Case study · May 2026 · anonymised at the firm's request

How a leading US tax-resolution firm turned $367 of voice-AI spend into 390 hours of recovered capacity.

In two months: roughly 7,900 outbound calls, 880 live conversations, and 23% ending in customer forward motion. Every one of the 64 hardship calls went to a human.

The headline

Two months, one flagship account.

$367total usage costMar–Apr 2026, per-minute usage only
390 hrshuman dialing replacedindustry-standard 3 min per attempt
64 / 64hardship flags to a human7.3% of conversations, 11.1% on collections
$23,000disputed-paid invoices surfacedapproximately; reconciliation, not collection

Coverage

The calls humans weren't getting to · 11.2% per-dial contact rate

Dial attempts7,860
Live conversations880
Unique customers reached557

Conversation → forward motion

A payment commitment, a scheduled callback, or active engagement

Warm-lead recovery (P1)29.4%
Collections (P2)19.2%
Combined23.3%

The ramp

P1 rate Dec 2025 → Apr 2026, while volume scaled 5×

Dec Apr 11.1% 28.9%
The work humans were doing badly

This wasn't headcount that needed hiring. It was work that needed redirecting.

Three pieces of the firm's revenue motion were being handled by humans, and handled badly: the under-$500 delinquent bucket nobody prioritised, the contracts that signed but never funded, and live inbound screening that sometimes disqualified qualified consumers.

“We have a lot of humans that are taking on raw call volume. Pay for a bot to do it because the bot's gonna be more consistent. It's not going to get insulted, it's not going to have a bad day of the week.”

EVP at the firm
What the firm sees

Every call, one click away.

Kusp ops dashboard refreshed 4 min ago
880conversations7,860 dials
23.3%forward motion29.4% P1 · 19.2% P2
$1.79per positive outcome$0.42 per conversation
390 hrshuman dialing replacedin two months

Call outcomes

Share of 880 conversations · Mar–Apr 2026

23.3% forward motion
Customer forward motion23.3% Hardship → human specialist7.3% No commitment yet69.4%

Warm-lead recovery rate

P1 forward motion, Dec 2025 → Apr 2026, while volume scaled 5×

35 20 5 11.1% 28.9% Dec Jan Feb Mar Apr

Why they can't pay

Payment barrier tagged on every call · relative shape, illustrative

Financial hardshipn/a
Timing inconveniencen/a
Disputes the amountn/a
Needs assistancen/a
Value uncertaintyn/a

How they felt

Sentiment scored per conversation · relative shape, illustrative

Cooperativen/a
Neutraln/a
Anxiousn/a
Frustratedn/a
Resistantn/a

Call explorer

Every call one click away: recording, transcript, and all 16 extracted fields

CallOutcomeSentimentPromiseConfidence
C-AL-DC-31F4Promised to payCooperativeMar 11High
C-AL-DC-62FDCallback scheduledNeutralMar 14Medium
C-AL-DC-88A1Hardship → humanAnxiousn/aEscalated

What's real here: the KPI strip, the outcomes donut and the recovery-rate trend are the published Mar–Apr 2026 production figures. The barrier and sentiment panels show which dimensions every call is scored on; their distributions are not published, so no percentages are given. Call identifiers are synthetic.

How we deployed

We earn the next use case. We don't big-bang.

  1. Week 1–2

    Discovery + script alignment

    Mapped how the firm's best closers actually worked the phone, call by call.

  2. Week 3

    Agent configuration

    Prompt build and routing setup: proven modules configured, not built from scratch.

  3. Week 4

    Soft launch

    Limited volume, full instrumentation. Errors surfaced on a real-time exception channel the same hour.

  4. Week 5

    Full rollout

    Dashboard live, two ops meetings a week from day one.

  5. Month 2–3

    Use case 2

    Phase 1 payment recovery, added once past-due had cleared its own ROI bar.

  6. Month 4+

    Customer-pulled expansion

    Spanish 24/7 inbound went live. The next use cases were the EVP mapping his own roadmap.

The verdict

Benchmarked against the vendor they had before.

“It's reciting back the case IDs perfectly. Able to have a full-blown conversation with custom fields and just being able to quote to a customer, hey, you owe $212.14. Instead of us having to use fixed values, is going to be really cool for us.”

EVP at the firm

“It's very clear who the winner is in this one.”

EVP at the firm, after benchmarking Kusp against a previous voice-AI vendor