Explainable AI for Borrowers: How to Force Models to Tell You What They Saw

AI for borrowers, if an automated model denies you or charges you a worse rate, you have power, but only if you ask for the right information in the right way. Most lenders hide behind “proprietary models.” That’s a stall, not a solution.

This guide gives the exact steps, scripts, and templates to force an explainable response, what to demand from the model team, how to act on the answers, and how to escalate if they dodge you.

The goal of AI for borrowers (crystal clear)

Get from the lender two things,

1. A concise, specific explanation of which factors caused the decision (not corporate PR).
2. A counterfactual or remediation checklist: “Change X, Y, Z and reapply/reevaluate.”

If you get that, you can fix errors, supply evidence, ask for manual review, or shop smarter. If you don’t get it, escalate.

What lenders should give you (the checklist of real explainability)

When you request an explanation, demand the following concrete items:

1. Top 3-5 factors the model used that most impacted the decision (e.g., “high DTI; recent NSF; device churn”).
2. A score or probability used in the decision (e.g., “Pd = 6.4%” or “Model score = 312 – threshold for auto-approve = 420”).
3. Which data sources produced those factors (credit bureau, bank feed vendor X, identity vendor Y).
4. The data snapshot used (date/time and the raw values the model saw).
5. Counterfactual explanation: the minimal change(s) that would have flipped the decision (e.g., “reduce credit utilization by 10% or add 2 months of payroll deposits”).
6. Model version & date – at least the model ID or version so you know what was used.
7. Whether a human reviewed or can review the decision and how to request manual review.

If the lender refuses to provide specific factors and data snapshots, treat their reply as weak and escalate.

Step-by-step playbook (do these in order – ruthless and precise)

Step 1 – Preserve evidence (right now)

Save everything: denial emails, screenshots of the offer, timestamps, copies of the application, and any supporting docs you uploaded. Do not delete anything.

Step 2 – Ask for the specific reasons (copy-paste email)

Use this exact template, don’t waffle.

Subject: Request for specific adverse action reasons and model explanation – [Full Name] – [Application ID / DOB / Last 4 SSN]

Hello,

I applied for [product] on [date] and received an adverse decision. Please provide, in writing:

1) The specific reasons that led to the adverse decision (top 3-5 actionable factors).
2) The consumer reporting or third-party vendors used (name and contact).
3) The data snapshot used for the decision (dates and raw values for the cited factors).
4) The numeric model output used (score or probability) and the threshold applied.
5) The model version or ID and the date it was run.
6) The minimal, actionable changes that would have changed the decision (counterfactuals).
7) Whether a manual review is available and how to request it.

Please respond in writing to this email within 14 days. If you withhold specific factors, explain the legal basis for refusal.

Thank you,
[Full name]
[Contact info]
[Application reference]

Step 3 – Expect one of three responses

• Good: They give top factors, data sources, and counterfactuals. Proceed to Step 4.
• Weak: Generic canned answer (“model score low”) – push back and repeat your request, citing the need for specific factors.
• Refuse: They cite proprietary rights, escalate (Step 6).

Step 4 – Verify & fix (action on the explanation)

If they give factors and data snapshots:

A. Check for errors. Does the snapshot show incorrect account balances, wrong employer, or stale data? If yes, dispute immediately with the vendor and the lender.
B. Follow the counterfactual. If they say “reduce utilization by 10%” or “add two months of payroll deposits,” act on it then reapply or request reevaluation.
C. Supply documentation. Email paystubs, bank screenshots, corrected vendor data, and request a manual re-underwrite pointing to those corrections.

Step 5 – Request manual re-evaluation (if you fixed issues)

Use this script,

“I have corrected/attached the following: [list]. Please perform a manual re-evaluation using this new evidence and confirm receipt. If you will not re-evaluate, please explain why and provide the escalation contact.”

Step 6 – Escalate if they stonewall

If the lender refuses to provide specifics or will not re-evaluate, escalate.

Options,

• File a formal complaint with the lender’s dispute team.
• File a complaint with your consumer protection regulator (describe the exact refusals).
• If the decision used a consumer credit report, obtain the report and dispute entries with the bureau/vendor.

(If you want specific regulator addresses or legal forms, get local legal/regulatory help, this guide is tactical, not a lawyer.)

Read: The Loan Pricing Engine: How Systems Decide If You Pay 9% Or 29% Interest

What to demand technically (for technically literate borrowers)

Ask for explainability formats you can actually use:

• Top-feature list (features + direction + weight). Example: “credit utilization: +0.42 impact; recent NSF: +0.30; device churn: +0.18.”
• Counterfactual statement: “If feature X were <= value Y, decision flips.”
• Score & threshold: numeric score and the threshold used for the decision.
• Data snapshot CSV: raw values for the features on the date used.
• Model version & training date: so you can know if the model is stale or recently retrained.

Companies may give some of this; push for as much detail as possible.

How to use the info once you have it (the tactical outcomes)

1. Fix errors (highest ROI)

If a vendor fed wrong data (wrong balance, wrong employer, duplicate negative tradeline), get it corrected. This often produces the fastest wins.

2. Make targeted changes (medium-term)

If counterfactual says “reduce utilization by 10%” – pay down cards. If it says “two more payroll deposits”, finish two pay cycles and reapply.

3. Negotiate (quick wins)

If the model penalizes you for device churn or channel, ask for manual pricing based on your corrected evidence: “I will set up autopay and provide proof, will you reprice?”

4. Change lender (practical)

If the model is a black box and won’t re-evaluate, shop different lenders, many use different models and data sources. Prequalify with soft pulls first.

Two powerful templates you’ll use repeatedly

1) Dispute to data vendor (short)

Subject: Dispute of data used in credit decision – [Name] – [Application ID]

I dispute the following item(s) in your dataset that were used in a credit decision by [lender]: [list inaccurate items]. Please provide your data snapshot, provenance, and correct/remove any errors.

Attached: documents proving correct values.

Please respond within 30 days.

Sincerely, [Name]

2) Escalation complaint (short)

Subject: Formal complaint – refusal to provide specific adverse-action factors – [Name]

I requested specific model factors and data snapshots on [date] and the lender refused / provided only generic reasons. This prevents me from correcting errors and exercising my consumer rights. Please investigate.

Attached: request copies and lender responses.

Thank you, [Name]

What lenders will claim and how to push back

• “Proprietary model” – response: you’re not asking for the model internals, only the top factors, data sources, and counterfactuals. That’s standard consumer request.
• “Too technical” – response: ask for plain-language versions of feature impacts and minimal-change guidance.
• “We can’t re-run model on new docs” – response: demand manual review and human-underwriter reconsideration.

Don’t accept high-level answers. Keep pushing.

Frequently Asked Questions

Final truth

AI models don’t have magical rights to secrecy. They make decisions out of math, and math can be explained, corrected, and contested. If you’re serious about getting fair treatment, don’t accept vague rejections.

Demand the top factors, the data snapshot, a counterfactual explanation, and a manual re-evaluation. Fix what’s broken, document everything, and never stop pushing until you get a clear path to remediation.

Read: Why Rural & Zip-Code Risk Still Matters And How To Beat Geographic Pricing