Pre-Qualifying Phone Leads With an AI Assistant: A Lender's Walkthrough

$7.50, under a minute, and one sentence of consent. That’s what AI loan prequalification costs on a live call when a rep has an assistant connected to our MCP server: the caller says yes to a soft pull, the rep runs one prompt with the phone number and the offer, and the assistant comes back with the caller’s name, address, FICO® Score 4, the factors behind it and a read on the phone line. If nobody’s found, it costs nothing.

What the assistant doesn’t do is decide. Your written score bands decide, the rep says the words, and the same lending rules apply as if the rep had typed the number into a web form. This walkthrough is for lenders and dealers who take financing calls and want the pre-qual done before the caller has time to get bored.

The call, in numbers

Here’s the shape of one call, using a fictional home improvement lender, Tamarack Home Improvement Finance, and a fictional rep, Gus Ostrowski.

Scale it up. Say Tamarack takes 300 financing calls a month, 260 callers agree to the check and 235 of those are found. That’s 235 × $7.50 = $1,762.50, plus the API plan at $49 a month (or $149 a year). If Gus and his team open tradelines on 40 borderline files, add $80. Call it $1,890 for the month. If 45 of those callers end up funding, the pre-qual cost is about $42 per funded loan. Your margin on a $16,000 installment loan will tell you whether that’s cheap. For most lenders it is.

These are example figures, not ours. Plug in your own call volume and close rate.

What the prompt does on an AI loan prequalification call

Our MCP server ships with a prompt called prequalify_caller. It takes two inputs: the caller’s phone number (required) and the offer you have in mind (optional, something like “$16,000 home improvement loan, 60 months”). Assistants that support MCP prompts usually list them next to the connector’s tools; in some apps it’s a menu item, in others a slash command.

When Gus runs it, the assistant receives a short instruction that says, in plain words: the caller at this number has agreed to a credit check for a credit application, look them up with the purpose credit_application, then summarize who they are, the score and its main reasons, card usage, on-time payments, collections and the fraud risk on the phone line, plus whether they fit the offer. It also says not to pull tradelines unless asked.

Read that first clause again. The prompt states that the caller agreed. Running it is the rep vouching for that. So the rule on the floor is simple: no OK, no prompt. If you’re new to how the purpose field works, our permissible purpose guide and the FCRA text itself cover it.

Here’s roughly what Gus saw for a fictional caller, Marisol Quintero, (555) 555-0173. The assistant’s wording will vary from one app to the next; the data won’t.

Marisol Quintero, Tucson, AZ (matches the name she gave). Mobile line, major carrier, fraud score 8. FICO Score 4: 688. Main factors: balances on revolving accounts are high compared with limits; length of time accounts have been open. Card usage 47%, 100% on time, no collections, 2 inquiries in the past year. Against your table, that’s Tier 2. A $16,000 loan fits under the $20,000 cap. I didn’t pull tradelines.

Before the lookup ran, the assistant also told him it would cost $7.50 if she was found. Our server’s instructions ask assistants to name the price before a paid call and to reuse a past lookup instead of paying twice, which is why it’s worth saving the lookup ID (more on that below).

Score bands to offers: give the assistant your table

An assistant knows your offer menu only if you tell it. Paste your band table into the conversation, or into whatever standing instructions your assistant supports, and ask it to name the band, not invent one. Here’s Tamarack’s, built for unsecured home improvement loans of $5,000 to $25,000. It’s illustrative; set yours from your own credit policy.

Three non-score checks run before the band counts:

• Name match. If the name on the file isn’t the person on the phone, stop. A different name is either a shared family line or someone using another person’s number.
• Fraud score. Tamarack sends anything 70 or higher to review, whatever the credit score.
• Address. If the caller is financing a roof, the address on file should plausibly be the house with the roof.

Notice what the table doesn’t contain: ZIP codes, names, ages or anything else that could stand in for a protected trait. Our post on loan lead scoring has a full point model if you want something richer than bands, and pre-qualifying leads with a phone number covers the non-AI version of this same workflow.

Soft pull wording that holds up

The words matter twice: once before the lookup, once after.

Before. Gus’s script at Tamarack:

“To see what you’d likely qualify for, I’d like to run a soft credit check through TransUnion. It won’t affect your credit score and other lenders won’t see it. Is that OK?”

He logs the yes with the time and the script version. Our consent form guide has wording for web and text, and the soft pull vs. hard pull explainer is worth sending reps who get asked “will this hurt my credit?” more than twice a day.

After. This is where reps drift. “You’re approved” is wrong. “You’re pre-approved” is usually wrong too, unless your program actually issues written commitments. What Gus says:

“Based on that, you look like a fit for up to $20,000. That’s a pre-qualification, not an approval. The full application comes next, and it may include a hard inquiry.”

One line we’d ban outright: “We didn’t really check your credit.” You did. It was soft, but it was a check, and the caller is entitled to know that.

What goes in the CRM

The assistant’s chat is not your system of record. Paste the summary into notes if you like, but these fields belong in structured columns:

Leave out what nobody needs: full account numbers from tradelines, the caller’s whole profile pasted into a free-text note, or a chat transcript shared to a team channel. Fewer copies of credit data means fewer places for it to leak. If you’re already routing leads by tier automatically, the webhook routing build shows how the same fields land in a CRM without anyone typing.

Separately: consent to check credit isn’t consent to text or autodial. If Gus wants to send Marisol the application link by text, that’s its own permission, covered in our TCPA guide for calling and texting leads.

Adverse action still applies

The assistant wrote the summary. Your company used the report. That’s who the law looks at.

Regulation B. A pre-qualification request can turn into an application without anyone filling out a form. The CFPB’s official commentary on what counts as an application says that if a creditor evaluates information about the consumer, decides to decline, and tells them, it has treated the inquiry as an application. Then the notice rules in 12 CFR 1002.9 kick in: generally 30 days to notify, with specific reasons. A smaller counteroffer the caller doesn’t take has its own clock (90 days).

FCRA section 615. When adverse action is based even partly on a consumer report, 15 U.S.C. § 1681m requires a notice that names the bureau, includes the score used and its key factors, and explains the right to a free report and to dispute.

For Tamarack, that means the “Below 580” row and some “Review” outcomes need a notice. So can Tier 3, when the caller wanted $16,000, you offered $12,000 and they walked away. Two practical points for AI calls specifically:

• The score factors aren’t your reasons. The assistant will happily repeat the factors that came back with the score. Your ECOA reasons are why you said no, which might also be income or loan size. Write them from your table, not from the chat.
• Don’t let the model phrase the decline. A friendly paraphrase on a live call can contradict the reasons in the notice you mail later. The rep says the decision; the notice template says the rest.

Our walkthrough of the ECOA adverse action notice goes field by field, and the wider question of FCRA rules when AI agents pull reports gets its own post. Usual caveat: this is practical information, not legal advice. Your counsel should see the script and the table before the first live call.

Setting it up for a team

Connecting takes a few minutes. Paste https://phonenumbercreditscore.com/mcp into the assistant as a custom connector, sign in on our page, and choose what that connection may do. Our MCP server announcement explains the pieces, the Model Context Protocol introduction explains the standard, and if your team uses Claude, the step-by-step for Claude walks through the clicks.

For a sales floor, the choices on our consent page that matter:

• Permissions. Lookups and reading past lookups. Leave tradelines off for most reps and let a manager open them.
• Test mode first. Test connections are free and return sample people. Numbers ending in 0422 come back as a thin file, 0423 as frozen and 0404 as no match, any area code, which is exactly the set of awkward calls a new rep should rehearse.
• How long it lasts. 30 days for a trial floor; longer once it’s settled.

Only API accounts can connect, and only people with the “Connect apps” permission (owners always can). Each connection shows up on API / Developer, then Apps & AI, where you can disconnect it immediately. Every call lands in your usage logs with the app’s name, and the account-wide spending limit applies to assistant lookups the same as API ones. Set it before the floor goes live.

One last habit worth building: every Friday, pull ten Tier 2 and Tier 3 calls at random and compare the band the rep quoted with the score in the CRM. If they match, good. If the assistant’s summary and the table ever disagree, the table wins, and you’ve found the call to coach. Full pricing and the tool list are on our API page.