skills/spend-management-analysis/SKILL.md

---
name: spend-management-analysis
description: Analyze legal e-billing, outside-counsel spend, law-firm timekeeping, or matter-economics files to diagnose staffing mix, budget variance, allocation, and supported AI-leverage opportunities. Use for legal-spend reviews, QBR prep, spend spikes, matter pacing, realization, or practice-group economics. Do not use for general procurement spend or legal advice.
---

# Spend Management Analysis

Turn uploaded legal billing or staffing data into an evidence-backed explanation of what happened, why it likely happened, and what the user can do next. Do not merely restate totals or recreate an e-billing dashboard.

Resolve bundled paths relative to this `SKILL.md` file.

## Start here

1. Inspect only the files the user provided. Classify the data as:
   - in-house/legal-ops data about outside-counsel spend; or
   - law-firm internal data about practice economics, realization, or staffing.
2. Check for LEDES before assuming CSV or Excel. If the first line starts with `LEDES1998B[]`, read [references/ledes-1998b.md](references/ledes-1998b.md) and use `scripts/parse_ledes.py` before analysis.
3. Determine the audience. Infer it only when the request clearly identifies the setting, such as QBR preparation or an internal legal-operations review. Otherwise ask one concise question before producing the full report: is this for an internal review, a conversation with outside counsel, or an executive/board audience?
4. Read [references/analysis-method.md](references/analysis-method.md), then normalize and analyze the data.
5. Before delivering results, read [references/report-and-export.md](references/report-and-export.md) and create the required report, workbook, and findings CSV.

## Core standards

- Normalize platform-specific columns into a consistent working table, but preserve enough source detail to trace findings back to the input.
- Base every material claim on the supplied data. State missing fields, thin samples, and analytical limits plainly; never fill gaps with invented values.
- Compare like with like when possible: similar matter types, phases, firms, practice groups, or periods actually represented in the input.
- Separate a likely staffing, rate, scope, or intake-process driver from the observed variance. Label inference as inference.
- Recommend actions; do not make binding staffing, vendor, or firm decisions.
- Treat attorney names, rates, client matters, and invoice details as sensitive. Repeat only the identifying detail needed to support a finding.

## Required outcome

Deliver:

1. An audience-tailored written report with an executive summary, ranked actions, staffing-mix findings, budget-variance findings, allocation findings, and supported AI-leverage opportunities.
2. A polished `.xlsx` workbook generated with `scripts/build_workbook.py`.
3. A flat findings `.csv` generated by the same script.

If a report section has no supported finding, say so briefly. Include AI-leverage opportunities only when task codes, activity codes, or line descriptions identify actual work categories. A matter name, practice area, role mix, or rate alone is not enough evidence for an AI-leverage finding. In the workbook payload, record whether AI leverage was `assessed` or `not_assessed` and include a short assessment note; do not treat an empty opportunity list as proof that an assessment occurred.

## Boundaries

- Do not access outside billing or timekeeping systems unless the user separately authorizes and provides an available integration.
- Do not assess fee reasonableness under ethics rules or give legal advice.
- Do not claim a trend without the relevant comparison period.
- Do not invent dollar or hour savings from AI. Keep AI opportunity estimates qualitative and require attorney review plus applicable confidentiality, privilege, and AI-use controls.

## Runtime fallback

The workbook helper requires Python 3 and `openpyxl`. If code execution or that dependency is unavailable and cannot be enabled within the current environment, still deliver the written analysis and a findings table, clearly disclose that the workbook could not be generated, and give the user the bundled command they can run locally.