The CDFI AI Evidence File Checklist

Why an AI evidence file matters for mission lenders

Your CDFI certification, awards, and impact reporting rest on a chain of attestations to the CDFI Fund and, ultimately, to the communities you serve. When AI assists the drafting or analysis behind those attestations, an evidence file proves a human owned the judgment — protecting both your certification and your mission credibility. This is a mission-integrity control, not a tech disclosure: the question a reviewer or examiner asks is "who decided, and on what basis?"

• Scope it to documents that feed obligations: the ACR, the TLR, Performance Progress Reports for applicable awards, impact narratives, and community-development outreach summaries.
• Recommended: keep one lightweight evidence row per AI-assisted deliverable — enough to reconstruct what AI touched and who verified it — rather than logging every keystroke.
• Suggested file language: "This evidence file records where AI tools assisted in preparing materials submitted to or relied upon for the CDFI Fund and for community-impact reporting. AI assistance does not replace the certifying official’s independent review; all figures, claims, and certifications reflect human verification."
• Not legal advice — confirm specifics against your Award/Assistance/Allocation Agreement and your award reporting instructions.

AI Evidence Row Template

Use one row per AI-assisted deliverable. The ledger should be simple enough to complete during normal file preparation and detailed enough to show source records, human review, verification, and reuse limits.

Build (and reuse) an AI use-case inventory

Before logging individual documents, list the AI uses you actually permit in mission work. A short inventory turns ad hoc tool use into governed, repeatable practice and gives you a stable vocabulary for every downstream evidence row. The AIEOG Lexicon frames AI governance as the policies, roles, and oversight that direct how AI is adopted and monitored.

• Recommended default prohibitions for mission lenders: no AI-generated eligibility or adverse-action reasoning, no synthetic statistics in impact claims, and no AI text inserted into a certification without a named reviewer’s sign-off.

1. List each approved AI use case in plain language (e.g., "draft impact-narrative first version," "summarize small-business outreach notes," "check ACR narrative for internal consistency").
2. For each, record the tool/model, the data category it may touch, and the explicit prohibition line (e.g., "no borrower PII," "no protected-class inferences," "no auto-generated certification language submitted without review").
3. Assign a human owner accountable for each use case and the verification step required before the output is used.
4. Note where outputs land (ACR narrative, PPR, impact report, outreach summary) so the inventory cross-references your evidence rows.
5. Review the inventory at least annually — align the cycle with your CDFI Fund reporting calendar so it is current when attestations are due.

Separate AI assistance from human judgment in the record

The single most valuable thing your file does is draw a clean line between what AI produced and what a person decided — the same principle CFPB enforces in lending (an institution cannot hide behind a model), applied to documentation.

• Default rule: every AI-assisted deliverable carries a one-line attribution note naming the AI role, the human reviewer, and the verification performed.
• Suggested file language: "Portions of this document were drafted with AI assistance. All data, eligibility determinations, impact claims, and certifications were independently reviewed and verified by [name/title] on [date]. The institution takes full responsibility for the accuracy of the final content."
• Distinguish three AI roles so reviewers know the stakes: drafting (low — language only), summarizing (medium — verify against sources), and analysis/recommendation (high — requires documented human re-derivation).
• Never let AI generate the certification or attestation language itself; the certifying official’s words and accountability must be human-authored.

Fair-lending and access guardrails (ECOA / Regulation B)

AI-assisted mission-lending documentation should preserve fair-lending and access integrity. For credit-related use, AI may not determine eligibility, pricing, adverse-action reasons, or certification language. Any protected-basis, demographic, Target Market, or proxy-variable data used for reporting or monitoring must be handled under approved reporting, compliance, and privacy controls — not casually introduced into AI prompts.

• Recommended: AI is permitted for drafting and summarizing, but credit decisions, eligibility calls, pricing, and adverse-action reasons must be human-determined and independently documented.
• For credit-adjacent work, reasons must be specific, accurate, and traceable to the file; AI may not invent, infer, rank, or rewrite the principal reasons actually determined by a qualified human reviewer.
• Log a fair-lending/access check on any AI-assisted analytical output: confirm protected-basis, Target Market, demographic, and proxy-variable data were used only under approved reporting, compliance, and privacy controls.
• Adaptable clause language: "No AI output was used to determine applicant eligibility, pricing, adverse-action reasons, or certification language. AI assistance was limited to drafting, summarization, consistency checking, or formatting QA of human-verified content."

Map AI evidence to CDFI Fund reporting instruments

Your obligations flow through named instruments, AMIS workflows, and your individual agreement; AI-assisted preparation of any of them belongs in the evidence file. CDFI Fund website templates can be useful references, but the latest templates and instructions should be pulled from AMIS and your award documents.

• ACR (Annual Certification and Data Collection Report): Certified CDFIs must submit the ACR to maintain certification. If AI helped draft narrative responses, log the row and confirm the certifying official reviewed every certification statement.
• TLR (Transaction Level Report): the TLR applies to new applicants and currently certified CDFIs; the version depends on award status, and active Assistance Agreement recipients continue the full-length TLR. AI must not generate or alter transaction-level data — if used at all (e.g., formatting QA), record it explicitly and verify against source systems.
• Performance Progress Report (PPR): active recipients generally submit the PPR in AMIS annually three months after fiscal year end; the CDFI Fund uses it to determine compliance with applicable Performance Goals in the Assistance Agreement.
• Material Events: active award or allocation recipients submit the Material Events form and supporting documentation within 30 days or as specified in the Award, Assistance, or Allocation Agreement. If AI assisted in drafting a notice, document human verification of the underlying facts and deadline.
• Recommended: for anything submitted to the CDFI Fund, the evidence row must name a human certifier; when an instrument or deadline is uncertain, defer to your grant agreement and reporting instructions and the CCME Help Desk rather than assuming.

Worked example — AI-assisted annual impact narrative

Here is the practice in miniature. Use this filled row as your template.

• Scenario: the Director of Impact uses an approved AI tool to turn de-identified, aggregate small-business lending data and field notes into a first-draft impact narrative for the certified Investment Area, plus a one-paragraph outreach summary from staff notes.
• Verification performed: the Director reconciled every figure to the loan-origination system, removed two AI phrasings that overstated job-creation outcomes, confirmed no borrower-level or protected-class data was used, and the certifying official independently approved the final narrative.
• Sample filled evidence row — Document: FY2025 Annual Impact Narrative (supports ACR) | AI role: first-draft drafting + summarization | Tool: [approved internal AI assistant] | Data: de-identified aggregate program data; staff field notes | AI prohibited from: borrower PII, certification language, synthetic statistics | Reviewer: Director of Impact | Verification: figure-to-source reconciliation; overstatement removed; fair-lending/access check passed | Certifying sign-off: CEO, 2026-03-04 | Retention: [X years] under the applicable agreement, Uniform Guidance record-retention rules, institution retention schedule, and any audit, litigation, or agency hold.

Retention, access, and audit-readiness

Evidence is only useful if it survives staff turnover and can be produced on request. Tie retention to your reporting cycle and keep the file simple enough that anyone can reconstruct what happened.

• Retention: [X years] under the applicable Award / Assistance / Allocation Agreement, Uniform Guidance record-retention rules, institution retention schedule, and any audit, litigation, or agency hold.
• Store evidence rows alongside the submitted instrument (ACR/TLR/PPR/impact report) so a reviewer can move from attestation to verification in one step.
• Control access: restrict the file to staff with a need to know, and never store borrower PII or protected-class data in it.
• Suggested file language: "AI evidence records for this reporting period are retained under the applicable Award / Assistance / Allocation Agreement, Uniform Guidance record-retention rules, institution retention schedule, and any audit, litigation, or agency hold, and are available to authorized reviewers and examiners upon request."

What good looks like / common mistakes

A strong AI evidence file reads like a mission-integrity ledger: clear human ownership, clean separation of AI from judgment, and no fair-lending or access blind spots. Most failures come from over- or under-documenting in the wrong places.

• What good looks like: every AI-assisted submission has a named human certifier, AI’s role is described in one honest line, figures are reconciled to source, and fair-lending/access checks are recorded.
• Common mistake (transparency): omitting AI involvement entirely, or burying it so a reviewer cannot tell what AI touched.
• Common mistake (fair lending): letting AI generate or influence eligibility or adverse-action reasoning, then being unable to give a specific, accurate, human-verifiable reason.
• Common mistake (access/accuracy): letting AI inflate impact or reach figures without reconciling to source data.
• Common mistake (over-documentation): storing borrower PII or protected-class data in the evidence file itself — keep the ledger about the process, not the people.

The CDFI AI Evidence File Kit

Use this checklist as the front door to a practical kit: the checklist, an editable evidence ledger, a CDFI reporting-instrument map, a source-reconciliation log, a fair-lending/access guardrail sheet, and a certifier sign-off page.

• Next step: Run the 30-minute CDFI AI Evidence File Review.
• Pick one AI-assisted grant, certification, or impact deliverable; identify what AI touched; reconcile claims to source records; assign the human certifier; and create the first evidence row.