Privacy and preparation
Legal invoice review dataset: tying policy, narrative and approval together
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A legal invoice exception task should join three records as they stood on the line date: the billing guideline clause and version the reviewer applied, the time-entry line with its narrative, codes and timekeeper role, and the approval chain from first flag to final decision, including pre-approvals, appeals and each approver's authority. Counsel screens narratives for privileged content and signs the dataset brief before anything is collected.
Document stack
Legal invoice review dataset
Key takeaways
- A flagged invoice line is only a usable exception when the rule, the entry and the decision can all be traced to dated records.
- The governing policy is the guideline version in force on the line date plus any matter-specific approval, not the guideline in force today.
- Time-entry narratives are the sensitive field; keep the administrative signal and remove the subject of the legal work.
- Approver authority is part of the label: a reduction reversed by someone with higher authority is a different outcome from one the firm accepted.
- Exceptions with missing approvals or pending appeals are labelled unresolved, never counted as violations.
- Counsel reviews the narrative screen and signs the dataset brief before collection, and the exclusion log travels with the data.
What belongs in a legal invoice exception task?
One task is one exception: a flagged line or invoice, the policy that was applied to it, the screened entry itself, and every decision made about it up to a final outcome. Strip away any one of the three and the record stops teaching the judgment. A guideline clause without entries is just a rule. Entries without the decision chain are text with no label. Decisions without the clause show that something was cut, but not why.
Legal departments generate these records whenever an e-billing system or a reviewer flags a line from outside counsel: block-billed time, a rate above the approved rate, staffing beyond what the guideline allows, administrative work billed as legal work. The reviewer reduces or approves, the firm sometimes appeals, and someone with authority closes it. A legal invoice review dataset built from that history is administrative evidence about billing, not legal advice, provided the narratives are screened.
Most departments already hold the parts in separate places: guidelines in a policy library, invoices in the e-billing platform, approvals in the platform audit trail and in email. The work is joining them at the line level with the right versions.
Which policy governed the line on the day it was billed?
The governing policy is the most specific written authority in force on the line date, not the version the department uses today. Outside counsel guidelines are revised, rate approvals are renewed, and matter budgets and staffing approvals are granted along the way, so a reviewer's reduction can only be judged against what applied at the time.
The ethics rules that sit underneath explain why many exceptions exist. ABA Model Rule 1.5 says a lawyer may not charge or collect an unreasonable fee or unreasonable amount for expenses, and ABA Formal Opinion 512, applying the earlier Opinion 93-379, says hourly billing must reflect time actually spent and that general overhead is not billed to clients. Guidelines turn those principles into specific, checkable rules, and states adopt their own versions of the model rules.
| Authority source | What it controls | Version evidence to keep | Typical conflict to resolve |
|---|---|---|---|
| Engagement letter | Billing basis, rate terms and the guideline it incorporates | Signed date and any amendment dates | Letter incorporates an older guideline than the current one |
| Outside counsel guidelines | Task rules such as block billing, minimum increments, staffing and expenses | Version identifier and effective date | Line dated before a new version took effect |
| Approved rate list | Rate per timekeeper or class for a period | Approval date and period covered | New rate billed before approval arrived |
| Matter budget or staffing approval | Exceptions granted for one matter | Approver, date and scope of the exception | Approval given by email and never recorded in the platform |
| Platform rule configuration | Which automated flags fire | Rule set version or change log | Rule engine stricter or looser than the written guideline |
What in a time-entry line is evidence, and what has to be screened?
The structured fields carry most of the administrative signal, while the narrative carries both signal and the most risk. LEDES.org describes LEDES 1998B as an ASCII, pipe-delimited format with 24 fields and the most widely used e-billing standard in US legal billing; LEDES 98BI extends the record to 52 fields. Lines usually carry UTBMS codes: the ABA litigation code set groups task codes into phases with L-prefixed codes and pairs them with A-prefixed activity codes, which not every firm uses.
At a glance
What in a time-entry line is evidence, and what has to be screened
The narrative is where confidentiality bites. ABA Model Rule 1.6 bars a lawyer from revealing information relating to a representation without informed consent or an exception, and requires reasonable efforts to prevent unauthorized disclosure; a department's own duties to its company and to opposing parties also apply. Federal Rule of Evidence 502 addresses when a disclosure waives privilege or work-product protection, but it does not decide whether material is privileged in the first place, so the screen happens before anything moves.
A workable screen keeps what the reviewer actually judged and removes what the work was about.
- Keep: units, line date, timekeeper class, task and activity codes, number of distinct activities in the entry, and whether separate time was given for each.
- Keep, normalized: activity verbs such as draft, review, call or travel, mapped to a fixed vocabulary.
- Remove: party, witness and expert names, case captions, docket numbers, the subject of advice and anything describing strategy.
- Replace with stable tokens: firm, timekeeper and matter identifiers, so patterns stay learnable without identities.
- Hold for counsel: any entry whose narrative cannot be reduced without losing the reason it was flagged.
Which approvals and decisions make up the outcome label?
The outcome is the last decision made by someone with authority to make it, and the dataset needs the chain that led there. Record each step as an event with the actor's role and authority level, so a model can learn the difference between a reduction the firm accepted and one an escalation overturned.
| Event | Fields to capture | Authority to record | Common gap |
|---|---|---|---|
| Automated flag | Rule identifier, rule set version, flag time | System, with the configuration owner | Rule changed without a version number |
| Reviewer adjustment | Units or line reduced, reason code, comment category | Reviewer role and approval limit | Free-text reason with no code |
| Pre-approval relied on | Approval reference, scope, date | Approver role | Approval lives only in an email thread |
| Firm appeal | Appeal date, appeal category, firm's stated basis in screened form | Firm billing contact role | Appeal text carries matter detail that must be removed |
| Appeal decision | Upheld, partially upheld or denied, final units | Decider role and whether it exceeds the reviewer's authority | Decision recorded as a new invoice, breaking the link |
| Closure | Final status and date | Platform audit trail | Invoice still open, so the outcome is not mature |
The professionally reviewed administrative invoice dataset brief
The brief is the document counsel and legal operations sign before collection; it fixes scope, authority and limits so the data can be checked against it later. Each section states a decision, names who signs it and lists what evidence is kept.
At a glance
The professionally reviewed administrative invoice dataset brief
| Brief section | What it states | Who signs | Evidence kept |
|---|---|---|---|
| Purpose and task definition | What one exception task is and what the dataset is for | Legal operations lead | Task definition with one sample record |
| Nonprivileged scope | Matter types, date window and fields included | Counsel | Field list with keep, normalize, remove or token rule per field |
| Source authority | Who in the company may approve release of these records | Counsel | Delegation or policy naming the approver |
| Policy versions | Every guideline, rate list and approval type used to label | Legal operations lead | Version register with effective dates |
| Amendments | Mid-matter changes to letters, rates or guidelines | Legal operations lead | Amendment log keyed to matter tokens |
| Approval capture | Which approval events are captured and from which systems | Legal operations lead | System list and audit export description |
| Narrative screen | Method used to reduce narratives and who checked samples | Counsel | Screen rules and a signed sample review |
| Exclusion record | Categories withheld and counts, never content | Counsel | Exclusion log by category |
| Outcome maturity | When an exception counts as final | Legal operations lead | Maturity rule and count of unresolved tasks |
| Limits | What the data cannot support, such as judging legal work quality | Counsel and legal operations lead | Written limits statement |
Illustrative worked example: a block-billed deposition entry across two guideline versions
This example is illustrative; the guideline clauses, durations and roles are invented. A department's guideline version 3 requires separate time for each task in entries longer than half an hour. Version 4, effective on the first of a later month, adds a rule that deposition preparation by more than one attorney needs prior approval.
An associate's entry, dated two weeks before version 4 took effect, bills 3.4 hours for preparing a deposition outline, reviewing a production and a client call, coded to a deposition task and a drafting activity. The rule engine, already configured for version 4, flags both block billing and unapproved second-attorney staffing.
The reviewer cuts 0.8 hours for block billing and rejects the staffing flag only after checking the line date against the version register. The firm appeals the block-billing cut, attaching a breakdown of the three tasks. A deputy general counsel, whose approval limit exceeds the reviewer's, partially upholds the appeal and restores 0.5 hours.
The finished task records the version 3 clause as the governing policy, the screened entry with three activities and no names, the four events with roles, and a final label of block billing, partially upheld on appeal. The exception worth preserving is the staffing flag: it is kept as a false positive caused by a rule configured ahead of its effective date, which is exactly the case a model needs to see.
Which source, amendment or missing record decides the label?
The line date picks the guideline version, and a written, matter-specific approval overrides the general rule where the guideline allows such approvals. When those two disagree, record both and let the approval decide only if it predates the line and covers its scope.
Missing evidence changes the label rather than being filled in. If a firm cites an approval the department cannot find, label the task unresolved authority instead of treating it as a violation. If an appeal is still pending or the invoice is open, the outcome is not mature and the task stays out of training splits until it closes. If the rule engine fired under a configuration that did not match the written guideline, label the flag as a configuration error and keep it.
Report the counts: tasks by final label, by guideline version, and how many are unresolved or immature. Those numbers show reviewers how much of the history actually has a settled answer.
Which invoice exceptions should never leave the legal department?
Some exception records stay inside even after screening, because the matter, not the field, carries the risk. Counsel decides the categories; the exclusion log records them by count.
- Matters under a legal hold or active litigation with the billing firm itself; holds and retention policies are covered separately.
- Government investigations and other matters where the existence of outside counsel work is itself sensitive.
- Matters handled for another party, such as insured defendants in insurance defense, where the department is not the only client.
- Entries whose narrative cannot be reduced without losing the reason for the flag.
- Fee disputes that turned into claims against the firm or by the firm.
- Any record the engagement letter or a protective order restricts beyond ordinary confidentiality.
Where licensing fits once the brief is signed
A signed brief with a version register, a narrative screen and an exclusion log is the material a licensing conversation starts from, because it answers the rights and privacy questions before records move. When a company licenses screened billing records through SourceX, SourceX carries the work from sourcing and rights review to contracting, preparation, delivery and payment.
A legal department finds out whether its company qualifies through a brief assessment that involves no invoice files. Before any record is delivered, the department approves the scope and terms and signs the agreement, so counsel's exclusions are fixed in writing first.
Frequently asked questions
What should a legal invoice exception dataset brief include?
It should define one exception task, the nonprivileged field scope, who has authority to release the records, the register of guideline and rate versions, how approvals are captured, the narrative screen, an exclusion log by category and the rule for when an outcome is final. Counsel and the legal operations lead sign their sections before collection.
Are law firm time-entry narratives privileged?
They can be, because they often describe the subject and strategy of the work. Treat narratives as confidential by default, keep only administrative signal such as activity type and units, and have counsel review samples of the screened output before anything leaves the department.
Which guideline version applies to a flagged line?
Use the version in force on the line date, plus any written matter-specific approval that predates the line and covers it. Keep the version register with effective dates so each label can be checked against the rule that actually applied.
How should appealed reductions be labelled?
Record the reviewer's adjustment, the firm's appeal and the final decision as separate events with each decider's role and authority. Label the task by the final decision, and leave tasks with pending appeals out of training splits until they close.
Can LEDES files alone train an invoice review model?
Not well. LEDES files carry what the firm submitted, while the review decisions, reasons, appeals and approvals sit in the e-billing platform's audit trail and in the department's policy records. The exception task needs both, linked at the line level.
Sources
- LEDES.org describes LEDES 1998B as an ASCII, pipe-delimited format with 24 fields and the most widely used US e-billing standard. Source
- LEDES 98BI is a pipe-delimited format with 52 fields in its current version. Source
- The ABA UTBMS litigation code set organizes task codes by phase and pairs them with activity codes, which not every office uses. Source
- ABA Model Rule 1.6 restricts revealing information relating to a representation and requires reasonable efforts to prevent unauthorized disclosure. Source
- ABA Model Rule 1.5 bars charging or collecting an unreasonable fee or unreasonable amount for expenses. Source
- ABA Formal Opinion 512, applying Opinion 93-379, says hourly billing must reflect time actually spent and general overhead is not billed. Source
- Federal Rule of Evidence 502 limits when disclosure waives privilege or work-product protection and does not decide whether material is privileged. Source
See if your company qualifies
A short company assessment. No data uploads are needed.