Most business arguments do not end because someone wins them. They end because someone gets tired, or louder, or cannot find the email. "Your agent promised me a refund." "No, it didn't." "It did — on Tuesday." Now you are digging through chat logs while a customer decides whether you are lying or just disorganized. Neither answer helps you. There is a third way to end that call, and it is deadpan on purpose: pull the record. What was asked, who authorized what, what the agent actually did, and what followed — read back, in order, from a record written at the time. Call that posture Receipts Mode. It is not a personality. It is what a governed chain sounds like when it answers the phone.
Governed AI agents: the dispute is now a regulated surface
Regulators have stopped treating the customer-service argument as a private matter. The New York City Department of Consumer and Worker Protection has proposed a rule for automated customer-service systems — AI chatbots, automated phone systems, and rules-based bots alike — that would require businesses to disclose at the outset that a consumer is dealing with an automated system, provide a specified means of reaching human assistance, and keep machine-readable records of covered interactions for at least three years (Digital Watch Observatory, Oct 28, 2026). The rule defines customer service to include refunds, billing matters, complaints, cancellations, fraud claims, and disputes, and it states that a business cannot avoid responsibility for unlawful conduct by attributing it to technology (Digital Watch Observatory, Oct 28, 2026). A public hearing is scheduled for November 19, 2026, with written comments due the same day (Digital Watch Observatory, Oct 28, 2026).
Read the three-year record requirement for what it is: the city is proposing that the record is the compliance. Not a promise to behave. Not a transcript you can summarize later. A machine-readable record, kept long enough for a dispute to surface and be settled against it. (Analysis: once a regulator specifies how long the record must live, "we think the agent said" stops being an answer.)
Custom CRM: a transcript is not a receipt
Here is the distinction most stacks miss. A transcript records what the model said. A dispute is about what the system did — and those are different files. Traceseal's August 2026 guide to AI agent audit trails puts it directly: chat transcripts are not an audit trail, because the questions that matter later are about actions — every tool call, what changed, the timing, and which steps ran under a human approval and who gave it (Traceseal, Aug 11, 2026). The same guide splits the problem in two: completeness — did you capture the right events — and credibility — can anyone other than you believe the record (Traceseal, Aug 11, 2026).
Credibility is where the deadpan comes from. A log written by the same system it describes, stored where the operator can edit it, cannot clear the operator when the operator is the party in question (Traceseal, Aug 11, 2026). Traceseal's proposed fix is mechanical: hash the artefacts, sign the record at execution time with a key held outside the agent's reach, and anchor it in a transparency log so the record provably existed before the dispute did (Traceseal, Aug 11, 2026). Its test is the one worth stealing: pick one agent action from last month and try to show a third party what happened without asking them to trust your database — if the exercise ends at "here is a row in our logging system," you have an audit trail, not evidence (Traceseal, Aug 11, 2026).
A custom CRM built for real operations has to pass that test on a phone call, in seconds. The customer's record, the authorization, the action, and the result — one chain, pullable while the customer is still on the line. That is the difference between arguing about Tuesday and reading Tuesday.
Bespoke software development: whose consent, whose record — October's court lesson
October also showed what happens when the record of permission is thinner than the record of the call. In In re Google Cloud Contact Center AI Privacy Litigation, Judge Rita F. Lin's October 22, 2026 order addressed callers to Home Depot whose calls were handled by Google's contact-center AI — a virtual agent that answered, transferred callers and transcripts to human representatives, and continued supplying transcription and suggested responses, as alleged in the complaint (Summit AI Notes, Oct 25, 2026, explaining the Oct 22, 2026 order). For two account holders, the court found the Home Depot terms they had accepted disclosed collection of customer-care recordings and sharing with service providers, so they had not plausibly alleged a lack of consent to Google's initial access (Summit AI Notes, Oct 25, 2026). For a third caller, Christopher Barulich, Google had not shown he assented to or was even aware of the policy, and the court rejected the idea that the common practice of recording customer-service calls was enough by itself to establish implied consent to sharing his call with an unannounced third party (Summit AI Notes, Oct 25, 2026). All claims were dismissed on a separate intent element, with leave to amend by November 12, 2026 — so the order is not a final ruling that the conduct was lawful in every respect (Summit AI Notes, Oct 25, 2026).
The operational lesson is not legal advice; it is record design. Consent turned on which caller, which terms, and what was disclosed — three facts a governed system should be able to produce per interaction, not reconstruct per lawsuit. Bespoke software development for voice agents has to treat permission as part of the record: who was told what, when, and what they agreed to, attached to the same chain as the action that followed.
AI agent governance: what Receipts Mode actually sounds like
So picture the call, with the chain underneath it. A customer says the agent promised a price, a refund, an exception. The voice agent does not debate, apologize into a discount, or improvise a new promise to end the discomfort. It pulls the record for that customer and reads the relevant entries back: the request as logged, the named human who approved or declined it, the action the system took, and the result. If the record shows the promise was made and authorized, the business honors it — the record cuts both ways, which is exactly why customers trust it. If the record shows no such authorization, the call is over as an argument and becomes a new request, handled on its own merits.
That restraint is becoming the industry pattern for money questions, even where the record is weak. Reporting in October 2026 described CarMax running 100% of inbound store and customer-experience-center calls through AI voice agents built with Sierra since May, per CEO Keith Barr on the company's September 29 earnings call — with the agents answering questions like store hours and stock, and passing the rest to associates (SmashOne, Oct 2026, summarizing PYMNTS reporting of Oct 2, 5 and 6, 2026). The same reporting's rule for small operators was blunt: anything about money — refunds, discounts, disputes — goes to a person (SmashOne, Oct 2026). And a vendor demo the same month showed both the promise and the limit of a calm voice: Rime's Prism launch film included an agent correctly declining to refund a $70 airline baggage charge it said was correctly applied — a promotional demo, not an independently verified transaction (SynthPeek, Oct 29, 2026).
(Analysis: those systems settle the call by routing it — the agent steps aside and a human decides. Receipts Mode settles it by reading it — the decision was already made, by a named human, under authorization, and the record is the decision. Routing is a safety valve. Reading is governance.)
One honest boundary: a record settles what happened; it does not make what happened right. If the chain shows your agent made a bad promise under a bad authorization, the record will prove that too, in the same deadpan voice. That is not a flaw in the design. That is the design — the record belongs to the facts, not to whoever would win the argument.
Phoenix custom software: four questions before your agent answers a dispute
If you are a Phoenix business owner shopping for custom software or governed AI agents that talk to customers, ask these before the first disputed call:
- Can your agent pull the record mid-call? If settling "you promised" means a human digging through logs tomorrow, you do not have receipts — you have homework. New York City's proposal would require keeping machine-readable interaction records for at least three years (Digital Watch Observatory, Oct 28, 2026); keeping them is step one, retrieving them on the call is the product.
- Does the record show actions, or just words? Transcripts record what the model said. Disputes turn on tool calls, changes, timing, and who approved what (Traceseal, Aug 11, 2026).
- Could the record have been edited after the fact? A log its own operator can quietly rewrite cannot clear that operator (Traceseal, Aug 11, 2026). Signing at execution time, with a key outside the agent's reach, is what turns a log into evidence.
- Is permission part of the same record? October's Google contact-center order turned on which caller accepted which disclosure (Summit AI Notes, Oct 25, 2026). If consent lives in a different system from the action, expect to reconstruct both under pressure.
A governed system makes these questions boring, because the chain was built to be read back. An ungoverned one makes the loudest person on the call the record.
Frequently asked questions
What is Receipts Mode?
It is the posture of a governed voice agent in a dispute: instead of arguing, escalating by default, or improvising a new promise, it pulls the audit record for that customer — request, authorization, action, result — and reads it back. If the record shows an authorized promise, the business honors it. If it does not, the argument ends and the customer's new request is handled on its merits.
Why is a call transcript not enough to settle a dispute?
Because a transcript records what the model said, while a dispute is about what the system did. Traceseal's August 2026 audit-trail guide argues a trail worth keeping records every tool call, what changed, timing, and human approvals — and that a complete log still needs credibility, since a record the operator can edit cannot clear the operator (Traceseal, Aug 11, 2026).
How long should customer-service interaction records be kept?
Regulators are starting to name numbers. New York City's proposed automated customer-service rule would require machine-readable records of covered interactions for at least three years, alongside disclosure that the consumer is dealing with an automated system and a specified route to human assistance (Digital Watch Observatory, Oct 28, 2026). The proposal is not final — a hearing is set for November 19, 2026 — but the direction is clear: retention is becoming a specified duty, not a preference.
Robbie Glover builds governed AI systems at Ronin Inc in Phoenix, Arizona. If you want a custom CRM where a disputed call ends with the record — request, authorization, action, and result, pullable on the call — talk to us.
