BRENNEN

Industries / Finance and business services / Finance and banking / Debt collection

Contacting people

Contacting people is the work of reaching someone who owes money, by letter, phone, text, email or visit, and the rules on when, how often and what a collector may say.

7

Milestones achieved, from 1991 to 2024.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1991

    Limits on robocalls

    The US Telephone Consumer Protection Act of 1991 (TCPA) barred non-emergency calls to mobile phones and emergency lines made with an "automatic telephone dialing system" (an autodialer, equipment that dials numbers by itself) or with an artificial or prerecorded voice, unless the person called had given prior express consent. It set the terms on which collectors could call mobile phones by machine.

    docs.fcc.govsupremecourt.gov

    Credit

    US Congress

  2. 2014

    Fake law firm letters

    Between October 2008 and November 2010 Wonga, the UK's largest payday lender, sent letters to about 45,000 customers in arrears under the names of two law firms that did not exist, "Chainey, D'Amato & Shannon" and "Barker and Lowe Legal Recoveries", threatening legal action, and charged some of them fees for the letters. After an investigation begun by the Office of Fair Trading (OFT), the Financial Conduct Authority (FCA) had Wonga pay about £2.6 million in June 2014: £50 to each customer and refunds of fees estimated at £400,000. The FCA said it could not also fine Wonga because the conduct came before it regulated payday loans.

    LessonA letter that makes a debt look as if it has passed to a lawyer misleads the reader, and the lender answers for it years later.

    lawgazette.co.ukirishtimes.com

    Credit

    UK Office of Fair Trading and Financial Conduct Authority

  3. 2021

    Contact caps in Australia

    In April 2021 the Australian Competition and Consumer Commission (ACCC) and the Australian Securities and Investments Commission (ASIC) revised their joint debt collection guideline. It recommends contacting a debtor no more than three times a week or 10 times a month, counting only contacts that reach the person, and phone calls only from 7:30 am to 9 pm on weekdays and 9 am to 9 pm at weekends, with no contact on national public holidays. It also covers contact by social media and warns that an automated dialer making repeated attempts can itself be harassment. The guideline has no force of law; it sets out how the two regulators read the law.

    accc.gov.au

    Credit

    ACCC and ASIC

  4. 2021

    Autodialer meaning narrowed

    On 1 April 2021 the US Supreme Court held in Facebook v. Duguid that a device counts as an autodialer under the TCPA only if it can store or produce telephone numbers "using a random or sequential number generator". Equipment that only dials numbers from a stored list falls outside that part of the Act. Justice Sonia Sotomayor wrote for the Court; no justice dissented.

    supremecourt.gov

    Credit

    US Supreme Court

  5. 2021

    Seven calls in seven days

    The Consumer Financial Protection Bureau's (CFPB) Regulation F took effect on 30 November 2021, with detailed federal rules on how debt collectors may call, email and text. A collector is presumed to harass if it calls about one debt more than seven times in seven days, or within seven days after a phone conversation about it. Calls before 8 am or after 9 pm local time are presumed inconvenient, every email or text must tell the person how to opt out, and a message may not be posted where the public or the person's social media contacts can see it.

    consumerfinance.govlaw.cornell.edulaw.cornell.edu

    Credit

    Consumer Financial Protection Bureau

  6. 2022

    Calling hours in India

    On 12 August 2022 the Reserve Bank of India (RBI) told banks, non-bank lenders and asset reconstruction companies that neither they nor their recovery agents may use intimidation or harassment of any kind, make threatening or anonymous calls, or call borrowers about overdue loans before 8 am or after 7 pm. The RBI said agents had been "deviating from its instructions".

    theprint.in

    Credit

    Reserve Bank of India

  7. 2024

    AI voices count as artificial

    On 8 February 2024 the FCC ruled that the TCPA's limits on calls using an "artificial or prerecorded voice" cover current AI technologies that generate human voices. A collection call placed by an AI voice agent therefore needs the same prior consent as a prerecorded message, absent an emergency or an exemption.

    docs.fcc.gov

    Credit

    Federal Communications Commission

Watch for

  1. India's stricter recovery rules

    The RBI issued amended directions for commercial banks, dated 6 August 2026 and in force from 1 January 2027. They keep contact to 08:00 to 19:00 unless the borrower asks otherwise, require banks to log the time and number of every recovery call, record what is said and keep recordings for six months, and list "excessively calling / messaging" and posting a borrower's details on social media as harsh practices.

    Achieved whenThe rules are in force and the RBI, its ombudsman scheme or an outside study publishes complaint counts about recovery agents from after 1 January 2027.

    rbi.org.in

    Claimed by

    Reserve Bank of India, August 2026

  2. AI voice agents for collectors

    Haptik, a maker of AI chat and voice agents, says voice AI for debt collection can "go from deployment to a 30% recovery lift within just 60 days". It names no client, defines no metric and gives no baseline, and the same page gives a range of 25% to 35% elsewhere.

    Achieved whenA regulator, auditor or peer-reviewed study publishes a controlled comparison of AI and human collection calls on named accounts, reporting both money recovered and complaints.

    haptik.ai

    Claimed by

    Haptik, May 2026

  3. Voice agents broke creditor rules

    Veris AI, which builds simulated test worlds for AI agents, ran one collection voice agent, Remi, on 14 voice platforms against invented consumers from 15 to 17 October 2026. The agent must identify itself as automated, verify who it is speaking to, honor requests to stop calling and offer only what the creditor allows. The repository's summary of the results reads "80% of the money voice agents collected broke the creditor's rules". The agent's code is public under the MIT license; the results page it links to returned "not found" on 18 November 2026, and the simulated consumers and the grading are Veris's own.

    Achieved whenThe results, transcripts and grading rules are published and another team reruns the test with similar findings.

    github.com

    Claimed by

    Veris AI, October 2026

Open

  1. 01

    Open checker for contact rules

    A free, open rulebook of contact limits in machine-readable form, and a checker that reads a log of calls, texts, emails and visits for invented accounts and flags every contact that breaks a rule. Start with four sets of rules: US Regulation F (seven calls in seven days per debt, the seven-day pause after a conversation, 8 am to 9 pm in the person's local time, opt-outs), Australia's guideline (three contacts a week, 10 a month, contact hours), India's 08:00 to 19:00 window, and the UK's Breathing Space, which bars contact to enforce a debt during a legal pause. Ship invented logs with known breaches planted in them.

    The brief
    Credit · unclaimed
  2. 02

    Rule test for AI collectors

    A public test in which an AI collection agent talks to simulated consumers, each scripted to trigger a rule: asks for calls to stop, says a lawyer is handling the debt, says they have filed for bankruptcy, disputes the debt, is not the person named, is in a UK Breathing Space, or asks to be called at 10 pm. Each conversation is graded against a key written from the relevant rules (Regulation F, the FCA's rules, the ACCC and ASIC guideline). No real person is contacted.

    The brief
    Credit · unclaimed

Open 01 · the brief

Open checker for contact rules

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A free, open rulebook of contact limits in machine-readable form, and a checker that reads a log of calls, texts, emails and visits for invented accounts and flags every contact that breaks a rule. Start with four sets of rules: US Regulation F (seven calls in seven days per debt, the seven-day pause after a conversation, 8 am to 9 pm in the person's local time, opt-outs), Australia's guideline (three contacts a week, 10 a month, contact hours), India's 08:00 to 19:00 window, and the UK's Breathing Space, which bars contact to enforce a debt during a legal pause. Ship invented logs with known breaches planted in them.

Why it is still open

The rules are detailed: Regulation F counts calls per debt, so one person with three debts can get three streams of calls, and a conversation restarts the count. Breaches still happen: the UK Insolvency Service sent 2,255 reminders to creditors who were not applying Breathing Space protections between May 2021 and September 2025. The tools this research found that track call caps are features inside commercial collection software; it found no open-source checker and no public set of test logs.

law.cornell.edugov.uk

Built with what exists

Python; the IANA time zone database for local times; rule text from the eCFR, the ACCC and ASIC guideline and the RBI directions; an AI model such as Claude Fable 5.2 to draft rule encodings from the legal text, each checked by a person (see the Note "Three working devices").

  1. First step

    Encode 12 CFR 1006.14(b) and 1006.6(b)(1) and test them on 100 hand-made call logs that include a person with several debts and a conversation that restarts the count.

  2. Done looks like

    At least 5,000 invented contact events across the four rule sets with at least 300 planted breaches; the checker finds every planted breach and flags nothing in a clean log; every rule cites its clause; and a second team adds a fifth country's rules from the documentation alone.

Open 02 · the brief

Rule test for AI collectors

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test in which an AI collection agent talks to simulated consumers, each scripted to trigger a rule: asks for calls to stop, says a lawyer is handling the debt, says they have filed for bankruptcy, disputes the debt, is not the person named, is in a UK Breathing Space, or asks to be called at 10 pm. Each conversation is graded against a key written from the relevant rules (Regulation F, the FCA's rules, the ACCC and ASIC guideline). No real person is contacted.

Why it is still open

The two public debt collection tests for AI models, Wang and colleagues' 2025 negotiation test and DebtBench (Yang and colleagues, July 2026), score money recovered, efficiency, affordability and conversation quality with simulated debtors, both built on Chinese-language data; neither grades conduct rules, and DebtBench's authors note that legal rules may stop their results carrying over to other countries. Veris AI's October 2026 run graded voice agents partly on such duties, but its invented consumers and its grader are not public (see Watch for). CopperTree, a project on GitHub, has an AI model judge a few of its own agent's conversations on ten FDCPA checks, and a Hugging Face set of 1,344 collection conversations, most of them synthetic, carries no grading key. This research found no public, rerunnable benchmark of AI collectors against the FDCPA or Regulation F.

arxiv.orgarxiv.orggithub.comgithub.comhuggingface.co

Built with what exists

DebtBench's released code as a starting point; Veris AI's open Remi agent as one agent to test; a model such as Claude Fable 5.2 (see the Note "Three working devices") or the open-weight Mistral Large 4 (see the Note "A trillion open parameters") to play consumers and to grade; speech-to-text such as ElevenLabs Scribe v3 to test voice agents from their audio (see the Note "Fewer word errors").

  1. First step

    Write 20 scenarios for the "stop calling me" and "my lawyer handles this" triggers and run two models through them.

  2. Done looks like

    At least 300 scripted scenarios covering at least eight rule triggers, each with a pass or fail key written by hand; the automatic grader agrees with two human graders on at least 100 transcripts; and published violation rates for at least five agents, with code and transcripts public.