Subodh KCSubodh KC/systems
let's talk →
AI Voice Operations

Kestrel Voice: A No-Code AI Receptionist Platform
Built for Real Business Calls

Deploy a hosted AI receptionist quickly — then add custom workflows, RAG, guardrails, integrations, testing, and managed support as your needs grow.

The Direct Answer

Most business owners are not looking for an AI voice development platform. They are looking for a practical outcome: answer my calls when I cannot, help the caller, book the appointment, transfer important calls, and show me what happened.

They do not want to assemble speech recognition, language models, text-to-speech, telephony, WebSockets, databases, vector search, call recordings, and analytics before the first customer can call. At the same time, they should not have to surrender control to a black-box receptionist that gives them little visibility into what instructions were active, where an answer came from, whether an appointment was actually created, why a call was transferred, what happened when an integration failed, how much the call cost, or what should be corrected afterward.

Kestrel is intended to be simple enough for a business to launch without coding, while retaining a path to deeper orchestration, custom integrations, managed deployment, and structured governance.

What Is Kestrel Voice?

Kestrel Voice is a hosted AI receptionist and voice-operations platform for businesses. A business can configure an agent to answer inbound calls, introduce the business correctly, respond to approved questions, capture leads, check appointment availability, create bookings, route or transfer callers, handle after-hours calls, detect configured urgent scenarios, use business knowledge through RAG, retain recordings and transcripts, summarize calls, and track outcomes.

The basic setup is handled through a dashboard rather than through source code. The onboarding flow covers business information, industry templates, voice and greeting selection, business hours, forwarding, services, FAQs, phone setup, and test calling. More complicated deployments — multi-location routing, custom CRMs, regulated workflows, complex scheduling, or consequential actions — can move into a supported customization or managed implementation process.

Where Kestrel Fits in the Voice-AI Market

Voice-AI products are not all trying to solve the same problem. A buyer will commonly encounter five approaches. This is a category comparison, not a claim that every platform fits one description.

Kestrel Voice Market PositionVoice-AI Market SpectrumDeveloperVoice APIMax flexibilityBuyer builds allVisual No-CodeBuilderFast flow designBuyer maintains flowsFixed AIReceptionistSimple packageLimited controlContact-CenterPlatformScale & QAEnterprise programManagedEnterpriseEnd-to-end deliveryLarge budgetKestrel VoiceNo-code launch + hosted runtime + custom orchestration+ RAG + guardrails + dashboards + supported customizationSelf-Service → Supported → Custom → ManagedBuyer ResponsibilityArchitecture, workflow, UI, opsDesign & maintain flowsLimited configurationImplementation programBudget & cycleBusiness configuration + ongoing reviewVapi, Retell SDKSynthflowAnswering servicesRetell, othersBland, FDEsKestrel's differentiation is not one isolated feature.It is the combination: fast setup + hosted infra + orchestration + controls + evidence + support
1

Developer Voice API

Strength: Maximum technical flexibility · Buyer: Architecture, workflow, UI, operations and support · Examples: Vapi, Retell SDK
2

Visual No-Code Builder

Strength: Fast workflow creation · Buyer: Designing and maintaining flows · Examples: Synthflow
3

Fixed AI Receptionist

Strength: Simple, packaged experience · Buyer: Limited configuration · Examples: Answering services
4

Contact-Center Voice Platform

Strength: Scale, QA and enterprise integrations · Buyer: Larger implementation and operating program · Examples: Retell, others
5

Managed Enterprise Service

Strength: End-to-end delivery · Buyer: Larger budget and deployment cycle · Examples: Bland, FDEs

Kestrel's differentiation is not one isolated feature. It is the combination: fast business setup, hosted voice infrastructure, custom orchestration, deterministic controls, managed knowledge, verified actions, fallback modes, operational evidence, and supported customization.

Kestrel's Core Position

A hosted AI receptionist that businesses can configure without coding, backed by a customizable orchestration and operations platform when the workflow requires greater control.

That position rests on six connected strengths:

Business Setup Without Voice Engineering

Configure through dashboard questions a business owner already understands — no speech providers, streaming audio, or telephony webhooks to manage.

Custom Python Orchestration Beneath the No-Code Layer

A custom runtime coordinates telephony, tenant config, conversation state, AI modes, deterministic rules, tools, transfers, recording, fallbacks, and post-call processing.

Deterministic Controls Around Generative AI

Three-layer adaptive architecture: hard interrupts, controlled fast paths, and AI reasoning — so AI is used where it adds value, not asked to improvise every business decision.

Business Knowledge With More Than a Single Prompt

Four knowledge layers: direct configuration, approved FAQs, RAG with refresh tracking, and live business tools — with a managed knowledge lifecycle.

Verified Business Actions

The AI reports success only after the business system confirms the result — not when it tries. Applies to bookings, transfers, CRM updates, and all consequential actions.

An Operations Dashboard, Not Merely a Call Counter

Call-level data including transcripts, recordings, outcomes, tool results, fallback status, cost, and a review workflow that turns real calls into corrections.

Business Setup Without Voice Engineering

A business should not have to understand the components of a realtime voice pipeline before testing an AI receptionist. The Kestrel setup experience focuses on questions a business owner already understands:

What is the business called?
What services does it provide?
When is it open?
Which areas does it serve?
What should the agent say when answering?
Which questions should it answer?
Where should calls be transferred?
Should it book appointments or take a message?
What happens after hours?
Which phone number should customers call?

Example: A local HVAC company

The owner can configure business name, normal business hours, emergency service hours, cities served, common problems handled, transfer number, appointment types, frequently asked questions, and after-hours response. The owner does not need to configure speech providers, manage streaming audio, or host a telephony webhook.

Custom Python Orchestration Beneath the No-Code Layer

Many voice agents are controlled primarily through prompts and visual automations. Kestrel also has a custom Python runtime that coordinates:

Telephony eventsTenant configurationConversation stateAI modes (Gather, Realtime, Adaptive, IVR)Deterministic rulesBusiness toolsTransfersRecording and transcript eventsFallback behaviorPost-call processing

The platform includes separate runtime paths for Gather, Realtime, Adaptive, IVR, outbound calls, recordings, transcripts, bookings, meetings, fraud controls, and post-call intelligence. Why does this matter? Because some requirements cannot be solved reliably by adding another sentence to the prompt.

Never confirm an appointment until the calendar succeeds
Stop the normal workflow when an urgent condition is detected
Change to a less expensive call mode after a usage limit is reached
Use a deterministic answer for business hours
Fall back when a realtime connection fails
Prevent one tenant's configuration from being used for another business

These are application and orchestration decisions — not merely conversational preferences.

Deterministic Controls Around Generative AI

The language model is useful for interpreting natural speech and conducting a fluid conversation. It should not independently control every decision. Kestrel's Adaptive architecture separates call handling into three layers:

Layer 1: Hard Interrupts

Emergency language, human-transfer requests, stop or opt-out requests, compliance phrases, critical account or security issues. Zero-latency, regex-based. The model is cancelled and a pre-written response is injected.

Layer 2: Controlled Fast Paths

Business hours, service area, common FAQs, basic routing, known business information. Keyword and lookup-based. The model is cancelled and a deterministic response is generated from tenant configuration.

Layer 3: AI Reasoning

Ambiguous caller language, multi-part questions, clarification, contextual follow-up, natural conversational responses. The model handles what benefits from interpretation.

The benefit is not that deterministic logic replaces AI. The benefit is that AI is used where it adds value, rather than being asked to improvise every part of the business process.

AI Voice Control Ladder

Click each level to see the capability, example, and minimum controls required. Higher levels carry greater risk and require stronger guardrails.
Lower riskHigher risk

Level 0: Inform

Example

Explain hours, services, or policies

Minimum Controls

  • Approved knowledge source
  • Tenant-bound configuration
  • No write tools

Read and write tools should be separated. The model may request a tool, but the application should enforce caller authorization, tenant scope, required fields, transaction limits, confirmation, idempotency, retry limits, and audit evidence. See the AI security tools for interactive agent-capability analysis or Kestrel Voice for a platform that implements these control levels.

Business Knowledge With More Than a Single Prompt

A receptionist needs to know more than a business name and greeting. Kestrel can combine several types of knowledge:

Kestrel Voice Knowledge StackKnowledge Stack — Four Layers1. Direct ConfigurationHours · Service areas · Transfer numbers · Business identity · Appointment settingsBest for: stable facts the business owner already knows2. Approved FAQsCommon questions with stable, pre-written answersBest for: predictable Q&A the business has already answered many times3. RAG & Website KnowledgeTenant-scoped sources · Chunks · Vector retrieval · Refresh tracking · Content hashesBest for: broader information from websites and documents — must be managed, not uploaded once4. Live Business ToolsAppointment availability · Customer status · Booking results · Current business recordsBest for: changing information that must be checked at call time, not cachedLifecycleApproved SourceIngestionRetrievalReviewRefreshReplace / DeleteRAG does not automatically make an answer accurate.If a website contains an outdated service area, the AI can accurately retrieve the wrong information.Knowledge must be managed — not uploaded once and forgotten.
1

Direct Configuration

Hours, service areas, transfer numbers, business identity, appointment settings. Best for stable facts the business owner already knows.
2

Approved FAQs

Common questions with stable, pre-written answers. Best for predictable Q&A the business has already answered many times.
3

RAG & Website Knowledge

Tenant-scoped sources, chunks, vector retrieval, refresh tracking, content hashes. Best for broader information from websites and documents — must be managed, not uploaded once.
4

Live Business Tools

Appointment availability, customer status, booking results, current business records. Best for changing information that must be checked at call time.

Knowledge Lifecycle

Approved SourceIngestionRetrievalReviewRefreshReplace / Delete

The distinction matters because RAG does not automatically make an answer accurate. If a website contains an outdated service area, the AI can accurately retrieve the wrong information. Kestrel's design therefore treats knowledge as something that must be managed, not simply uploaded once and forgotten.

Verified Business Actions

A natural-sounding call is not valuable if the agent does not complete the requested task. Consider an appointment request.

Verified Action Flow: Weak vs ControlledWeak WorkflowControlled WorkflowHear "Tuesday afternoon"Assume 2:00 p.m.Attempt calendar callSay "confirmed" — even if tool failsResult: Caller shows upat a time that was never bookedProblemsNo caller name capturedNo phone confirmationNo availability checkNo result verificationClaims success without proof1. Capture caller name2. Confirm phone number3. Identify service & location4. Check actual availability5. Offer valid options6. Repeat selected date & time7. Obtain caller confirmation8. Execute booking9. Inspect authoritative result10. Report success only if confirmedResult: Booking confirmedonly when the system says soSuccess = tool confirms, not AI tries

Weak Workflow

  • Hear "Tuesday afternoon"
  • Assume 2:00 p.m.
  • Attempt calendar call
  • Say "confirmed" — even if tool fails

Controlled Workflow

  • 1. Capture the caller's name
  • 2. Confirm the phone number
  • 3. Identify the service and location
  • 4. Check actual availability
  • 5. Offer valid options
  • 6. Repeat the selected date and time
  • 7. Obtain confirmation
  • 8. Execute the booking
  • 9. Inspect the authoritative result
  • 10. Report success only when the booking system confirms it

The AI should not claim that an action succeeded because it tried. It should claim success only after the business system confirms the result.

This principle applies equally to:

TransfersMessagesCRM updatesAppointment changesRefundsPaymentsAccount modifications

An Operations Dashboard, Not Merely a Call Counter

Businesses need to know more than how many calls the AI answered. A useful voice-operations dashboard should help answer: Who called? What did they need? How was the call handled? Was an appointment created? Did the transfer succeed? Was a fallback used? Which model or mode handled the call? What did the call cost? What should be corrected?

Operational Data Model

Calls and call status
Transcripts
Recordings
Outcomes
Interaction mode
Appointment results
Summaries
Tags
Fallback status
Model metadata
Latency information
Lead scoring

The review workflow can also turn a real call into a correction by adding a missing FAQ, an operational note, or a new review item.

Improvement Loop

CallReviewIdentify failureCorrect knowledge or ruleRetestImprove

The dashboard is therefore part of the operating system, not merely a reporting feature.

Failure and Degraded Modes

Realtime AI connections, calendars, transfers, and business APIs can fail. Kestrel's architecture includes a four-level degradation system and circuit-breaking with a Gather fallback when the realtime path cannot continue.

Level 0: Full AI

Frontier model with realtime voice synthesis. Full conversation quality. The default mode when all services are healthy.

Level 1: Fast AI

Cost-efficient model with standard voice synthesis. Reduced latency and cost. Used when the frontier model API is degraded or rate-limited.

Level 2: Rule-Based

State machine only. No language model. Handles greetings, basic intent routing, and voicemail capture. Used when AI APIs are unavailable.

Level 3: Human Transfer

Transfer to a human or voicemail. Last resort. Used when all automated modes have failed.

FailureResponse
Realtime WebSocket failsFall back to Gather (turn-based)
3 WebSocket failures in 5 minCircuit breaker trips → Gather mode
AI API degradedDegrade to Fast AI or Rule-Based
All AI unavailableHuman transfer or voicemail
Tool failsAI reports failure, offers alternative
Emergency detectedImmediate transfer, no AI deliberation

When one AI component fails, the entire customer journey should not disappear with it.

Learning From Calls

Kestrel Voice learns from call patterns. The learning pipeline promotes recurring fast-path patterns into active fast actions — but with guardrails.

1

Record

Every fast-path match is recorded with tenant ID, intent, user input, and timestamp. This is the raw signal for pattern detection.
2

Promote

Patterns that appear above a minimum occurrence threshold are promoted to active fast actions. Guardrails filter the promotion: only whitelisted intents, blocked phrases excluded, length limits enforced.
3

Review Window

Promoted patterns enter a time-delayed review window. Email reminders are sent to the tenant before auto-approval, giving them an opportunity to reject. This is a notification window with opt-out.
4

Activate

After the review window passes, patterns become active in future calls. The tenant can deactivate any pattern from the dashboard at any time.

Learning Guardrails

  • Intent whitelist — Only approved intent types can be promoted
  • Blocked phrases — Patterns containing blocked phrases are excluded
  • Length limits — Pattern text must be within configured bounds
  • Minimum occurrences — Patterns must appear multiple times before promotion
  • Tenant-scoped — Patterns are never shared across tenants

Beyond Voice: SMS, Video, and WebRTC

Voice is the primary channel, but Kestrel Voice also supports SMS, video sessions, and browser-based calling.

SMS

Outbound SMS via Twilio with internal (system notifications) and user (business messaging) modes. Inbound SMS handles TCPA-required STOP, HELP, and START keywords through a dedicated webhook.

Video Sessions

Video consultation sessions via Daily.co API. Session management includes creation, participant tracking, recording, and status lifecycle. Tenants can create video sessions for remote consultations.

WebRTC

Browser-based calling via Twilio Client. A prebind session binds the tenant ID before the device connects, ensuring deterministic tenant resolution for browser-initiated calls. Enables click-to-call from the dashboard.

Security Architecture

Security in a voice agent platform spans telephony, API, data, and compliance. Kestrel implements defense-in-depth across six domains.

Telephony Authentication

All Twilio webhook endpoints validate request signatures. Fail-secure in production — validation bypass is blocked outside development environments.

API Key Management

Three-tier system: admin keys (constant-time comparison), developer keys (SHA-256 hashed, tenant-scoped), and internal service keys for platform-internal communication.

Spam and Fraud Protection

Score-based spam enforcement with reputation memory. Tenant-configurable sensitivity levels. Four whitelist types for false-positive protection.

Toll Fraud Prevention

Country-specific call rate limiting with daily, hourly, and maximum duration limits. High-risk country detection for IRSF prevention.

Recording Consent

TCPA-compliant call recording consent. Area code to state mapping detects two-party consent states. Consent prompts generated as TwiML before recording begins.

Tenant Isolation

Five concentric layers: ASGI middleware, parameter-scoped queries, Postgres row-level security, server-side auth extraction, and internal API authentication.

Cost Visibility and Guardrails

Voice costs can grow through long calls, realtime model usage, repeated tools, multiple call legs, recording, transcription, failed retries, and abuse. Kestrel tracks model and telephony activity at the call level, including turns, tool calls, interruptions, and cancelled responses.

Cost Per Verified Outcome

This creates the foundation for measuring a more meaningful business metric: total voice and AI cost divided by verified completed outcomes. Examples include cost per appointment booked, cost per qualified lead, cost per successful transfer, and cost per resolved call. That is more valuable than reporting cost per minute alone.

Tenant-Specific Runtime Configuration

Each business can have its own configuration across every dimension:

IdentityPromptGreetingVoiceServicesHoursKnowledgeTransfer destinationEmergency behaviorAppointment configuration

Prompt Resolution Hierarchy

Tenant-specific override
Industry template default
Safe fallback

The prompt-resolution hierarchy also records information about the source and version of the prompt used. This helps diagnose questions such as: Did the correct business prompt load? Did the system use a generic fallback? Which instructions handled this call? Was the wrong business identity inserted?

Kestrel Compared With Common Voice-AI Approaches

This is a category comparison, not a claim that every platform fits one description.

ApproachMain StrengthBuyer ResponsibilityKestrel's Position
Developer voice APIMaximum technical flexibilityArchitecture, workflow, UI, operations and supportBusiness product first, with customization underneath
Visual no-code builderFast workflow creationDesigning and maintaining flowsAdds owned Python orchestration and supported deployment
Fixed AI receptionistSimple, packaged experienceLimited configurationDeeper business configuration and integration paths
Contact-center platformScale, QA and enterprise integrationsLarger implementation and operating programMore accessible starting point
Managed enterprise serviceEnd-to-end deliveryLarger budget and deployment cycleSupports both self-service entry and managed customization
Kestrel VoiceNo-code launch plus controlled expansionBusiness configuration and ongoing reviewHosted runtime, custom orchestration, RAG, guardrails, dashboards and support

A Practical Kestrel Call Scenario

Imagine a homeowner calling an HVAC company at 8:15 p.m.: "The furnace is making a strange noise, and I can smell something burning." A generic voice bot may continue through its normal intake script. A Kestrel deployment can be configured to follow a more controlled path.

1

Resolve the correct business

The called number is associated with the correct tenant and its name, greeting, instructions, emergency contacts, services, and business hours.

2

Evaluate priority conditions

The phrase involving a burning smell triggers a configured urgent scenario before normal appointment reasoning continues.

3

Deliver the approved response

The agent provides the business-approved safety wording rather than inventing technical advice.

4

Transfer or alert

The call moves toward the configured emergency contact or human fallback.

5

Preserve evidence

The business can later review the recording, transcript, trigger, transfer result, call outcome, and any follow-up action.

The important feature is not that the AI had a dramatic conversation. It is that the system followed a business-defined operational response.

Three Ways to Use Kestrel

Find Your Deployment Path

Select a use case below to see which deployment path fits, or click a path directly to explore what it includes.

Select your primary need

Self-Service AI Receptionist

Configure through the dashboard — no coding required

Best For

  • Basic answering and FAQs
  • Lead capture and message taking
  • After-hours coverage
  • Simple appointment workflows
  • Human transfer

What's Included

  • Business identity and greeting
  • Voice and tone selection
  • Business hours and after-hours rules
  • Services and service areas
  • Approved FAQs
  • Transfer destinations
  • Phone number setup
  • Test calling
Typical setup time:Minutes to hours

Every path starts with the hosted dashboard. You can begin with self-service and move to supported customization or managed deployment as your needs grow — without rebuilding from scratch.

Kestrel and HAIEC

Kestrel operates the voice workflow. HAIEC supports the assessment and assurance-readiness process surrounding more sensitive deployments. For a compliance-scoped implementation, the two work as a coordinated process.

HAIEC Helps Assess

  • Intended use
  • Affected users
  • Data
  • Jurisdictions
  • Applicable requirements
  • AI risk
  • Security exposure
  • Testing requirements
  • Evidence gaps
  • Known limitations

Kestrel Implements

  • The call workflow
  • Business identity
  • Instructions
  • Knowledge
  • Tools
  • Transfers
  • Emergency behavior
  • Recording configuration
  • Fallbacks
  • Monitoring
  • Call evidence

HAIEC helps determine what the deployment should control and document. Kestrel operates the configured voice workflow and produces operational evidence. The relationship is not an automatic technical certification or a guarantee of compliance.

Who Is Kestrel Best For?

Kestrel is a strong fit for businesses that:

Miss calls during work or after hours
Depend on phone-based leads
Need appointments or callbacks
Want a hosted solution rather than a development project
Need more control than a generic answering bot provides
Want to review recordings, transcripts, and outcomes
Expect their workflow to become more customized over time

Examples include:

Home-service businessesProfessional servicesConsultantsProperty managersMedical or dental practices (with reviewed deployment)Local and regional service providersMulti-location businessesSmall SaaS or support teams

When Kestrel May Not Be the Right Starting Point

  • A large global contact center with numerous geographic regions, existing CCaaS systems, formal procurement certifications, and contractual SLA guarantees
  • A company that wants to build its entire product and interface from voice APIs
  • A business that needs only simple voicemail replacement without custom orchestration

Kestrel's strongest position is between these extremes: a business-ready product that can become more customized and controlled without forcing every customer to begin as a software project.

Frequently Asked Questions

Is Kestrel Voice an AI receptionist?

Yes. Kestrel can answer business calls, respond to approved questions, capture caller information, support appointments, transfer calls, use business knowledge through RAG, and retain operational call records including transcripts, recordings, summaries, and outcomes.

Can Kestrel answer calls after hours?

Yes. Businesses can configure business hours, after-hours responses, callback behavior, voicemail, transfer paths, and other fallback options. After-hours handling is part of the standard configuration — no code required.

Can Kestrel book appointments?

Kestrel supports appointment workflows and calendar-related configuration. A properly configured workflow checks availability, confirms caller details, executes the booking action, and reports success only after the authoritative system confirms the result — not when the AI attempts the booking.

Can Kestrel use information from my website?

Kestrel supports business descriptions, FAQs, website sources, and RAG-based retrieval with tenant-scoped storage, content hashing, and refresh tracking. Website information should still be reviewed and refreshed because retrieval does not guarantee that the underlying content is current.

Can Kestrel transfer calls to a person?

Yes. Transfer destinations and transfer triggers can be configured through the dashboard. A custom deployment can also define what happens when the destination does not answer — including voicemail, callback, or a secondary contact.

Can I use my existing business number?

Depending on the phone arrangement, the business may use forwarding, conditional forwarding, a new Kestrel number, or a supported porting process. The setup wizard guides you through the appropriate option for your situation.

How is Kestrel different from a fixed AI answering service?

A fixed answering service may be faster to configure but offers limited control. Kestrel is designed to support deeper business instructions, RAG with refresh tracking, verified tool execution, operational evidence, customized routing, emergency rules, and a supported path to custom integrations — without requiring the business to start as a development project.

How long does it take to set up a basic AI receptionist?

A basic answering, lead-capture, or FAQ agent can be configured within minutes when phone provisioning and account setup complete normally. The onboarding wizard covers business information, industry templates, voice and greeting, business hours, forwarding, services, FAQs, phone setup, and test calling.

Can Kestrel handle multiple locations?

Yes. Multi-location routing is available through supported customization. Each location can have its own configuration, services, hours, and transfer destinations, with routing rules that direct callers to the correct location.

What happens if the AI cannot answer a question?

If the AI cannot answer from its configured knowledge, it can take a message, offer a callback, transfer to a human, or provide a deterministic fallback response — depending on the tenant configuration. The call review workflow also allows the business to identify gaps and add missing FAQs or knowledge sources afterward.

Does Kestrel work with my existing phone system?

Kestrel can work alongside an existing phone system through forwarding, conditional forwarding, or a dedicated Kestrel number. The platform also supports sequential ring — ringing the business phone first and falling back to AI only if no one answers within a configured timeout.

Is Kestrel fully production-ready for every business?

Readiness depends on the configured workflow. A basic FAQ agent is materially different from an agent handling healthcare information, outbound marketing, payments, or account changes. Each deployment should be tested according to its actual risk. HAIEC can support a structured assessment for compliance-sensitive use cases.

AI Receptionist Setup Checklist

Get a practical checklist covering business configuration, knowledge sources, deterministic rules, verified actions, fallback planning, security controls, testing scenarios, and operational review for deploying an AI receptionist.

No spam. Unsubscribe anytime.

For the complete production architecture behind Kestrel Voice, read AI Voice Agent Architecture: How Kestrel Voice Works. For ten failure modes and a six-phase deployment plan, read Why AI Voice Agents Fail in Production. To try the product, visit Kestrel Voice.

Ready to Deploy an AI Receptionist?

Start with a self-service setup or schedule a managed deployment consultation. Kestrel Voice handles the call. HAIEC handles the assurance.

Let's Talk →