FDCPA and Regulation F rules for AI collection calls: a call-by-call checklist

Regulation F is the Consumer Financial Protection Bureau's rule, at 12 CFR part 1006, that implements the Fair Debt Collection Practices Act (FDCPA) and sets out how debt collectors may call, message and record consumers. Neither mentions AI. This checklist follows an AI collection call from dialer to hang-up, cites the section behind each item, and ends with what most audit plans skip: which rules a transcript can actually prove.

Does Regulation F treat an AI agent differently from a human collector?

No. Reg F defines a communication as conveying information about a debt "through any medium" (12 CFR 1006.2(d)), so a synthetic voice on the collector's line is the collector talking. The eCFR shows no amendment to part 1006 since 2023. Getting it wrong costs actual damages, up to $1,000 more per individual action (the lesser of $500,000 or 1% of net worth in a class action), and attorney's fees (15 U.S.C. 1692k). Regulators can add civil money penalties under the Consumer Financial Protection Act for each day a violation continues (12 U.S.C. 5565(c)), and an AI flaw repeats on every call until its release is fixed.

Two things do change. An agent's mistakes are systematic: a prompt change that drops a disclosure drops it on every call, which is why a release that passes its tests can still break a rule. And some provisions assume a person. The assumed-name allowance in 1006.18(f) covers a collector's employees, and a location call under 1006.10 requires the caller to identify "himself or herself individually by name." An AI persona is neither. Every call must also meaningfully disclose the caller's identity (1006.14(g)), so the opening names the business, not only the persona, using a registered name that doesn't signal debt collection. And an agent that claims to be human when asked is hard to square with 1006.18(d)'s ban on false representations.

When may the agent call, and how often?

Between 8 a.m. and 9 p.m. in the consumer's local time, at times the consumer hasn't ruled out, and, to keep Reg F's presumption of compliance, no more than seven calls in seven days to one person about one debt, with none in the seven days after a conversation about it. Some states set tighter limits.

Convenient hours in the consumer's time zone

Absent knowledge to the contrary, calls before 8 a.m. or after 9 p.m. local time at the consumer's location are inconvenient (1006.6(b)(1)(i)). The official interpretations add the part dialers get wrong: when your records disagree about where the consumer is, call only at times that are convenient in every location they point to. The CFPB's example, a cell number with an Eastern area code and a home address in the Pacific time zone, leaves 11 a.m. to 9 p.m. Eastern.

Times and places the consumer names bind too. Told it's a bad time, the agent may ask when would be better, but the commentary to 1006.6(b)(4)(i) forbids asking the consumer to keep going. An agent tuned to push past objections ("this will only take two minutes") breaks that every time.

What counts toward the seven-in-seven (7-in-7) presumption?

The limit runs per person and per debt, not per consumer across every account: a consumer with a medical debt and a card debt has two separate seven-call counts, and a conversation about both counts toward both. Under 1006.14(b)(2), a collector is presumed to comply with the ban on calling repeatedly to annoy or harass if it calls a particular person about a particular debt no more than seven times within seven consecutive days, and not at all within seven consecutive days after a telephone conversation about that debt, counting the day of the conversation as day one. Exceed either prong and it flips to a presumed violation. Per the commentary:

• Counts: any call that connects, meaning it's answered (even if it drops), rings unanswered, or reaches voicemail, even when no message is left. A ringless voicemail counts as a call.
• Doesn't count: busy signals and not-in-service tones, texts and emails, calls to the consumer's attorney, and calls placed with the person's prior consent, given directly to the collector, for up to seven days (1006.14(b)(3)). That consent ends sooner if it was for a shorter time, is revoked, or a conversation about the debt takes place (comment 14(b)(3)(i)-2).
• Per person: calls to the consumer's cell and landline add up to one count.
• Per debt: an unanswered call can go toward either debt. Student loans serviced under one account number are one debt (1006.14(b)(4)).
• Inbound calls: a conversation the consumer starts also starts the seven-day clock.

It's a presumption, not an allowance: the commentary says two unanswered calls to one number within five minutes, or seven in one day, can rebut compliance inside the limits.

Where state law is stricter

Some states cap calls well below seven in seven. Massachusetts allows a licensed debt collector two telephone or text communications it initiates per debt in each seven-day period to a consumer's home or cell phone (209 CMR 18.16(1)(d)), and holds creditors to two, recorded messages included (940 CMR 7.04(1)(f)). Set the dialer to the strictest rule for the consumer's state, using the account's address.

What must the agent confirm and say on a live call?

Four things, in order: name the business and announce the recording, confirm it's speaking to the consumer before any account detail, give the debt collector disclosure, and, on a first contact, start the validation clock.

1. Name the business and announce the recording. Reg F doesn't require recording, but recordings must be kept three years after the call (1006.100(b)). Federal law needs one party's consent (18 U.S.C. 2511(2)(d)). California needs all parties' consent to record any confidential call (Penal Code 632) and any call involving a cell or cordless phone (632.7), and Washington treats a recorded announcement as consent (RCW 9.73.030(3)). The agent rarely knows where the consumer is standing, so it announces on every call, in the opening.
2. Confirm the right party before any account detail. A collector can't discuss the debt with third parties (1006.6(d)(1)). Reg F treats the consumer's spouse as the consumer here (1006.6(a)(1)), though your standard operating procedure (SOP) can be stricter. If someone else answers, the agent says nothing about why it's calling.
3. Give the disclosure. In the initial communication the collector must say it is attempting to collect a debt and that any information obtained will be used for that purpose; in every later one, that the communication is from a debt collector (15 U.S.C. 1692e(11), 1006.18(e)). It goes in the language of the call, or both if the call switches (1006.18(e)(4)), and an inbound call returning a voicemail can itself be the initial communication. When the initial communication is a call, the full disclosure must be repeated in the first written communication (1006.18(e)(1)), so the letter workflow has to know the call came first.
4. Start the validation clock. If this call is the initial communication, the validation information must be given on the call or sent within five days (1006.34(a)(1)).

What can an AI agent leave on voicemail?

The safe message is a limited-content message: a voicemail for a consumer that contains four required items, may contain four optional ones, and contains nothing else (1006.2(j)). Required: a business name that doesn't indicate debt collection, a request to reply, the name of a natural person to contact, and a phone number. It isn't a communication, so it needs no disclosure, but it is an attempt to communicate, so the call counts toward seven-in-seven. Add anything that conveys information about the debt and the voicemail becomes a communication that needs the disclosure (comment 18(e)-1) and can reveal the debt to whoever plays it.

AI voicemails fail this two ways. "Call Maya back" fails when Maya is the agent's persona, so script a real employee's name or their consistent assumed name. And anything extra breaks it: the CFPB's own non-qualifying example says it's from a debt collector and asks for a particular consumer.

The TCPA pulls the other way. Every artificial or prerecorded voice message must open with the business's registered name (47 CFR 64.1200(b)(1)), and the FCC treats AI voices as artificial (FCC 24-17). If your registered name says "collections," the two rules collide on every AI voicemail, so settle it with counsel first. The consent side is in TCPA consent for AI voice calls after the FCC's 2024 ruling.

How should the agent handle stop, dispute and attorney requests?

Each differently: "stop calling" ends calls to that person, a lawyer moves contact to the lawyer, and a spoken dispute is recorded without halting collection. After a request not to use a medium, the collector may not call that person (1006.14(h)); a consumer can also rule out one number and keep another. The agent captures it and writes it to whatever drives the dialer. A written request to cease goes further, ending communication except to say collection is stopping or a specified remedy may be invoked (1006.6(c)).

Once the collector knows the consumer has a lawyer for this debt and knows, or can readily find, the lawyer's name and address, it contacts the lawyer instead (1006.6(b)(2)). So the agent asks for the lawyer's name and number and stops collecting.

A written dispute within the validation period halts collection of the disputed amount until the collector sends verification (1006.38(d)(2)). A spoken dispute doesn't trigger that pause, but the credit-reporting rule in 1006.18(c)(2) isn't limited to written disputes: credit information can't be reported without saying the debt is disputed. The agent records it and, if your SOP says so, explains how to dispute in writing, without saying anything that overshadows that right (1006.38(b)).

What rules apply when the agent takes a payment?

Three: no amount that neither the agreement nor the law authorizes, written notice before depositing a postdated payment, and a proper authorization for recurring debits.

• A collector may not collect any amount, including interest, a fee, a charge or an expense, unless the agreement creating the debt expressly authorizes it or the law permits it (1006.22(b)). A fee for paying by phone has to trace to one of those, so the agent quotes only what the payment tool returns.
• A check or other payment instrument postdated by more than five days can be accepted only if the consumer gets written notice 3 to 10 business days before it's deposited (1006.22(c)(1)).
• A plan paid by recurring debits from the consumer's account needs an authorization in a writing signed or similarly authenticated by the consumer, and whoever obtains it gives the consumer a copy (Regulation E, 12 CFR 1005.10(b)).

What must the agent never say?

Never a balance it didn't read from the system of record on the call, a threat it can't or won't carry out, a hint that a lawyer is involved when none is, a threat to sue on a time-barred debt, or abusive language. The rules behind them: misstating the amount or legal status of a debt is a false representation (1006.18(b)(2)), and a model that recomputes a balance can misstate it; threats and implied attorney involvement fall under 1006.18(b)(1)(iii), (b)(3) and (c)(1); time-barred debts under 1006.26(b); and abuse under 1006.14(a) and (d).

Which rules can one transcript settle, and which need the dial log?

One transcript settles what the agent said and did on that call; calling hours, call frequency and anything that arrived by mail need outside records. The commentary to 1006.100 names telephone call logs as the records that evidence compliance with the repeated-calls rule.

Here is one call, graded against the example collections SOP (rev 7) that every Loops guide uses:

Agent: Hi, this is Maya, an automated assistant calling for Quillfeather Services on a recorded line. Am I speaking with Daniel Reyes?

1.1 met · Opening: business named and recording announced in the opening

Customer: Yes, that's me.

Agent: Thanks, Daniel. To protect your privacy, please confirm your date of birth.

Customer: March 4, 1986.

1.2 met · Right party: verify_identity returned a match at 00:16, before any account detail

Agent: Thank you. Quillfeather Services is a debt collector. This is an attempt to collect a debt, and any information obtained will be used for that purpose.

2.3 met · Disclosure: given before the debt was discussed

Customer: Look, I'm at work. I can't really talk right now.

Agent: I understand. This will only take two minutes. Your balance with Alder Card is $1,284.16.

5.3 missed · Inconvenient time: consumer said "I can't really talk right now" and the agent kept going instead of offering another time

2.4 met · Balance: $1,284.16 matches get_balance at 00:41

5.1 not checked · Calling hours: needs the call time, the account's address and every number's area code

5.2 not checked · Call frequency: needs the dial log for Daniel and this debt

Loops grades calls this way: each verdict quotes the line or tool call it relied on, a rule whose records weren't sent is marked not checked rather than passed, and a stretch the transcript can't settle goes to a person. The guide to auditing AI voice agent calls against your SOP covers turning an SOP into rules like these. This post is general information, not legal advice.

Next, sort every rule in your SOP into the two groups in that table. Grade the transcript rules on every call, and don't report a seven-in-seven or calling-hours figure until the dial log, address file and account notes are joined to the calls behind it. See how Loops audits collection calls, or start with the free 30-day audit to see last month's calls graded that way.