Build a control record for every outsourcing engagement

This pack turns approved contracts, statements of work, invoices, service reports, governance records, and plans into a cited engagement register with visible gaps and human approvals.

Start with ten engagements. Choose high-spend, high-risk, renewal-near, escalation-heavy, or operationally critical work. Improve the method before scaling to the portfolio.

1. Pick the approved execution path

Path A: approved enterprise AI workspace

Create a dedicated workspace, load this pack and one approved batch of source files, then run the included system and run prompts. Keep file-grounding and citations enabled where available.

Path B: company-controlled repository and coding agent

Use a restricted, versioned repository or controlled document workspace with an enterprise coding agent approved for the data classification. Keep source files immutable and write outputs into a separate folder.

Path C: spreadsheet-first fallback

If source documents cannot be uploaded, have engagement owners populate the inventory and source manifest. Run the approved assistant only on the sanitized sheets. A hosted spreadsheet and bot is acceptable only after security, identity, retention, and data-processing approval.

Do not load: passwords, private keys, production credentials, privileged legal advice, employee medical information, or any record your approved environment is not permitted to process. Record restricted sources in the manifest and route the needed fields to a human reviewer.

2. Collect the source set

supplier and engagement lists
master agreements, statements of work, work orders, and amendments
invoice or purchase-order extracts
service-level reports and quarterly business reviews
transition and operating plans
governance, RACI, and escalation documents
risk, security, privacy, continuity, and exit records
approved savings or improvement plans

3. Run the five passes

Inventory sources and authority.
Identify engagement boundaries.
Extract facts with citations and confidence labels.
Populate the control record and gaps.
Run no-guess QA, then route decisions to named human owners.

4. Pack contents

File	Size
`01_AGENT_SYSTEM_PROMPT.txt`	1,777 characters
`02_RUN_PROMPT.txt`	1,424 characters
`03_FOLLOW_UP_PROMPTS.txt`	1,669 characters
`04_source_manifest.csv`	144 characters
`05_engagement_inventory.csv`	230 characters
`06_engagement_control_register.csv`	372 characters
`07_assumptions_and_gaps.csv`	133 characters
`08_decision_log.csv`	161 characters
`09_SECURITY_AND_HUMAN_REVIEW_CHECKLIST.txt`	1,205 characters

5. Copy the system prompt

Agent system prompt

You are an enterprise outsourcing engagement-control analyst.

Your job is to create a reviewable control record for each engagement from the approved source files and the templates in this pack.

Non-negotiable evidence rules:
1. Never invent, complete, or normalize a material fact without evidence.
2. Every populated material field must cite a filename and page, section, sheet, or row when available.
3. Label each material field CONFIRMED, INFERRED, CONFLICTING, or MISSING.
4. Preserve conflicts. Do not silently choose the newest, most favorable, or most complete value.
5. Treat supplier, contract, and engagement as different objects. One supplier may have multiple engagements.
6. Split engagements when outcome, consuming organization, geography, pricing model, owner, contract scope, or risk profile differs.
7. Do not provide legal conclusions. Route contract, privacy, security, employment, regulatory, and tax questions to qualified human reviewers.
8. Do not expose secrets, credentials, personal data, or restricted information in outputs.
9. Do not modify source files.
10. Stop and report a blocked item if the approved source set is insufficient for a defensible result.

Required outputs:
- source_manifest.csv
- engagement_inventory.csv
- engagement_control_register.csv
- assumptions_and_gaps.csv
- decision_log.csv
- a concise executive summary of coverage, conflicts, gaps, and decisions required

Quality gate:
- every populated material field has a citation;
- all inferences are labeled;
- all conflicts are visible;
- missing is never converted to none;
- engagement splits are explained;
- financial totals reconcile where the source data permits;
- no engagement is marked approved until a named human owner records approval in decision_log.csv.

6. Copy the run prompt

Five-pass run prompt

Use the instructions in 01_AGENT_SYSTEM_PROMPT.txt and the CSV schemas in this pack.

Work in five passes.

PASS 1 - SOURCE INVENTORY
Inventory every approved source file. Record date, document type, apparent authority, restrictions, duplicates, unreadable files, missing appendices, and supersession clues in source_manifest.csv.

PASS 2 - ENGAGEMENT IDENTIFICATION
Create provisional engagements. Do not assume one supplier equals one engagement. Explain every split and record confidence.

PASS 3 - EVIDENCE EXTRACTION
Populate the inventory and control register from cited evidence only. Use CONFIRMED, INFERRED, CONFLICTING, or MISSING for confidence. Put unsupported assumptions and unresolved contradictions in assumptions_and_gaps.csv.

PASS 4 - CONTROL BUILD
For each engagement, populate outcome, scope, baseline, evidence source, commercial model, term, owners, service levels, governance cadence, risk controls, change control, continuity, exit, and open initiatives when supported.

PASS 5 - NO-GUESS QA
Check citations, confidence labels, conflicts, missing fields, engagement boundaries, and financial reconciliation. Produce a concise executive summary with:
- source coverage;
- engagement count;
- evidence-supported field percentage;
- top conflicts;
- top missing controls;
- decisions required;
- recommended human reviewers.

Do not mark any record approved. Human approval belongs in decision_log.csv.

7. Human acceptance gate

No output becomes operating truth until a named engagement owner reviews boundaries, resolves source conflicts, confirms owners and baselines, and records approval in the decision log.