply the rules in a fixed order and stop at the first that fits, labeling each amendment line separately because one change order often mixes causes. The order matters: checking for administrative changes and planned options first prevents ordinary paperwork from being counted as a scoping failure.

At a glance

Labeling rules a second reviewer can reproduce

Run a second reviewer on a sample of at least a few dozen change orders, blind to the first reviewer's labels, and report agreement by label. Disagreements usually trace to a missing first-mention record, which is a data gap to fix, not a judgment to average away. The rules:

  • Rule 1. If the line has no link to a signed baseline version, label it unresolved and exclude it from training.
  • Rule 2. If the amendment line changes no scope, label it administrative, concession or remedy, and record which.
  • Rule 3. If the baseline contains the option or phase being exercised, label it planned option or phase.
  • Rule 4. If a dated external event caused the change, label it outside change.
  • Rule 5. If a stated baseline assumption was contradicted by a dated delivery record, label it failed assumption and record which party owned the assumption.
  • Rule 6. If the requirement's first dated mention precedes signature and the baseline omits it, label it omitted requirement; if the baseline names it but the seller's own sizing inputs implied more, label it under-sized line.
  • Rule 7. If the first dated mention follows signature and the buyer raised it, label it new buyer requirement.
  • Rule 8. If no first-mention record exists, label it unresolved rather than defaulting to new buyer requirement.

Does a project with no change orders prove the scope was right?

No. Silence can mean the seller absorbed the gap, the buyer tolerated it, or the dispute moved into the renewal, so a quiet project is not a positive example of good scoping. Treat it as unlabeled unless delivery records show the sold lines were completed within their stated quantities and assumptions.

The reverse error is just as common. Overruns are ordinary in project work: a 2004 Simula publication by Moløkken-Østvold, Effort and Schedule Estimation of Software Development Projects, summarizes surveys in which 60 to 80 percent of software projects overran effort or schedule. A change order that follows an overrun is therefore weak evidence by itself; the label needs the dated sizing inputs or first-mention record that show the gap was visible at signing.

A won and expanded deal can still rest on a poor scope, and a lost renewal can follow a sound one. That is why the labels describe the scope against its baseline rather than the deal's commercial outcome, and why the outcome record sits beside the label instead of deciding it. For how delivered hours split between the parties after a handoff, the sales handoff dataset page covers who carried each gap.

Writing the dataset request: counts, fields and an acceptance test

Ask suppliers first for counts that show whether the evidence exists, then for the fields, then for a test that proves the labels can be reproduced. A useful request states the intended use, such as training a quote review assistant or evaluating a statement-of-work drafting agent, because the use decides how many unresolved cases are tolerable.

At a glance

Writing the dataset request: counts, fields and an acceptance test

SourceX sources datasets for AI buyers on request from U.S. businesses rather than holding stock, and buyers describe scope, volume, language, format, timing and licensing needs through its request form. The supplying company approves the scope and terms and signs before any records are delivered. What to put in a request for scope correction records:

  • Counts: closed projects or orders with a retained signed baseline; amendments with prior versions recoverable; amendments with a first-mention record; the date range covered.
  • Fields: the five records in the evidence map, with roles instead of names and amounts normalized if the supplier prefers.
  • Labels: the seven causes plus unresolved, applied per amendment line, with the reviewer and rule number recorded.
  • Acceptance test: a held-out sample relabeled by the buyer's reviewer from the same records, with agreement reported by label and a stated minimum for the scope-correction labels.
  • Exclusions: deals without a recoverable baseline, and any records the supplier's customer contracts do not permit it to license.

Limits of the seven-cause labels, and the evidence still missing

The labels depend on records many sellers do not keep: discovery calls are often unlogged, estimate worksheets live in personal files, and approvals are given orally, which IFRS 15 explicitly allows. Expect a large unresolved share in early samples and report it rather than filling it with guesses.

Other limits are structural. The record comparison draws on one open-source ERP, one billing system's documentation, one federal data system and one accounting standard; other systems may store reasons in custom fields that would change the export picture. Accounting treatment varies with judgment and with the framework a seller reports under. No sample of change orders has been labeled for this article, so no share of scope corrections is reported.

The next evidence to collect is a permissioned labeling study: a defined population of closed deals, two blind reviewers, agreement by label, the unresolved share as its own denominator, and a comparison of finance's treatment with the reviewed cause. The design is stated here; results are not yet published. Sources were checked on 27 October 2026: Tryton through a clone of its GitHub mirror, and the IFRS Foundation, FPDS, Zuora and Simula pages through search excerpts on their own domains.

Frequently asked questions

What is a sales scope correction dataset?

It is a set of change orders or contract amendments, each linked to the signed baseline it modified, the earliest record of the requirement and the delivery event that raised it, and labeled by cause. It lets an AI system learn which gaps in a quote or statement of work led to corrections, as distinct from amendments that reflect new demand or paperwork.

Is every change order evidence of poor scoping?

No. Many change orders add work the buyer asked for after signing, exercise an option the contract already contained, respond to an outside change or adjust dates and billing terms. Only changes whose requirement was on record before signature, or whose quantity was visibly under-sized at the time, point to a bad initial scope.

Which fields should a change order record keep for AI training?

Keep the baseline version identifier, each changed line with old and new values, approval dates and approving roles, whether the price was settled, and the link to the requirement and delivery records. Replace personal names with roles and quote only the sentence that states the requirement.

Can accounting treatment tell new scope from a correction?

It can help. Under IFRS 15, a modification that adds distinct goods or services at stand-alone selling prices is a separate contract, while other changes adjust the existing contract. That treatment reflects judgment, so use it to cross-check a reviewed label rather than as the label itself.

Why label each amendment line instead of the whole change order?

One change order often bundles an omitted requirement with new buyer requests and an updated billing date. A single label would hide the correction inside the growth, so each line gets its own cause and rule number.

Sources

  • Tryton's sale amendment module stores amendments with a date, free-text description, draft or validated state and action lines (taxes, payment term, parties and addresses, warehouse, or an existing line's product, quantity, unit, unit price and description); validation raises the sale's revision and replaces draft shipments and invoices. Source
  • Tryton's sale history module historizes sales and sale lines and adds a revision counter that is appended to the sale number. Source
  • IFRS 15 paragraphs 18 to 21 define a contract modification, which may be called a change order, variation or amendment; allow approval in writing, orally or by customary practice; cover scope changes approved before pricing; and treat a modification as a separate contract only when it adds distinct goods or services at stand-alone selling prices. Source
  • The FPDS Reason for Modification element describes the type of modification, with codes including additional work under a new agreement, supplemental agreement for work within scope, funding only action, change order and exercise of an option. Source
  • The FPDS user manual version 1.5 (October 2025) lists Reason for Modification as required on modifications and not applicable on base awards. Source
  • Zuora's order actions include create subscription, add, update and remove product, change plan, terms and conditions, suspend, resume, renewal, owner transfer and cancellation, at rate plan or subscription level. Source
  • A 2004 Simula publication by Moløkken-Østvold, Effort and Schedule Estimation of Software Development Projects, summarizes surveys finding that 60 to 80 percent of software projects encounter effort or schedule overruns. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies