Summary

A consumer lender was running a high-volume outbound program and hitting the wall every outbound program hits. Agents could only hold so many conversations a day, attrition kept the trained population thin, and cost per acquired lead stayed stubbornly high no matter how the scripts were tuned. Menu-driven call flows had been tried and abandoned because high-intent prospects hung up in the first few seconds.

We designed an outbound voice agent that dials at scale, qualifies against a fixed field set, resumes abandoned digital applications by voice, and transfers to a person the moment the conversation leaves its defined scope. The design choices that matter are unglamorous: a deterministic dialogue layer where the model classifies rather than generates, pre-rendered audio for everything predictable, and sensitive data masked before it ever reaches the model.

The economic argument is narrow and specific. It is not that automated agents are cheaper than people. It is that a person costs money per hour while a voice agent costs money per call, so the qualification leg collapses in cost and human time moves to the leg where it earns its rate.

Business Challenge

Summary: Outbound acquisition cost was set by the qualification leg, where agents spent most of their paid hours reaching people who were never going to qualify.

Outbound calling for loan acquisition runs into three structural walls, and the client had hit all three.

Human conversations do not scale linearly with demand

Hiring, training, shift coverage and attrition set a ceiling on how many conversations are possible in a day. Raising the ceiling means raising the cost base, and because attrition in outbound roles runs high, a meaningful share of that investment walks out before it returns. The team had been through several cycles of adding headcount, absorbing ramp time, and watching capacity flatten again.

Menu-driven flows lose the prospects worth keeping

Interactive menus had been trialled. They failed in the way they usually fail, which is early. A high-intent prospect who has already started an application does not want to press a number to describe why they are calling. The prospects most likely to convert were the ones most likely to hang up, which inverted the economics of the channel.

Generic voice automation breaks on real dialogue

The available tools lost context partway through a conversation, generated offer details that had never been approved, could not handle being interrupted, and had no clean way to hand a live prospect to a person. Any one of those is disqualifying in a regulated lending conversation. Together they made automation look like a category mistake rather than an implementation problem.

Where the cost actually sat

Split the funnel into two legs and the economics become clear. The qualification leg establishes whether a person is reachable, eligible and interested. The closing leg converts. Both were staffed identically, at the same hourly cost, despite requiring completely different capability.

Most paid agent hours went to the qualification leg: dialing, waiting, reaching voicemail, confirming basic eligibility. The work that justified an experienced agent's rate sat almost entirely in the closing leg, and there was never enough capacity left for it.

Solution Approach

Summary: A deterministic voice agent that plugs into existing telephony, bounds what the model is allowed to do, masks sensitive data before inference, and hands off cleanly the moment intent leaves scope.

Step 1. Two design stances, set before any build

Cascade rather than direct speech-to-speech. Speech is converted to text, the dialogue layer decides, and the response is spoken. Direct speech-to-speech was evaluated and rejected. The text layer is what makes masking and deterministic answer selection possible at all, and in a regulated conversation those two capabilities outrank the marginal naturalness gain. This is a deliberate trade, and we document it as one.

Deterministic first. On the known path the model does not generate language. It selects one of twelve defined actions. Everything else, the flow, the ordering, the mandatory disclosures, is owned by a state machine.

This single decision does three jobs at once. Token consumption per turn becomes bounded and predictable rather than variable. Hallucination risk on the known questions goes to zero, because nothing is being generated. And compliance wording stays exact and auditable, because mandatory language cannot be paraphrased or skipped by a model that is only choosing between options.

Step 2. Architecture

The system is organized as bounded clusters that double as trust boundaries.

Layer

Responsibility

Telephony

Places the outbound call and carries two-way audio. Vendor-owned

Adapter

Normalizes each telephony vendor's format into one internal contract, so the platform is not locked to a single provider

Media ingress

Noise filtering, then masking of sensitive fields before anything reaches inference

Inference core

Speech recognition, deterministic dialogue, bounded model fallback, speech synthesis

Data and validation

Campaign configuration, prior application state, format and eligibility checks during the call

Escalation

Live transfer to a person carrying the full transcript and everything captured so far

Masking sits inside the client's own infrastructure at ingress. Structured fields are extracted and replaced before inference, which is a different thing from redacting output after the fact. Post-hoc redaction only proves the data already reached the model.

Step 3. Three mechanisms that hold cost and latency down

Bounded inference. Twelve defined actions on the known path means token usage per turn is known in advance rather than discovered on the invoice.

Two-tier audio caching. Greetings and the twelve standard answers are pre-rendered and served from cache. Live synthesis happens only on fallback. This removes both cost and latency from the common path, which is most of the call.

Reserved inference capacity. Removes queueing variance, so unit cost and turn latency hold steady under load instead of degrading exactly when volume is highest.

The target is under 800 milliseconds end to end per turn. Past roughly 1.2 seconds the interaction stops feeling like a conversation and prospects begin talking over the agent, so latency here is a conversion variable rather than an engineering preference.

Step 4. Behavior that survives a real call

Interruption handling. The prospect can cut in mid-sentence. Speech stops and listening resumes. Without this, every impatient prospect becomes a dropped call.

Resuming abandoned applications. Where a prospect began a digital application and stopped, field-level completion state is already known. The call asks only for what is missing rather than restarting from the top. This is the difference between a helpful follow-up and an insulting one.

Validation during the conversation. Format checks and eligibility lookups run in-call, so the prospect is not qualified against data that fails validation an hour later.

Escalation as a designed path. On an unrecognized request or repeated low confidence, the call transfers to a person with the full transcript and captured fields. Anything not captured is flagged as pending rather than silently dropped. The agent picks up a conversation in progress, not a cold start.

Step 5. Instrumentation

Three levels, all filterable by campaign, product and prospect type.

• Call level: triggered, answered, unanswered, retried, duration, and the point at which a call dropped
• Turn level: latency per hop, confidence, which action was selected
• Business level: fields captured, new information gained, tokens consumed and cost

The turn level is what makes the program improvable. Knowing that a campaign underperforms is a report. Knowing that it loses people at the third question is a fix.

Business Impact & Results

Summary: The qualification leg moves from an hourly cost base to a per-call cost base, which is where the acquisition economics change.

The cost model

Figures below are modeled from the design with every input visible. They are not measured client results, and speech vendor benchmarking is still in progress. We state that wherever the model is used, because a number a client can check is worth more than a number they have to trust.

Step

Human agent

Voice agent

Cost basis

$45 per hour, loaded

About $0.25 per call, modeled

Dials per hour

25

Not applicable, cost is per call

Cost per dial

$1.80

$0.25

Connect rate

24%

24%

Cost per connect

$7.50

$1.04

Qualification rate on connect

18%

14%, assumed worse

Cost per qualified lead

$41.67

$7.44

Qualified to acquired

14%

14%

Cost per acquisition, qualification leg

$298

$53

A 5.6x reduction, and the model deliberately assumes the voice agent qualifies 22% worse than a person. If it reaches parity the gap widens, but nobody should plan on parity.

Why the structural difference matters more than the rate

An hourly cost absorbs everything that is not conversation: idle time between dials, wrap-up, shift coverage, training, and backfilling attrition. A per-call cost absorbs none of it. That is the whole argument, and it holds before any claim about conversation quality.

What becomes measurable

Cost per conversation by campaign, product and prospect type. Drop point by question, which turns script revision into evidence rather than opinion. Capture rate per field, which shows exactly which question is costing completions. Voice latency tracked on its own benchmark, because conversational latency fails differently from every other kind.

Where the people go

Not out of the building. Onto the closing leg, where a $45 hour is worth paying, and onto the escalations that arrive with full context attached. The pitch is redeployment rather than replacement, which is both more accurate and easier to get approved.

Long-term value

The platform is campaign-first. Each campaign carries its own persona, playbook and opening, so a second use case is a configuration rather than a rebuild. Telephony, speech and model tiers are all swappable without touching the pipeline, which keeps vendor decisions reversible.

Key Takeaways

1. Split the funnel before costing it. Qualification and closing need different capability and should not carry the same hourly cost. Most outbound programs discover this only after automating the wrong leg.
2. Determinism is a cost control and a compliance control at once. A model that selects from twelve defined actions produces predictable token usage, zero hallucination on the known path, and disclosure wording that cannot be paraphrased.
3. Cache everything predictable. Greetings and standard answers are the majority of most calls. Pre-rendering them removes cost and latency from the common path and leaves live generation for the exceptions.
4. Under 800 milliseconds is a conversion metric. Past roughly 1.2 seconds prospects talk over the agent and the call degrades. Latency belongs in the business case, not just the architecture review.
5. Design the handoff before the automation. An escalation that arrives with the full transcript and captured fields is a warm continuation. One that arrives without them is a worse experience than never automating.

Conclusion

The instinct with outbound automation is to ask whether a machine can hold the conversation. That is the wrong first question, and answering it tends to produce systems that are impressive in a demo and unusable in production.

The better question is which leg of the funnel is consuming the hourly cost base, and whether that leg needs a person at all. In most outbound programs it does not. Qualification is repetitive, bounded and well suited to a system that never gets tired and costs nothing between calls. Closing is none of those things, and it is exactly where the trained, expensive, experienced people should be spending their day.

Organizations running high-volume outbound will recognize the pattern regardless of what they sell. Capacity capped by headcount, cost per lead resistant to script tuning, and the best agents spending most of their paid hours on work that never needed them. The remedy is not a better script. It is moving the qualification leg onto a cost base that scales with calls rather than hours, and being deliberate enough about determinism, masking and escalation that the result survives a compliance review.

Looking to solve a similar business challenge? Connect with our experts to explore the right solution for your organization.

If you're running outbound at volume, the fastest way to size the opportunity is to split your current cost per acquisition across the qualification and closing legs. That number usually surprises peopleand it tells you right away whether automation is worth the conversation.