Tilicho Labs
Banks, card issuers, lenders, credit unions, insurers

Voice agents for financial services

AI agents that call and answer your customers on the phone, by SMS and on web chat, in more than 25 languages, including Spanish. They finish the task during the conversation: confirm a flagged transaction, take a promise to pay, complete an account application, renew a policy. Tilicho builds and runs them inside your own Google Cloud project, starting from accelerators for 27 financial services use cases.

Auto loan payment, 6 days past dueEnglish
Sample call
AgentCustomerSystems
Loan record read: auto loan, payment $412.36, due Oct 12
AgentHi, am I speaking with Michael Turner?
CustomerYes, this is Michael.
Right person confirmed
AgentI'm the AI assistant from ABC Bank. This call is recorded. Your auto loan payment of $412.36 was due on October 12, and we haven't received it yet.
CustomerI know. My paycheck came in late this month. I can pay on Friday.
AgentThanks. Can I note Friday, October 23, for the full $412.36?
CustomerYes, Friday works.
Logged in the loan system: promise to pay $412.36 by Oct 23
AgentDone. I'm texting you a payment link now. I will never ask for your card number, PIN or password.
SMS sent: payment link
Follow-up call scheduled: Oct 24, 10:00 am local time, if unpaid
Risk alert read: $1,249.99, online electronics merchant, 3 minutes ago
AgentHello, am I speaking with María Hernández?
CustomerYes.
Right person confirmed
AgentI'm the AI assistant from ABC Bank, and this call is recorded. A charge of $1,249.99 at an online electronics store was just attempted on your credit card. Did you make it?
Customer¿Perdón? ¿Puede hablar en español?
Language switched to Spanish
AgentClaro. Hace tres minutos se intentó un cargo de 1,249.99 dólares en una tienda de electrónica en línea. ¿Usted hizo esa compra?
CustomerNo, yo no compré nada.
AgentEntendido. Ya bloqueé la tarjeta y el cargo no se aprobó. Le enviaremos una tarjeta nueva, y nuestro equipo de fraude la va a llamar hoy. Nunca le voy a pedir el número de su tarjeta ni su PIN.
Card blocked in the card system, replacement ordered
Fraud case opened, callback booked with the fraud team today
Card record read: delivered Oct 17, not yet active
AgentGood morning. Am I speaking with Jessica Park?
CustomerYes, speaking.
Right person confirmed
AgentI'm the AI assistant from ABC Bank. This call is recorded. Your new credit card was delivered yesterday. Would you like to activate it now? It takes about a minute in our app.
CustomerSure. Where do I do it?
AgentI'm texting you the link. Open it, then tap Cards and Activate. Please don't read your card number or PIN to me. I'll stay on the line.
SMS sent: app link to activate the card
CustomerOkay, it says the card is active.
Card status read: active
AgentIt's active. Your statement closes on the 5th of each month. Would you like to set up autopay while you're in the app?
Tagged in CRM: activated on call, autopay offered
Policy record read: auto policy, 6-month premium $1,386, renews Nov 1
AgentBuenas tardes. ¿Hablo con el señor Luis Ramírez?
CustomerSí, soy yo.
Right person confirmed
AgentSoy la asistente de inteligencia artificial de ABC Insurance. Esta llamada se graba. Su póliza de auto se renueva el 1 de noviembre. La prima para seis meses es de 1,386 dólares.
CustomerSí, pero quiero agregar a mi hijo. Ya tiene su licencia de manejar.
AgentAgregar un conductor cambia la prima, así que un agente con licencia lo atiende. ¿Lo comunico ahora?
CustomerSí, por favor.
Transferred to a licensed agent with the summary: renewal Nov 1, add one driver, newly licensed
Tagged in CRM: transferred, change of coverage requested
Names, amounts and dates are examples.
In production in
LendingReal estateHigher educationHealthcareE-commerceHome services
See the use cases
Comparison

Human calling team and voice agent

Human calling teamVoice agent
Calls in a dayAbout 100 per caller50,000+ across the campaign
First contactUp to 42 hours after the inquiryWithin minutes
LanguagesThe languages your team speaks25+, switching during the call
ScriptVaries by person and by dayThe approved script on every call
Follow-upDepends on the caller's notesScheduled until the outcome is reached
RecordA sample of calls is reviewedRecording, transcript and outcome tag on every call
CapacityHiring and training, with 30% leaving each yearAdded in a day, removed after the peak
Capabilities

What the agent can do

In conversation

  • Answers and places calls, 50,000+ in a day
  • Speaks 25+ languages and switches when the customer does
  • Stops when interrupted and picks up from there
  • Stays clear on a noisy line
  • Speaks with a natural tone and empathy, in a voice you choose
  • Remembers earlier calls and chats with the same customer

Actions during the call

  • Reads and updates the CRM, account, loan or policy record
  • Sends links, documents and confirmations by SMS or email
  • Books slots: loan officer meetings, branch visits, advisor callbacks
  • Raises and updates tickets
  • Finds the nearest branch or ATM for the customer
  • Answers from your product documents and FAQs

Outbound campaigns

  • Campaigns from an uploaded list or from your CRM
  • Targeting by segment, product, days past due or due date
  • Retries for no-answers and scheduled callbacks
  • Automatic follow-ups until the outcome is reached
  • Answering machines and voicemail detected

After every call

  • Recording and full transcript
  • Summary with the key points and next actions
  • Customer sentiment
  • Outcome tag written to your system
  • Summary message to the customer and to your team

With your people

  • Live transfer with the transcript summary attached
  • Signs of distress route the call to a person
  • Approval from a supervisor on Slack while the caller holds
  • Callback booked for a named team member

Connected to your stack

  • Salesforce, HubSpot, Zoho
  • Twilio, Telnyx, Plivo, Genesys, SIP trunks
  • Zendesk, Freshdesk, Google Calendar, Calendly, Cal.com
  • 5,000+ more apps through Zapier
  • Any other system through its API
One call

Before, during and after the call

1

Before the ring

The number is matched to the account, loan, card or policy record: the amount due, the date, the last conversation and any promise already made. Consent, the Do Not Call list, the customer's local time and attempt limits are checked before the dial.

2

During the call

The right person is confirmed before any account detail is spoken. The agent states the amount and the date from the record, takes the promise or the instruction, and texts the payment, application or renewal link while the customer is on the line.

3

When a person is needed

Disputes, hardship, changes of coverage and any request for a person go to your team live, with the transcript summary and the account history on screen. A supervisor can approve an exception on Slack while the caller holds.

4

After hang-up

Outcome tag, transcript and recording are written to the account, loan, card or policy system. The promise date or callback is scheduled, and a confirmation goes to the customer by SMS or email.

Use cases

27 accelerators for financial services

Each agent starts from an accelerator, so a use case is on live calls in about one month.

Cards

Account opening and onboarding

Lending

Live in one month

Ready in the acceleratorDay 1
  • The call flow and prompts for the use case
  • The fixed opening: right person, who is calling, AI and recording disclosure
  • Outcome tags and the number it is measured by
  • Integrations for telephony, SMS, CRM and ticketing
  • A test set of calls for the common situations
Done with youWeeks 1 to 4
  • Your numbers and telephony connected
  • APIs to your card, loan, policy or core system
  • Your approved script, product rules and languages
  • Compliance and information-security sign-off
  • Staging tests on your data, then the first live calls
Live in one month
2 weeks
when your systems are on standard integrations and the script is already approved
1 month
the usual case: one or two system APIs and a compliance review
2 months
when a core system needs a new API or the security review runs long
Compliance

How every call opens

The opening is fixed in the script. It is the same on every call, in every language.

1

Right person

The agent confirms it is speaking to the customer before it mentions any account detail.

2

Who is calling and why

Your organization's name and the purpose of the call, in the words your compliance team approved.

3

AI assistant, recorded call

The agent says it is an AI assistant and that the call is recorded.

4

The conversation

Only then does the script move to the amount, the date or the application.

Calling window
12 am8 am9 pm12 am

Calling hours

Calls are placed only inside the hours you set, in the customer's local time zone, for example 8 am to 9 pm. That is the window for telemarketing calls, and the window Regulation F uses for collection calls. Outside it the system does not dial.

Attempt limits

A limit on attempts per customer that you set, for example the seven calls in seven days that Regulation F uses for collection calls. Every attempt is recorded, answered or not.

Consent and Do Not Call

Under the TCPA, calls in an artificial voice to mobile numbers need the customer's prior express consent, and the FCC ruled in February 2024 that AI-generated voices are artificial. Consent is checked before the dial. Telemarketing calls are screened against the National Do Not Call Registry and your own list.

Personal data

Customer financial data is handled under your GLBA information security program. PII is masked in transcripts and logs, and access is by role. The agent never takes a full card number, PIN or password, so card data stays out of the call.

Your own cloud project

The agent runs in your Google Cloud project, in a US region. Recordings, transcripts and outcome tags stay in your storage, retrievable by account or policy number.

Transfer to a person

Disputes, hardship, changes of coverage and any request for a person are transferred live, with the transcript summary.

Engineering

How the agent is built

Speech to speechOne model hears and speaks. The fastest replies, and the caller's tone carries through. The default for live calls.
Caller audioGemini 3.8 LiveAgent audio
CascadedThree models in sequence. Each step can be inspected, and the model cost per minute is lower. Used where the script is strict.
Caller audioChirp 3 speech to textGemini 3.8 FlashGemini 3.8 Flash TTSAgent audio

One turn of a call

CallerAgent Speech detected (VAD)A pause, not the endEnd of turnAgent repliesCaller interruptsAgent stopsAgent answers the new point
Voice activity detection (VAD)Separates the caller's speech from line noise and silence. Only speech is sent to the model, so silence is neither processed nor paid for.
End-of-turn detectionTells a pause in the middle of a sentence from the end of it, so the agent does not reply while the customer is still reading out a number.
Barge-inWhen the caller interrupts, the agent stops speaking at once and the words it did not say are dropped from the conversation.
Answering machine detectionDecided from the audio in the first seconds, before a language model is involved, so voicemail is not counted as a contact.
Noise handlingCars, hands-free kits and speakerphones are common on these calls. The agent stays clear without asking the customer to repeat.
Context compressionA live model re-reads the call so far on every turn. Older turns are summarized, so reply time and cost stay flat on long calls.

Reliability

In-region deployment

The agent runs in the Google Cloud region closest to your customers. This is fixed before the build, since reply time cannot be tuned afterwards.

Language mirroring

The agent mirrors the customer turn by turn, so it follows the customer's language and not a setting.

Supervisor and task agents

A supervisor routes to small agents that share the state of the call. The script grows by adding agents.

Supervisor
VerificationPaymentDispute

Model fallback

A health-checked fallback model in the same region takes over mid-call, with the state of the call intact.

Warm transfer

Your team member sees the intent, the transcript summary and the captured details before speaking.

Closing the loop

An analyzer agent keeps each account open until it reaches an end state: paid, promise kept, or handed to your team.

Architecture

Where it sits

Your systemsOwned by you
Core, loan, policy and CRM systemsCore banking, loan origination and servicing, card platform, policy administration, Salesforce, HubSpot, Zoho. Two-way sync through APIs.
TelephonyTwilio, Telnyx, Plivo, Genesys or SIP trunks, on your existing numbers.
Payments and applicationsYour payment links, your online application and your identity verification journey.
Voice layerBuilt and run by Tilicho
Call handlingCall transport, latency control, live call state, prompts and tool calls, PII redaction.
Speech modelsKnows when the caller has finished, hears an interruption, detects a voicemail.
MeasurementResponse time, calls handled without a person, conversions, and what to fix next.
Google CloudModels and compute
ModelsGemini 3.8 Live speech to speech, or speech to text, Gemini 3.8 Flash and text to speech in sequence.
Agent runtimeGemini Enterprise Agent Platform: the runtime and its security boundaries.
Your projectPer-customer isolation in a US region. Recordings and data stay in your storage.
Integrations
TelephonyTwilio, Telnyx, Plivo, Genesys, SIP trunks
CRMSalesforce, HubSpot, Zoho CRM
SchedulingCal.com, Calendly, Google Calendar
TicketingZendesk, Freshdesk, Re:amaze
Messaging and operationsSMS, web chat, Slack, Gmail, Zapier
E-commerce and customShopify, Google Sheets, MCP Filesystem, custom REST APIs

Anything not on this list, including your core, loan or policy system, is connected through its API during pilot setup. Tilicho builds and maintains it at no extra cost.

The models

Gemini 3.8 LiveSpeech to speech. One model listens, reasons and speaks in real time. This is the model on the call.
Gemini 3.8 FlashReasoning in the cascaded pipeline, and the analysis after each call.
Chirp 3 and Gemini 3.8 Flash TTSSpeech to text and text to speech for the cascaded pipeline.
Gemini Enterprise Agent PlatformThe agent runtime and its security boundaries, inside your own Google Cloud project.
Questions

Common questions

Which languages does the agent speak?

More than 25 languages, including English, Spanish, Mandarin, Vietnamese, Tagalog, Korean, Arabic, French, Russian and Portuguese. The agent switches language during the call when the customer does, including between English and Spanish.

What happens when the customer wants a person?

The call is transferred live to your team with the transcript summary, so the customer does not repeat the case. You decide which cases always go to a person, such as disputes, hardship and changes of coverage.

How are consent and calling rules handled?

Consent, your Do Not Call list, the customer's local time and attempt limits are checked before every dial, using the rules your compliance team sets. Every attempt is recorded.

Where are recordings and customer data kept?

In your own Google Cloud project, in a US region, and in your storage. Tilicho runs the agent inside that project.

Do we have to change our telephony or CRM?

No. The agent connects to your existing numbers and systems. There are more than 50 integrations, and other systems are connected through their APIs during setup.

How is the pilot judged?

By the number you choose in week 0, for example promises to pay captured or applications completed, reviewed every week against your current baseline.

See it on one of your use cases

A 30-minute session: a live call with the agent, then a walk through one use case of yours, such as fraud alerts, early-stage collections or account opening.

Write to
sharada@tilicho.in