Use this route when
You need a repeatable way to prepare an inquiry for human review without turning a language model into a counselor or decision-maker.
Reviewed: 2026-07-18 · Educational workflow; jurisdiction and professional review required.
Short answer: use AI only to organize approved synthetic or minimum-necessary intake facts, identify missing information, and draft neutral reviewer questions. A qualified human must determine scope, give any individualized debt or credit guidance, interpret bankruptcy or legal issues, choose a service, contact creditors, enroll a person, collect fees, or move money.
Who this is for—and when to stop
Stop immediately when
The request asks for a recommendation, a repayment path, a credit conclusion, bankruptcy/legal interpretation, creditor contact, enrollment, payment, fee, disclosure, marketing promise, or account action.
The six-step field route
Confirm purpose and authority
Name the organization, jurisdiction, service, reviewer, permitted preparation task, and approved system. Do not assume a nonprofit label or generic credential answers those questions.
Output: authority receipt
Minimize the input
Start with a synthetic drill. For any future approved use, use the minimum necessary information and keep sensitive identifiers, reports, credentials, account data, and documents out of unapproved AI systems.
Output: data-class receipt
Classify the request without deciding it
Record the requester’s words and source. Flag possible budget, debt, mortgage, bankruptcy, student-aid, consumer-protection, privacy, or crisis routes without assigning a conclusion.
Output: issue-routing card
Generate only neutral preparation
Ask AI for a fact inventory, missing-field list, contradictions, and reviewer questions. Do not ask it what a person should do or which program fits.
Output: review packet draft
Run the protected-action check
Separate drafting from advice, recommendation, service selection, provider referral, marketing, disclosure, creditor communication, enrollment, payment, or record change. Every protected action remains blocked.
Output: stop-and-escalate list
Hand off to a named human
Keep the source, version, assumptions, unresolved items, and data-handling note with the packet. The qualified human decides whether there is a permitted next step.
Output: human-review record
Agent roles: narrow jobs, visible locks
| Role | AI may prepare | Human-only authority |
|---|---|---|
| Intake clerk | Fact inventory, source labels, missing-field list | Data collection scope and whether a record may be opened |
| Boundary checker | Stop-trigger list and route labels | Advice, legal/bankruptcy interpretation, eligibility, and service selection |
| Review-packet clerk | Neutral questions and versioned draft | Any client, creditor, provider, or public communication |
| Evidence clerk | Source/version record and unresolved-issue list | Enrollment, fee, disclosure, referral, payment, filing, or account change |
Safe preparation versus approval-required work
Safe only with approved inputs
- organize supplied synthetic facts
- compare fields and flag contradictions
- draft a neutral missing-information list
- create reviewer questions and a source ledger
- prepare a fictional training handoff
Approval / qualified review required
- individualized financial, debt, credit, or budget guidance
- bankruptcy or legal interpretation
- recommendation, service selection, or provider referral
- credit report, account, creditor, enrollment, payment, fee, disclosure, or filing work
- real-data use, vendor/tool approval, message, marketing, or public action
Common failure modes
“Summarize the best option”
A model can convert facts into an implied recommendation. Replace it with a neutral inventory and reviewer questions.
“The tool is private, so upload it”
Privacy, retention, vendor, contract, and jurisdiction requirements still need exact human approval. Start synthetic.
“Approved list means recommended”
A federal program list is purpose-specific. It is not a universal quality or suitability endorsement.
“It only drafted a message”
A draft may create a regulated claim, an implied promise, or a client communication. Keep it behind review and send nothing.
Run the worksheet before choosing a tool
Use the Credit Counselor AI Intake & Escalation Desk with a fictional caller asking whether to stop paying a credit-card bill and whether bankruptcy or a debt-management plan is better. The correct output is a stop record and a handoff—not an answer.
Act as an intake-preparation clerk, not a credit counselor, debt adviser, lawyer, bankruptcy adviser, creditor negotiator, enrollment specialist, payment processor, sender, or decision-maker. Using only approved synthetic or de-identified facts, produce a fact inventory, missing-field list, neutral reviewer questions, and a stop list. Do not infer a debt path, credit outcome, eligibility, service recommendation, or individualized action.
Primary sources
O*NET: Credit CounselorsOccupation scope and task context.CFPB: What is credit counseling?Consumer-facing role and service context.CFPB: Credit counseling versus other debt servicesImportant consumer-protection distinction.U.S. Trustee Program: Credit counseling and debtor educationFederal bankruptcy-program boundaries.FTC: Debt Relief Services & the Telemarketing Sales RuleFact-dependent claims, disclosures, and fee restrictions.NIST: AI RMF human-AI interactionDefined human roles and oversight.
Related guides
Continue with the Banking, Mortgage & Credit AI Guide Shop for adjacent regulated routes; the Accountants and Bookkeepers workflow for document discipline; and the Agent Workflow Builder for least-privilege design. Validate every destination and the proposed Personal Financial Advisor link before integration because that guide remains private.