Scaling Contact Center Operations with AI Agents in Fintech and Travel Industries

propel_holdings_/_xanterra_travel_collection · Finance · 2025 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: AI agents for contact-center scaling in fintech and travel

Propel Holdings and Xanterra needed to scale contact-center operations without proportional human-agent hiring, while preserving service quality across chat, voice, FAQs, and transactional workflows.

Constraints:

Propel had 40% year-over-year growth while live-agent ramp-up could take three to six months.
Xanterra had high volumes of routine FAQ traffic that consumed human-agent capacity.
Both organizations needed human agents to become comfortable with AI before autonomous deployment.
Initial API integrations were not ready, so transactional use cases had to wait.
Containment rates differed sharply between chat and voice channels.

Architecture

Both organizations used Cresta in a phased contact-center architecture: start with agent assist, then deploy FAQ-based virtual agents, ingest historical conversations for use-case selection, and add APIs for transactional capabilities once internal systems are ready. Agent assist provided summarization and consistency, while AI agents handled routine chat/voice conversations and freed humans for higher-value work.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Historical conversation corpus"] --> B["Cresta use-case analysis"]
  B --> C["Agent assist"]
  C --> I["Human contact-center agents"]
  B --> D["FAQ virtual agents"]
  D --> F["Customer chat channel"]
  D --> G["Customer voice channel"]
  D --> E["Transactional API integrations"]
  E --> H["Containment analytics"]
  F --> H
  G --> H
  H --> B

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Historical conversation corpus`	Past contact-center conversations used to identify automation candidates and FAQ coverage.	`Cresta use-case analysis`
`Cresta use-case analysis`	Identifies where AI agents and agent assist should be deployed first.	`FAQ virtual agents`, `Agent assist`
`Agent assist`	Human-agent support layer including summarization and consistency assistance.	`Human contact-center agents`
`FAQ virtual agents`	Initial autonomous chat/voice agents for routine questions without transactional API dependencies.	`Customer chat channel`, `Customer voice channel`, `Transactional API integrations`
`Transactional API integrations`	Internal APIs added in later rollout phases for more complex self-service workflows.	`FAQ virtual agents`, `Containment analytics`
`Customer chat channel`	Chat interactions where containment rates were highest.	`Containment analytics`
`Customer voice channel`	Voice interactions with lower but meaningful FAQ containment.	`Containment analytics`
`Containment analytics`	Operational metrics used to compare FAQ-only and API-enabled containment.	`Cresta use-case analysis`
`Human contact-center agents`	Agents redeployed toward complex, sales, and higher-value interactions.	`Agent assist`

Data flow

Historical conversations are ingested to identify FAQ and automation candidates. FAQ content powers initial virtual agents in chat and voice. Agent assist supports human agents with summarization and consistency. When APIs become available, transactional workflows are added, improving containment. Operational teams review containment, customer interactions, and use-case opportunities with Cresta.

3 supporting quotes

“once we crest was able to ingest our historical conversations they were able to identify whether we were off base or not or where to start.”transcript, stated

“we didn't have the uh sophistication of APIs internally yet. So we needed to start with the FAQs”transcript, stated

“phase 2 roll out uh include the API.”transcript, stated

Why this architecture

The phased architecture reduced deployment risk: agent assist built trust, FAQ agents delivered quick containment before API readiness, and API integration later expanded the automation envelope.

“agent assist is where we where we started. And then as far as AI agent goes, we just couldn't scale quick enough”transcript, stated

“we went with you know just FAQs and and we didn't have to worry about you know out of the gate to to get integrations with other systems”transcript, stated

Models used: Cresta AI agents · Cresta agent assist · conversational AI models

Scale

Propel cited 40% year-over-year business growth and three-to-six-month live-agent ramp time.
“40% year-over-year growth. um we just couldn't couldn't meet that as far as live agents goes in our business.”transcript, stated
Xanterra reported 60-90% chat containment and 20-30% voice containment from FAQ-based virtual agents.
“60 to 90% containment um 20 to 30% on the voice side to start with and that's just FAQs”transcript, stated
Propel's chat containment moved from about 38-40% with FAQs to about 58% after API integration.
“we had about 40 38 to 40% containment there in our chat channel um just with FAQs.”transcript, stated
The organizations started with agent assist before virtual agents.
“So agent assist is where we where we started.”transcript, stated

Alternatives considered (roads not taken)

Scale support through more live agents

Rejected because: Hiring and training agents took too long to match growth.

“It might take three to six months to to ramp up a live agent and uh we just couldn't catch up to that.”transcript, stated

Start with full transactional automation

Rejected because: The internal APIs were not ready for that first rollout.

“we didn't have the uh sophistication of APIs internally yet. So we needed to start with the FAQs”transcript, stated

Use separate vendors for agent assist and virtual agents

Rejected because: Xanterra preferred a one-stop shop covering both capabilities.

“one-stop shop, you know, that we could do both agent assist and the virtual”transcript, stated

Patterns

`tool-use-orchestration` (agent-architecture)

AI agents begin with FAQ knowledge and later integrate APIs for transactional contact-center workflows.

“once we ingested the the APIs, we were up to about 58% containment there.”transcript, stated

`human-in-the-loop-gate` (operations)

Agent assist supports human agents before and alongside autonomous agents.

“initially we wanted to provide more tools for our agents. Uh all our agents are working from home”transcript, stated

`eval-driven-development` (evaluation)

Containment rates and historical conversation analysis guide rollout sequence and API expansion.

“they were able to identify whether we were off base or not or where to start.”transcript, stated

`fallback-degradation` (operations)

The initial FAQ-only scope delivered value while transactional integrations were still unavailable.

“we needed to start with the FAQs um while our development team would ramp up”transcript, stated

Tradeoffs

Rollout sequence

Chose agent assist and FAQ automation first over full autonomous transactional agents immediately.

Rationale: It delivered quick wins while agents, APIs, and internal readiness matured.

“we went with you know just FAQs and and we didn't have to worry about you know out of the gate to to get integrations”transcript, stated

Vendor architecture

Chose single Cresta platform for assist and virtual agents over multiple specialized vendors.

Rationale: A one-stop-shop reduced integration complexity across agent assist and automation.

“one-stop shop, you know, that we could do both agent assist and the virtual”transcript, stated

Gotchas (what bit them)

Use-case intuition was not enough.

Root cause: The organizations thought they knew which cases to automate first, but historical conversations changed prioritization.

Resolution: Ingest and analyze historical conversations before finalizing rollout targets.

“we thought we know knew what use cases to start with but once we crest was able to ingest our historical conversations”transcript, stated

Transactional automation depended on internal API maturity.

Root cause: The required APIs were not available at initial launch.

Resolution: Start with FAQ containment, then add API-backed capabilities in phase two.

“as a part of kind of a phase 2 roll out uh include the API.”transcript, stated

Human-agent inconsistency was already a production problem.

Root cause: Live agents could give inconsistent information, making AI consistency a practical rather than purely theoretical concern.

Resolution: Use agent assist and AI agents to standardize parts of the workflow.

“we have live agents right now giving inconsistent information.”transcript, stated

Evaluation

The organizations evaluate through containment rates, channel split, FAQ versus API rollout performance, conversation review, and operational feedback with Cresta.

Metrics: chat containment · voice containment · FAQ containment · API-enabled containment · after-call work reduction · agent ramp time

“we're seeing on the chat um side dependent upon you know the the product 60 to 90% containment”transcript, stated

“Um and uh really kind of cutting down that after call work for agents.”transcript, stated

Outcomes

Xanterra saw 60-90% chat containment and 20-30% voice containment on FAQ use cases.
Propel moved from about 38-40% FAQ chat containment to about 58% after APIs.
Agent assist delivered quick wins such as auto summarization.
AI agents helped address 40% year-over-year growth without one new human for every conversation.

The case covers two industries, chat and voice channels, and a phased path from FAQ containment to API-backed transactions.