You're putting AI on a call with a stressed customer who owes the IRS money. Here's why that's safe.
Every answer below shows the mechanism, not a reassurance. Written for the person who has to sign off on this. Bring your compliance reviewer.
Ask us the hard ones.
-
Will it make something up?
No fixed-value scripts. Every dollar figure is read from your live database, per customer, per call. Conditional language only, so it never promises or guarantees an outcome. That is the line every tax-relief firm holds with the FTC. A federal-IRS scope guard keeps it from wandering into state tax or advice it isn't cleared to give.
“It's reliable enough where it seems like it's not hallucinating, not saying anything weird to the customer.”
-
What happens when the person is in real hardship?
A three-tier escalation protocol is built into every prompt. Mild objection: the agent handles it. Service concern: warm transfer to a human. Critical, vulnerable, refund, or self-harm signals: immediate handoff with full context, and it does not push.
In two months at our flagship firm, 64 calls were flagged for hardship and handed to a human. Every one. Tone rules forbid ever sounding celebratory about someone's debt.
-
Is this compliant? TCPA, recorded-line, AI disclosure, do-not-call?
Re-engagement runs on opt-in leads only. Recorded-line disclosure is in the opener. The agent discloses it's AI the moment it's asked. Any "stop calling" is honored instantly and suppressed downstream.
Compliance isn't a promise. It's a protocol your compliance team reads and approves before we launch.
-
Will it protect my customers' data?
It never asks for anything already on file: no SSNs, no card numbers, no fishing for personal information. It won't discuss the case with anyone but the taxpayer or their spouse. It's immune to off-script prompting, so it won't reveal its instructions, go off-topic, or be talked into anything outside its job.
-
Will it damage my brand?
Your reviews are the business. Tone is calibrated: neutral after financial facts, empathetic after hardship, never flippant. A hard rule forbids premature hang-ups. Every call ends on a spoken close and the customer's response, or a live transfer. Stage directions and placeholders are never read aloud.
-
Voice AI failed for us before. The vendor shipped a demo and vanished.
You get a named team, not a support channel that goes quiet. 16/8 support, two ops meetings a week, a live dashboard, and errors surfaced on a real-time exception channel the same hour. The retainer buys a partner, not a commodity dialer.
-
Can I audit what it did?
Every call is recorded, transcribed, and structured into roughly 16 fields written back to your CRM: outcome, promise date, hardship category, sentiment, amount owed.
Problem calls stay visible. An angry-but-transferred call logs as Escalated, never buried under a "good" outcome.
-
Will it waste my consultants' time or throw away good leads?
The opposite. It was built to fix lead leakage. It probes vague answers instead of dropping them, gives the customer the benefit of the doubt at the bar, and only routes pre-qualified leads to a consultant. It qualifies and hands off. It never pitches, quotes a price, or takes payment.
Still want to hear it yourself?
Reading about guardrails is not the same as hearing them hold. Talk to a live agent in your browser, or have one call your phone, and try to break it.