
Compliance + audit trail for AI work
An audit-trail format regulators and lenders will thank you for. Anchored to the MHG SBA 7(a) loan process, a real in-flight financing.
Paper trail, why
Mile High Golf is in the middle of an SBA 7(a) loan process for the Denver NC flagship site at 7521 Eastern Medical Dr. The loan hasn’t closed. MHG is preparing the application package (financial projections, site analysis, business plan, market research) with significant AI assistance from the SUMMIT agent. Curtis Woodie and Olivia Merlock are doing the venue ops work. Sarah Cooley at Ascent RE is advising on the site.
When the underwriter reviews this package they’ll ask: how did you arrive at these numbers? Who reviewed this analysis? What are the assumptions behind this projection? If the answers exist only in an agent session that was never saved, or in a session transcript that can’t be cleanly presented to a lender, the loan review stalls.
AI assistance creates a compliance gap that human-only work doesn’t have. A human financial analyst who builds a projection owns it. Their name is on it, their methodology is documentable, review of the inputs is assumed. An agent that builds a projection isn’t an analyst. The operator who submits that projection to a lender needs to be the human who can answer for it. The audit trail is what makes that possible.
Same discipline applies across the TruPath portfolio. QC provisional patent application (patent-sensitive, attorney review required). Crave Athletics NDA (already signed, CIPHER confirms status). Investor materials (CEO review required before any distribution). Anywhere AI-assisted work goes to an external party with scrutiny rights, the audit trail is the governance mechanism.
Where regulators and lenders actually look
SBA 7(a) underwriters aren’t looking for AI disclosures. The SBA doesn’t currently require them. They’re looking for the things they’ve always looked at. Reproducibility, accuracy, and the human who stands behind the numbers.
| What they check | What they’re actually asking | What you need |
|---|---|---|
| Projection reproducibility | Can you show us how you got from inputs to these numbers? | Named assumption set file + reproduction test |
| Assumption defensibility | Are the inputs to your projection reasonable for this market? | Comparable venue data, cited sources, human domain review (Curtis/Olivia on ops assumptions) |
| Human accountability | Who reviewed and stands behind this document? | Review record with name, date, attestation |
| Source traceability | Where did the market data come from? | Cited sources in the document; not “agent research” |
The audit trail doesn’t need to disclose AI assistance. It needs to demonstrate human accountability. The “AI-assisted, human-reviewed” standard satisfies all four checks. The human reviewer can reproduce the projection from its inputs, defend the assumptions, and has their name on the attestation.
Audit trail format
Four artifact types in the MHG compliance context
Type 1, AI-assisted projection. Any financial forecast, revenue model, or cost projection with AI assistance. Required accompanying artifacts: input assumption set with data sources and date, human review record with reviewer name and attestation, sensitivity analysis noting what changes the output by more than 10%. The Denver 18-month revenue projection is this type.
Type 2, AI-assisted research or analysis.Market analysis, site comparison, competitive landscape, demographic research. Required: methodology note (what data sources, what criteria), human review record, limitations note (what this analysis doesn’t cover). The Denver NC site selection analysis is this type.
Type 3, AI-drafted document. Business plan narrative, letter of intent, grant application, executive summary. Required: human review with tracked changes, final version approved by named human, version history (v1 draft, v2 after review, final submitted). The MHG concept plan delivered to the Denver landlord is this type.
Type 4, AI-assisted compliance checklist. SBA eligibility verification, grant requirements cross-check, regulatory requirement mapping. Required: source document references for every checklist item, date last verified, human sign-off. The Lincolnton TIP and Facade grants (likely lost in the site change) need a fresh review against the new Denver site before writing them off.
Compliance trigger list in CLAUDE.md
Every session topic that produces an external-facing artifact should trigger artifact-creation before the session ends. MHG compliance triggers: SBA, lender, underwriter, loan application, financial projection, revenue model, site analysis, grant, lease, term sheet. QC compliance triggers: patent, provisional, IP assignment, investor materials, PIIA, NDA terms, ACL. Parley triggers: academic publication, community claim about product capability, grant application.
The trigger list goes in CLAUDE.md as a hard rule. “For any session producing content that touches [compliance trigger list], save the output as a named artifact before ending the session. Artifact minimum: title, date, author/reviewer, status (draft/reviewed/submitted).”
Three failure patterns
How compliance audit trails fail for AI-assisted work. Each failure is recoverable before submission. None recoverable after.
AI-drafted document with no human review record
Projection inputs not preserved
Agent session used as the compliance record
All three failures trace to treating sessions as records. Sessions are ephemeral. They end, and their content may not be recoverable in a form a lender can review. Artifacts aren’t ephemeral. The discipline: compliance-trigger sessions produce artifacts before they end.
Trail completeness check
Run this before submitting any document set to an external party. Takes about 15 minutes for a standard loan packet and surfaces every document that isn’t compliance-ready.
Run a compliance audit trail completeness check on this
document or document set.
Context: [describe what you're preparing and for whom]
Recipient: [lender / regulator / investor / partner]
AI assistance used: [which documents were AI-assisted?]
For each AI-assisted document:
1. REVIEW RECORD: Is there a record of human review?
- Reviewer name, date, changes made, attestation?
- If not: who should review before this goes external?
2. INPUT PRESERVATION: For projections or analyses —
are the input assumptions saved as a separate artifact?
- Is the projection reproducible from the assumption set?
- If not: what assumptions are missing?
3. ARTIFACT VS SESSION: Is the output saved as a
dated, named artifact — or does it only exist in
a session that may not be recoverable?
- If session-only: extract and save before proceeding.
4. ATTRIBUTION: Is the AI-assisted status of each
document disclosed appropriately?
(Lender/regulator requirements vary — some require
disclosure, some don't. Check the specific requirement.)
Output: a gap report — which documents are compliance-ready
and which need work before external submission.Checks review records, input preservation, artifact-vs-session, and attribution for every AI-assisted document. Run before external submission.
Annotated transcript
Building the audit trail for the MHG SBA loan packet. Two turns. One on overall structure, one on the specific format for the Denver revenue projection assumption set.
Two operator turns, two agent turns. Result: three-layer audit trail structure, specific assumption set format for Denver revenue projection, review routing through Curtis and Olivia.
- operator
SUMMIT is helping me prepare the MHG SBA 7(a) loan packet. The application includes financial projections and site analysis that were AI-assisted. What's the audit trail I need to maintain so this holds up to lender scrutiny?
- agent
Three things a lender will look at, and what you need for each:
1. Financial projections. Lenders verify that the numbers are reproducible and the assumptions are defensible. For every projection in the packet: save the input assumption set (revenue per lane/hour, occupancy rate, pricing model, cost assumptions) as a separate named file alongside the projection. Date it. The test: given the assumptions file, can you reproduce the projection numbers? If yes, you have a compliance-grade projection. If no, you have a number.
2. Site analysis. For the Denver NC decision (7521 Eastern Medical Dr) — what's the record of the analysis? If SUMMIT did traffic analysis, demographic research, or competitive landscape work, that output needs to be saved as a dated artifact with human reviewer noted. Not the session — the artifact. Sessions are ephemeral; a lender can't review a session.
3. Human review attestation. Every AI-assisted document in the loan packet needs a review record: who reviewed it, when, what was changed. The attestation is: 'I have reviewed this document and attest to its accuracy.' This is the 'AI-assisted, human-verified' standard that holds up under scrutiny.
Critical note: the MHG SBA 7(a) loan is in-flight — it has not closed. Do not represent any numbers or commitments in this packet as confirmed without explicit CEO review. Any SBA or lender communication routes to SUMMIT then CEO before any response.
- operator
What's the right format for the assumption set on the Denver revenue projection? We're projecting 18 months from opening.
- agent
Here's the format. Save this as a named file —
MHG-Denver-Revenue-Assumptions-v1-YYYY-MM-DD.md— alongside the projection spreadsheet:MHG Denver Revenue Projection — Input Assumptions Version: 1.0 Date: [date] Prepared by: [name or 'SUMMIT assisted, TruPath Ventures reviewed'] Reviewer: TruPath Ventures Review date: [date] Status: DRAFT — not for external distribution without CEO sign-off Operating parameters: - Bay count: [N] golf simulator bays - Cornhole lanes: [N] (QC-instrumented at opening) - Operating hours per day: [N] hours - Operating days per year: 365 minus estimated closures Revenue assumptions: - Simulator bay: $[N]/hour avg, [N]% occupancy - Cornhole lane: $[N]/hour avg, [N]% occupancy - Food & beverage: $[N] per party avg - Private events: [N]/month avg, $[N] each Ramp assumptions: - Month 1-3: [N]% of steady-state occupancy - Month 4-6: [N]% of steady-state - Month 7-12: [N]% of steady-state - Month 13-18: steady-state Cost assumptions: - [line item]: $[N]/month - [line item]: $[N]/month Data sources: - [cite: comparable venue data, market research, DDI buildout estimate, etc.] Sensitivity notes: - [What would change this projection by more than 10%?]Fill in the numbers after Curtis Woodie and Olivia Merlock review the operating assumptions — they have the venue ops context. The projection is only as good as the assumptions, and the assumptions need human validation from people with venue experience, not just agent-generated estimates.
Audit trail template
Minimum viable audit trail for any AI-assisted document going to an external party. Copy the header block and append it to every compliance-trigger artifact.
COMPLIANCE AUDIT TRAIL HEADER
(attach to every AI-assisted document going external)
DOCUMENT TITLE: [descriptive name]
VERSION: [1.0, 1.1, etc.]
DATE CREATED: [YYYY-MM-DD]
PREPARED BY: [name — if AI-assisted, note: "SUMMIT assisted"]
AI ASSISTANCE: [yes/no — which agent or tool]
REVIEW RECORD
Reviewer: [full name]
Review date: [YYYY-MM-DD]
Changes made: [describe changes, or "none — accepted as drafted"]
Attestation: "I have reviewed this document and attest that
to the best of my knowledge it is accurate and complete."
Signature: ___________________ Date: ___________
INPUT SOURCES
(for projections and analyses only)
[ ] Source 1: [name, date, URL or file path]
[ ] Source 2: [name, date, URL or file path]
[ ] All assumptions saved in: [filename of assumption set]
REPRODUCTION TEST (for projections only)
[ ] Given the assumption set above, can the projection
numbers be reproduced? YES / NO
[ ] If NO: assumption set is incomplete — do not submit.
LIMITATIONS
[What does this document not cover?
What should the reader know about its scope?]
STATUS
[ ] DRAFT — internal only
[ ] REVIEWED — ready for CEO review
[ ] APPROVED — CEO approved, ready for external submission
[ ] SUBMITTED — sent to [recipient] on [date]
COMPLIANCE TRIGGERS PRESENT
[ ] SBA / Lender — requires CEO review before submission
[ ] Patent / IP — requires CIPHER review + attorney
[ ] Investor — requires CEO review and approval
[ ] Regulatory — verify current requirements before submit
Prompt kit
Three prompts. Completeness check, review attestation, session-to-artifact extraction. I keep mine in the CLAUDE.md compliance section.
Run a compliance audit trail completeness check on this
document or document set.
Context: [describe what you're preparing and for whom]
Recipient: [lender / regulator / investor / partner]
AI assistance used: [which documents were AI-assisted?]
For each AI-assisted document:
1. REVIEW RECORD: Is there a record of human review?
- Reviewer name, date, changes made, attestation?
- If not: who should review before this goes external?
2. INPUT PRESERVATION: For projections or analyses —
are the input assumptions saved as a separate artifact?
- Is the projection reproducible from the assumption set?
- If not: what assumptions are missing?
3. ARTIFACT VS SESSION: Is the output saved as a
dated, named artifact — or does it only exist in
a session that may not be recoverable?
- If session-only: extract and save before proceeding.
4. ATTRIBUTION: Is the AI-assisted status of each
document disclosed appropriately?
(Lender/regulator requirements vary — some require
disclosure, some don't. Check the specific requirement.)
Output: a gap report — which documents are compliance-ready
and which need work before external submission.Write a human review attestation for this AI-assisted document.
Document: [name and description]
Prepared with AI assistance: [yes/no, which agent/tool]
Reviewer: [name]
Review date: [date]
Changes made in review: [describe or "none — accepted as drafted"]
The attestation must state:
1. What document was reviewed
2. That the reviewer read the full document
3. What the reviewer attests to (accuracy? completeness?
methodology? all three?)
4. Any material limitations or qualifications
(e.g., "projections are estimates and may not reflect
actual performance" for financial documents)
Format: a short paragraph, suitable for filing
alongside the document. Sign with reviewer name and date.Extract and save this session output as a compliance-grade artifact.
Session topic: [describe what was produced]
Compliance trigger: [SBA / lender / patent / regulatory / investor]
Content to save: [paste or describe the session output]
Format the artifact as:
TITLE: [descriptive name]
DATE: [today]
PREPARED BY: [name or "SUMMIT/AXIOM assisted, [reviewer] reviewed"]
REVIEWER: [name]
REVIEW DATE: [date or "PENDING — must be reviewed before external use"]
STATUS: DRAFT or REVIEWED
VERSION: 1.0
[content]
INPUT SOURCES: [what data or documents were used?]
LIMITATIONS: [what does this analysis not cover?]
NEXT STEPS: [what needs to happen before this is final?]
Output: the formatted artifact, ready to save as
[TITLE]-v1-[DATE].mdApply this — audit one external document
~45-minute exercise. Tag, check, preserve, attest. Do it before the document goes external.
Your first compliance-grade AI artifact
Each step takes 10 minutes. Progress saves automatically.
- 01Identify every AI-assisted document in your current compliance-sensitive workflow. Tag each as 'reviewed' or 'needs review.'For MHG: the SBA loan packet, the Denver concept plan, the revenue projections, the DDI buildout scope. For QC: the patent package, the investor materials, the Crave NDA. Start with the most external-facing documents.
- 02Run the audit trail completeness check on one document that will go to an external party in the next 30 days.The check takes 10 minutes. Finding a gap at this stage is infinitely better than finding it during lender due diligence.
- 03Create an input assumption set for your most important AI-assisted financial projection. Test that the projection is reproducible from the assumptions.If you can't reproduce the projection from the assumptions, the assumptions are incomplete. The lender will find this faster than you think.
- 04Define the compliance-trigger topics for your CLAUDE.md: what subjects require a saved artifact before the session ends?For TruPath: SBA, lender contact, patent filing, investor materials, lease terms. Any session touching these produces a saved artifact — not just session history.
- 05Write the review attestation for one AI-assisted document that hasn't been reviewed yet.The act of writing the attestation forces the review. You can't attest to something you haven't read. That's the mechanism.