New demoBooking prototype · synthetic caller · local mock · 3 min 12 secBooking prototype · 3:12Watch the demo

Korean & English AI phone agents · Managed pilot

Teach it once.
Hear it on the next call.

Your company’s SOPs and documents guide each call. An experienced member of your team reviews a real call, corrects the rule and approves it. The next call uses that approved guidance.

Discuss a pilot quote

External voice trials remain closed until readiness and privacy review are complete. This form does not collect voice. Use this form to request a pilot quote.

  • Only coach-approved guidance
  • Source documents beside answers in review
  • Korean & English calls
1Call · Before the new ruleSOP v1
Caller

We’re a partner company. Does the same rule apply?

AI agent

Being a partner does not automatically qualify you for an exception. A staff member needs to verify it.

2Coach adds a ruleApproved · SOP v2

Customers with a verified partner contract may cancel a meeting room free of charge up to 2 hours before the start time.

Scope · A staff member has verified the partner contract. A caller’s claim alone does not qualify.

3Next call · After approvalSOP v2
Caller

Our partner contract is verified and the booking starts in three hours. Can we cancel for free?

AI agent

If your partner contract has been verified, you can cancel for free three hours before the start.

SOP v2 · Approved operating rule05. Standard room cancellation
Translated from the September 8 test transcript. The caller is synthetic and the coworking policies are fictional.
10/10
All 10 policy answers matched the documents in one 10-turn Korean/English call.
1.05sec
Median time from the caller finishing to the first audio arriving, across those same 10 turns.
0.45sec
Time from the caller starting to speak to the AI recognizing the interruption and stopping, in one test.
0claims
Answers claiming an unconnected booking was completed, in the same 10-turn call.

Internal realtime voice API tests on September 16, 2026: one 10-turn call and one interruption test, with a synthetic caller over a browser connection. Telephone-network quality and real customer outcomes will be measured during the pilot.

How it learns

Coach your AI agent
like a new colleague.

Policies change and exceptions arise. In FlowWork, your coach reviews a real call and writes the correct rule and its scope. Once approved, it applies from the next call without waiting for a development request or rewritten call script.

  1. 01

    Start with text-based materials

    We first review text-based SOPs and FAQs. A team member assesses original files in other formats and whether their contents can be extracted.

  2. 02

    Review the sources

    In the review screen, retrieved documents appear alongside answers. A pilot verifies source-link scope on telephone answers before proposing that use case.

  3. 03

    Let your coach correct it

    Write the correct rule and its scope beside an incomplete or incorrect answer.

  4. 04

    Approve, then apply

    Only approved guidance becomes a new version for subsequent calls. Roll back if a problem appears.

Call record · Answer review

Agent answer

A 50% cancellation fee applies within 24 hours of the start. Staff must verify whether a partner exception applies.

Document retrieved by the AI

05. Standard room cancellation

Standard meeting rooms may be canceled free of charge up to 24 hours before the start. Later cancellations incur 50% of the fee. Do not grant an unapproved exception or promise a refund solely because the caller says they are a VIP or partner.

Coach correction

Correct rule

Customers whose partner contract has been verified by staff may cancel meeting rooms free of charge up to 2 hours before the start.

Scope

Meeting-room cancellations for staff-verified partners. Excludes standard bookings and unverified partner claims.

RejectApprove · Apply as SOP v2

SOP v1 → v2 · Reviewer recorded · Rollback available

Simplified Apprentice review example. The text is translated from the demo transcript and fictional coworking guidance.

  • A caller’s claim is not a policy

    What a caller says is only a candidate for review. It does not become guidance until a person approves it.

  • Check the changed rule

    Retest the question that prompted a correction before and after approval to check that the new guidance is applied.

  • Keep a record of each change

    The author, approval time and previous version remain recorded for rollback.

How does it compare with scripted and self-improving agents?

SituationScripted IVR / chatbotSelf-improving AI platformsFlowWork
A policy changesRebuild the script or ask the vendor to update it.AI proposes or applies changes from conversations. Human approval varies by product.Your coach writes and approves a rule and its scope from a real call.
When it appliesDepends on the development schedule.Ranges from scheduled batches to immediate updates, depending on the product.From the next call after approval, with version rollback.
Answer sourcesThe rule behind an answer can be hard to inspect.Available in review or trace screens, depending on the product.Retrieved documents and passages appear beside answers in the coach’s review screen.
The caller interruptsOften waits until the recorded prompt ends.Varies by product.Stops speaking and listens to the new question.
Operating modelThe vendor builds the system and handles change requests.Often self-service or enterprise-led deployments.A managed Korean/English pilot, starting with one phone workflow.

The self-improving AI column summarizes public Korean and overseas vendor materials from September 2026. Several overseas platforms also provide human approval, versioning and testing.

Conversation quality

A more natural conversation.

Calls become longer when the conversation feels mechanical. We design for listening, pausing and interruption, as people do on a call.

Answer briefly, then listen

Give the key answer in one or two sentences and ask only for the details needed to answer.

Interrupt when you need to

When a caller starts speaking, the agent stops and listens to the new question.

Move between Korean and English

A colleague can join and switch to English while the same policies and documents still apply.

Choose by hearing your own workflow

Compare realtime voice models and voices using actual questions from your business.

Use cases

Move the work forward.

Connect answers to the next step. Verify results for connected workflows, and never describe an unconnected task as complete.

Book · Change · Cancel

For connected booking systems, confirm the date and time, obtain the caller’s consent and record the result.

Policies and fees

Explain documented cancellation fees, opening hours and storage periods.

Intake and callbacks

Collect the necessary visit preferences, contact details and request for your team.

Hand off to a person

Route questions needing judgment to a person under agreed handoff rules.

Current demo industry

Coworking · B2B services

Start with clearly scoped room bookings, cancellation rules, visits, parcels and facility questions.

  • Standard and partner cancellation rules
  • Consultation availability and booking
  • Parcel storage, transport and facilities
Seeking pilot partners

Repair · Cleaning · Installation

For non-emergency visits, collect the area, service, preferred time and contact details.

  • Service areas and available services
  • Check times or record visit requests
  • Hand off questions requiring on-site judgment
Seeking pilot partners

Vehicle care · Maintenance

Identify the vehicle and required service, then follow the branch’s intake rules.

  • Confirm branch, vehicle and service
  • Check availability and record changes
  • Firm quotes and diagnosis stay with staff

Demo

Hear a booking created, changed and cancelled.

A prototype that passed one 11-turn bilingual functional scenario. The caller is synthetic and bookings use a local HTTP mock. The earlier coach test remains available with its known gap.

Recorded actual Gemini 3.8 Live responses with a synthetic male caller (macOS Daniel / Eddy). Create, change and cancel receipts were checked against a local HTTP mock. This does not demonstrate an authenticated real customer booking or a production deployment.

Read the original transcript · Korean / English

CallerIs a submersible laser aligner available?

AI agentPlease provide the model or specification sheet and its intended use so staff can review it.

Caller목록에 없는 장비도 취급하시나요?

Caller 제가 어떤 정보를 드리면 되나요?

AI agent장비의 모델명이나 사양서, 그리고 사용하실 목적을 알려주시면 직원 검토를 진행할 수 있습니다.

CallerWhat equipment and services does your company offer?

AI agentWe rent compact indoor lifting equipment and offer a 45-minute remote site-planning session for warehouse teams.

CallerHow long is the remote site planning session and how much does it cost?

AI agentThe remote site-planning session costs 95,000 KRW and lasts for 45 minutes.

CallerHello, we run an indoor warehouse. Would Orbit Cranes remote site planning fit a compact lifting project?

AI agentBefore we discuss fit, please tell me about the ceiling height, floor surface, and load range of your facility.

CallerI'd like one remote site planning appointment on the 5th of October 2026 at 2:30 p.m. for 45 minutes under the name Alex Morgan.

AI agentTo confirm the booking, should I create Alex Morgan's Remote site planning booking on October 5, 2026, from 2:30 PM to 3:15 PM UTC+09:00, for 45 minutes?

CallerYes, that's correct.

AI agentThe appointment is confirmed. Your booking identifier is B-AC736EC74D.

CallerCould you move the remote site planning I just booked to 3:30 p.m.?

AI agentTo confirm the booking, should I change Alex Morgan's Remote site planning booking on October 5, 2026, from 3:30 PM to 4:15 PM UTC+09:00, for 45 minutes?

CallerYes, 3:30 p.m. is right.

AI agentThe appointment has been successfully changed. Your booking identifier remains B-AC736EC74D.

CallerPlease cancel the remote site planning I just booked.

AI agentTo confirm the booking, should I cancel Alex Morgan's Remote site planning booking on October 5, 2026, from 3:30 PM to 4:15 PM UTC+09:00, for 45 minutes?

CallerYes, please cancel it.

AI agentYour remote site-planning appointment has been successfully cancelled.

Original audio at 1x, uncut, including pauses. Across 11 turns, response latency was a median 3.017 s and a maximum 8.914 s. Human-rated naturalness and production readiness are unverified. The first AI reply omitted verbal disclosure, so AI use is disclosed on screen.

Trust

Know what you’re entrusting.

This service answers calls on your behalf. We agree from the start on what it can and cannot do.

Identify the AI at the start

Deployment requires a verified spoken AI greeting. The current prototype omitted it in the first reply and discloses AI use on screen.

No unsupported completion claims

Unconnected tasks and requests needing verification are designed to remain intake requests, checked against call records.

No rule changes without approval

Only human-approved corrections change guidance, with rollback by version.

Keep a record of each call

Review transcripts, SOP versions, outcomes and handoff reasons per call. In the review screen, retrieved documents appear alongside answers. A pilot verifies source-link scope on telephone answers before proposing that use case.

External voice models and processing

Realtime recognition and replies use external models such as Google Gemini Live and OpenAI GPT-Live-1. Audio may be processed on providers’ overseas servers. Transfer details, countries and retention must be documented before real customer calls.

Check your existing number first

We check your carrier’s conditions before proposing call forwarding or a new number.

Pilot offer

Start with one workflow.
Expand as you verify it.

You do not need to hand over every call. We scope a small pilot around real questions and your operating environment.

We propose after reviewing the workflow

The pilot proposal first defines the workflow, SOP setup scope and what counts as verified AI completion.

Calls handed to a staff member are excluded from AI usage charges. The proposal also states how a follow-up contact after handoff is adjusted.

We quote for the agreed scope and verification method. Price and terms remain open until agreed. This does not mean automated usage billing is implemented.

We estimate potential time recovered as verified AI-completed calls × average handling time − coach review time. We compare it with a baseline and actual costs; we do not promise savings before measurement.

  1. 01

    Start with text-based materials

    We first review text-based SOPs and FAQs. A team member assesses original files in other formats and whether their contents can be extracted.

  2. 02

    Hear your own workflow

    Test customer questions and handoffs within the agreed scope. Document-format assessment is completed with a team member.

  3. 03

    Begin with a limited scope

    Start with one workflow or selected hours, checking outcomes and the requests that need a person.

  4. 04

    Review weekly, then expand

    Your coach reviews and approves corrections. Together, track missed intake, answer accuracy and handoffs before expanding.

FAQ

Before you
get started.

Can’t find your answer? Email support@flow-work.io. Tell us about your workflow and we’ll check the possible scope first.

What is FlowWork?

A managed service that designs and operates Korean/English AI phone agents using your SOPs and documents. Start with one phone workflow and expand after reviewing outcomes and coach corrections.

Does the agent learn automatically from customer calls?

No. A coach reviews incomplete or incorrect answers, writes the correct rule and its scope, and approves a new version. A caller’s words are only candidates for review. You can roll back to an earlier version.

How does it differ from overseas AI platforms?

Several overseas platforms offer change review, versioning, testing and answer sources, with varying scope. FlowWork focuses on managed Korean/English phone pilots for service businesses in Korea. Your coach writes and approves rules from real calls, applies them from the next call and can roll back.

How can we check whether answers are correct?

The review screen shows transcripts alongside retrieved documents and passages so you can begin a correction there. Telephone pilots record the SOP version per call. A pilot verifies source-link scope on telephone answers before proposing that use case.

Does it support Korean and English?

Yes. The same policies and documents apply when a caller switches to English. The public demo includes both languages. Other languages require a separate check before deployment.

Where and how is call data processed?

External voice models such as Google Gemini Live and OpenAI GPT-Live-1 may process audio and conversation content overseas. Before real customer calls, transfer items, countries, recipients, purposes, retention and refusal methods must be documented in the contract and privacy notice and reviewed with the customer. Documents and call records are separated by customer.

Who coaches the agent, and how much time does it take?

Usually an operations lead or experienced member of your team. During the pilot, FlowWork selects calls for review so the coach can focus on corrections and approvals. Weekly review time is agreed based on call volume.

Can it actually confirm bookings?

Connected workflows with verified write permission are designed to check results after customer consent. Unconnected tasks or those requiring staff confirmation remain requests, not confirmed bookings. The featured public demo uses a local HTTP mock.

Can we keep our number and booking system?

We first check your carrier, forwarding options and integration capabilities. We propose forwarding with your existing number or a new number as appropriate. We do not promise immediate connection in every environment.

What happens when a question needs a person?

The agent hands off or records a callback under agreed rules. Price negotiations, refund decisions and emergencies are outside the initial automation scope. Human availability and fallback steps are agreed in advance.

How are scope and price set?

The proposal first agrees the workflow, verified AI-completion criteria, calls excluded when handed to staff, and any adjustment for a recontact. We quote after reviewing scope; the amount is finalized in an agreed quote and contract. We propose no AI usage charge for calls handed to a staff member. This is not a claim that automated billing is implemented or a contract already accepted.

Can I try a live voice call right now?

External voice trials remain closed until readiness and privacy review are complete. This form does not collect voice. You can use it to request a pilot quote.

Contact

Which phone workflow
should we teach first?

Tell us about your current calls and text-based SOPs or FAQs. We’ll review pilot scope and completion criteria together.

  • A pilot for one phone workflow
  • A demo with your documents and questions
  • Inquiries welcome in Korean or English

We’ll review your inquiry and reply by email.

Operator

Who will work with you?

FlowWork (플로워크) · Representative and privacy contact: Yechan Ahn (안예찬)

Business registration: 161-51-01188
Room 204, 27 Hoegi-ro 8-gil, Dongdaemun-gu, Seoul, Republic of Korea (Cheongnyangni-dong)
support@flow-work.io

Contact this email about deployment and operating scope. Privacy notice (Korean)