FPT AI

A one-sentence summary

FPT AI is a comprehensive artificial intelligence platform designed for enterprises; its products include multilingual AI agents, chatbots for various channels, Voicebots for call centers, speech synthesis and recognition, document OCR, eKYC solutions, and employee training tools.

Tool Introduction

FPT AI is operated by Vietnam’s FPT Smart Cloud, and it serves areas such as customer experience, digital employees, operational automation, and sales processes. It is not a single model or a consumer-friendly chat interface; rather, it is a set of products that can be delivered through consoles, APIs, SDKs, and enterprise solutions.

The platform has been in development since 2017; its products are designed for use by businesses in Vietnam and Southeast Asia, and they also support languages such as English, Japanese, and Indonesian. Different products come with varying quotas, prices, and technical requirements, and it is necessary to check the specifics of each service when making a purchase.

Overview of the product portfolio

FPT AI Agents

Construction of multilingual AI agents

FPT AI Agents enables companies to create custom agents using natural language and configurable components; according to the product documentation, a basic agent can be developed in about 10 minutes. However, for actual production projects, it is still necessary to prepare knowledge, permissions, tools, and test sets.

Five core components

An agent is composed of Knowledge, Business Process, Tools, Guardrails, and Persona, which are responsible for managing knowledge, processes, external capabilities, security boundaries, and interaction roles respectively. The absence of any one of these components can result in correct responses but an inability to complete tasks, or in the ability to operate the system without proper authorization controls.

Business data and task execution

Agents can process structured and unstructured corporate data, and integrate with business systems to carry out specific tasks. The operations they perform should be based on the principle of least privilege, require approval, be idempotent, and include audit logs; reliance solely on generation models is not acceptable.

Content generation and knowledge Q&A

The platform can generate emails, reports, summaries of conversations and processes based on an organization’s knowledge, and it can also serve as an assistant for employees or customer service agents to answer questions. The actual reliability is determined by the updating of this knowledge, tracking of citations, and removal of outdated content.

FPT AI Chat

Natural Language and Generative Dialogue

FPT AI Chat integrates intent recognition, entity identification, keywords, dialogue management, a knowledge base, and generative AI to enable multi-turn conversations and more natural responses. Companies can use example sentences and business-specific knowledge to train the system continuously, but automatic learning still requires human oversight.

Omni-channel access

Chatbots can be deployed on websites, as well as on platforms such as Facebook Messenger, Zalo, Instagram, WhatsApp, and Viber. Each of these platforms has its own rules regarding templates, conversation windows, privacy, and messaging costs.

Emotion and Conversation Analysis

The system can analyze customer emotions, conversation content, and interaction data to assist teams in optimizing scripts and service processes. Emotional recognition may be affected by language differences, cultural factors, and sarcastic expressions; therefore it should not be used to determine how to deal with high-risk customers.

Collaboration between knowledge base and human customer service

The industry knowledge base can be updated to enable 24/7 responses, and robots can help reduce repeated inquiries. Complex complaints, identity disputes, transaction authorizations, and regulatory issues should be referred to human agents promptly, with the full context of the conversation retained.

FPT AI Engage

Incoming and outgoing Voicebot calls

FPT AI Engage is designed for call centers and can handle various inbound and outbound tasks such as customer service, welcome calls, surveys, reminders, verification, and telemarketing. The level of concurrency, the number of available numbers, recording capabilities, and compliance requirements for outbound calls must be determined based on the specific needs of each enterprise project.

Speech recognition and synthesis

Voicebot combines automatic speech recognition, dialogue management, and text-to-speech technology; its public product page offers over 100 different Vietnamese voice options. Noise, accents, cross-talk, proper nouns, and telephone line issues can affect the accuracy of speech recognition.

Call analysis

The platform can track call content, status, and various metrics, helping businesses identify customer needs and issues within their processes. Recordings, transcriptions, voiceprints, and emotional data are considered sensitive information; it is necessary to clearly specify their purpose, as well as the terms regarding access and storage.

Voice print authentication

The Engage scenario includes Voice Biometrics for Authentication, which can be used as part of the authentication process in call centers. Voice biometrics cannot serve as the only security element; it must be combined with factors such as liveness checks, device information, transaction risks, and alternative verification methods.

FPT AI Voice Maker

Natural Vietnamese speech

Voice Maker converts text into synthesized audio, offering male and female voices as well as accents from the northern, central, and southern regions of Vietnam. Punctuation, text normalization, and context influence pauses, stress, and tone.

Content and accessible applications

It can be used for audio content, application announcements, chatbots, call centers, as well as text-to-speech functions for users who are blind or elderly. For important notices, it is essential to read out names, numbers, amounts, and addresses correctly, in order to avoid misunderstandings caused by mispronunciations.

Real-time interface integration

The text-to-speech API can be integrated into applications and service workflows; requests in the free tier may be queued and are subject to daily limits, while paid plans offer a higher conversion speed. Long texts should be split into segments, cached, and have consistent voice parameters.

Speech to Text

The speech recognition service offers trial and commercial packages on a per-minute basis, and can be used for call transcription, media subtitles, voice commands, and subsequent analysis. The accuracy of this service needs to be evaluated based on factors such as the actual accent, sampling rate, noise levels, multiple speakers talking at once, and the vocabulary specific to a particular field.

The 60-minute trial period is available on an annual basis and comes without technical support; the business package includes standard technical support. For projects exceeding 10,000 minutes, it is necessary to contact sales to determine the price, throughput, real-time streaming capabilities, and SLA.

FPT AI Read

Document preprocessing and classification

FPT AI Read can identify the different files within a set of documents and apply operations such as cropping, rotation, correction, and image quality improvement. Preprocessing helps with OCR, but content that is severely blurred, obscured, or of low resolution may still not be recoverable.

OCR and field understanding

The product page states that it can handle PDFs, handwritten text, tables, signatures, and stamps, and uses OCR and natural language processing to extract relevant business fields. The claim of an accuracy rate of up to 98% is not a guarantee for all documents and fields.

Cross-verification and manual correction

The system can check for validity, duplicates, and contradictions within the same set of documents, and it allows for corrections through human collaboration. For amounts, identification numbers, contract terms, and key dates, confidence thresholds and manual verification are still necessary.

Template library and customization

The FPT AI Read documentation provides over 50 ready-made template formats for use in the banking, finance, insurance, and logistics sectors; it is also possible to customize templates or create new models based on specific business processes. The number of available templates does not mean that all countries and layout styles are supported, so it is necessary to test them using samples first.

FPT AI eKYC

Surface of the document, QR code, and NFC

eKYC is capable of reading information from the surface of identification documents, QR codes, and NFC chips, as well as cross-referencing the data in these fields. The availability of specific functions depends on the type of document, chip, device, and local regulations.

Face comparison and liveness detection

The system offers face recognition, comparison, search, and various liveness detection functions to counter attacks involving photos, videos, and deepfakes. Both false rejections and false acceptances can occur; therefore, manual and alternative verification methods are necessary.

Document anti-fraud

The platform claims to be able to detect various signs of fraud and quality issues on the surfaces of over 30 different types of documents, and it can also verify Vietnamese chip-based identity cards by integrating with administrative population databases. The capabilities of such databases are constrained by the customer’s qualifications, authorization levels, and regulatory requirements in Vietnam.

Web and mobile SDKs

eKYC provides SDKs for web and mobile platforms, along with a management portal for monitoring processes, customizing modules, and performing reconciliations. When integrating this solution, it is necessary to avoid exposing keys on the client side, and explicit consent must be obtained for the use of cameras, NFC, and biometric data.

FPT AI Mentor

FPT AI Mentor employs adaptive micro-learning techniques to tailor short lessons and repetitive exercises based on an employee’s strengths, weaknesses, and performance from the previous day. The content can be in the form of text, images, audio, and short videos, and it provides dashboards for both employees and managers.

Corporate knowledge can be transformed into questions, briefings, and customized training programs, but the content generated automatically should be reviewed by subject matter experts. There must also be transparent rules for determining employee rankings and skill profiles, to prevent the use of incomplete data in making personnel decisions.

Create a chatbot from scratch

1. Register for an FPT AI Console account, create a Conversation project, and obtain a robot token.
2. Define the robot’s uses, supported languages, channels, topics it is prohibited from addressing, and conditions for transferring to a human agent.
3. Organize frequent questions, corporate knowledge, intents, entities, keywords, and example sentences.
4. Design multi-round dialogues, exception branches, authentication, and business interface calls.
5. The testing environment includes typos, spoken language, mixed languages, ambiguous questions, and adversarial inputs.
6. Connect to websites or messaging channels, and set up Webhooks for coordination with human customer service.
7. After going live, monitoring tracks for issues, incorrect answers, the rate of transfers to human agents, and user feedback.

Tutorial on integrating a text-to-speech API

1. Enable the Speech service in the console and create a secure API key.
2. Select the voice, regional accent, and output format, and first test it with a short text.
3. Normalize text for dates, phone numbers, amounts, abbreviations, and foreign words.
4. Submit the text from the server; do not include the keys in web pages or mobile clients.
5. Read task results and download or play audio, handling queuing, errors, and timeouts.
6. Cache duplicate content, track character usage, and set budget limits.
7. Before official release, native speakers listen to the key content and test the accessibility of the experience.

Document recognition and eKYC implementation process

1. Specify the country, type, version, fields, daily processing volume, and error cost of the document.
2. Prepare authorized and anonymized samples of normal, blurred, reflective, obscured, and suspected fraudulent content.
3. Select the OCR, document recognition, face comparison, liveness detection, QR code, or NFC module.
4. Establish collection guidelines, quality checks, and user consent processes through APIs or SDKs.
5. Measure the field accuracy, false rejection rate, false acceptance rate, latency, and the amount of manual review separately.
6. Configure thresholds, manual queues, retries, evidence storage, and appeal channels.
7. After going live, monitor changes in document versions, attack methods, and model performance.

Process for launching AI agents

1. Choose a business task that has clear boundaries, allows for rollback, and is supervised by a human operator.
2. Create an approved knowledge base, specifying the effective date, responsible person, and confidentiality level.
3. Configure business processes, tool permissions, roles, and Guardrails.
4. Different authorization levels are set for reading, writing, sending, paying, or performing identity-related operations.
5. Verify the security boundaries using normal, abnormal, privilege escalation, and prompt injection tests.
6. Run it in shadow or advisory mode first, and then gradually enable automatic execution for low-risk tasks.
7. Ongoing audit results, tool calls, costs, failures, and manual intervention.

Which users are it suitable for

• Vietnamese and Southeast Asian companies: require AI capabilities in Vietnamese, English, Japanese, or Indonesian.
• Customer service and call centers: Develop text Chatbots, Voicebots, and omnichannel services.
• Banks, insurance, and financial institutions: handling documents, account opening, anti-fraud measures, and business-related paperwork.
• Retail and e-commerce team: Automatically answers inquiries, provides recommendations, sends reminders, and conducts customer surveys.
• Developers and integrators: Connect to enterprise systems via APIs, Webhooks, and SDKs.
• Content and Accessibility Team: Generating audio in various Vietnamese accents.
• Human Resources and Training Department: Develop personalized micro-learning platforms and capability dashboards.

Typical use cases

• Automatically answer questions regarding orders and services on websites, Zalo, WhatsApp, or Messenger.
• Use Voicebot to handle welcome calls, payment reminders, satisfaction surveys, and basic customer service.
• Convert Vietnamese articles, notices, navigation prompts, or e-learning materials into natural speech.
• Convert call recordings into text, and then analyze hotspots, emotions, and quality assessment metrics.
• Identify invoices, forms, signatures, seals, and logistics documents and enter them into the business system.
• Complete the reading of Vietnamese identification documents, face comparison, liveness verification, and anti-fraud checks.
• Create dedicated AI agents that can access corporate knowledge, invoke tools, and execute processes.
• Short training sessions and spaced reviews are delivered daily based on employees’ knowledge gaps.

Product advantages

• The product portfolio covers various business processes such as conversations, voice, documents, identity management, intelligents, and training.
• Vietnamese voices, regional accents, identification documents, and integration with local services all have a clear regional focus.
• Chat supports multiple international messaging channels, providing a unified interface for text-based customer service.
• Engage covers both incoming and outgoing calls, and provides call statistics and analysis.
• Read includes preprocessing, classification, OCR, cross-checking, and manual correction, rather than merely returning the original text.
• eKYC can incorporate documents, facial recognition, liveness checks, QR codes, and NFC modules.
• Conversation, Speech, and Vision offer some features at no cost or at a pay-as-you-go rate.
• It provides APIs, Webhooks, as well as SDKs related to Chat and eKYC, to facilitate system integration.

Usage restrictions and precautions

• Most enterprise products do not have fixed public prices; a demonstration, a needs assessment, and a customized quote are required.
• Different services are billed based on requests, characters, minutes, or a contract, and the quotas for each type cannot be used interchangeably.
• Free Speech requests may be queued, have time limits and daily usage restrictions, and no technical support is available.
• Public data on accuracy and efficiency are product metrics specific to certain scenarios, and they cannot replace testing with customer samples.
• Chat and agents may make mistakes or perform actions beyond their authorized scope; therefore, Guardrails and manual intervention must be implemented.
• Voice, voiceprints, facial features, identity cards, and NFC involve sensitive personal data and require strict compliance obligations.
• Multilingual support does not mean that all functions, industry-specific terms, and accents are handled with the same level of accuracy.
• The administrative population database is integrated for eligible Vietnamese businesses; it should not be considered a global public query tool.
• Public GitHub repositories do not mean that the entire FPT AI platform is open source; some of these repositories lack explicit licensing terms.
• The privacy policy does not specify a fixed period for data retention for each type of AI product; such details must be agreed upon in the corporate contract.

Price of chatbots

Although the enterprise plan specifies the requested volume and the price per additional unit, the base contract price still needs to be determined through quotation. It is necessary to clarify in the quote whether costs related to channel messaging, generative models, integration, customization, and SLAs are charged separately.

Text-to-speech pricing

Voice recognition pricing

Prices of document recognition and other products

The price verification date is August 22, 2026. The rates for Vietnamese dong, any additional costs in US dollars, taxes, fees associated with the distribution channels, as well as the corporate pricing, may change; the final amounts shall be those indicated in the control panel and in the contract.

How to understand the billing unit?

Supported languages and channels

Supported platforms and connection methods

API integration and key security

The FPT AI Console is used to create, manage, and monitor various service projects; public interface documentation is available for Conversation, Speech, and Vision. Developers should store robot tokens and API keys in a server-side key management system.

• Create keys and quotas separately for development, testing, and production environments.
• Do not embed the key in browsers, mobile packages, public repositories, or logs.
• Verify the signature or sender of the Webhook to prevent duplicate callbacks from triggering repeated business operations.
• All external inputs are checked for length, type, permissions, and malicious content.
• Set timeout, rate limiting, backoff retry, idempotent keys, and cost alerts.
• Manual or secondary confirmation is required for operations related to identity, payment, sending, and data modification.
• Regularly rotate keys and audit usage, failures, regions with abnormalities, and high-risk calls.

Privacy and data protection

FPT Smart Cloud’s privacy policy applies to websites, digital platforms, applications, and services; it may collect account contact information, corporate data, interaction content, device details, IP addresses, cookies, system logs, and usage history. In specific projects, session data, audio files, documents, and identity information uploaded by customers are also processed.

Personal data may be shared with entities within the FPT ecosystem, as well as with technology and support service providers, and with authorities authorized by law; it may also be transferred across borders in the course of providing services. The statement only outlines the principle of keeping the data for \"the period necessary to achieve the intended purpose or as required by law,\" without specifying a fixed retention period for each type of data.

AI training and user rights

The privacy policy guarantees that personal data will not be used to train AI models without proper consent or without undergoing the necessary anonymization process. Companies should still distinguish, in the project contracts, between service-related inference, quality improvement, dedicated training, and the training of models intended for use by other customers.

Data subjects have the rights to be informed, to access their data, to have it corrected or deleted, to limit or object to its processing, to withdraw their consent, and to file complaints. In cases involving eKYC, voice recognition, or employee profiling, companies must also provide mechanisms for manual review and appeal.

GitHub, SDKs, and open-source status

FPT Corporation’s official GitHub repository provides information on the Chatbot SDK and the AI Marketplace interfaces. The Chatbot repository contains Java and Python source codes as well as examples for sending messages to the robot engine and processing the responses from Webhooks.

The AI Marketplace repository lists interfaces for large language models, vision-language models, multimodal systems, Embeddings, Reranking, speech recognition, fine-tuning, and Responses. The fact that these details are publicly accessible does not mean that the model weights, the backend platforms, or the commercial services are open source.

Basic information

Recommendation score

Recommendation score: 4.3 / 5. FPT AI offers a comprehensive range of products for Vietnamese speech processing, document handling, call center services, and multilingual solutions for businesses in Southeast Asia; it provides various options, ranging from free API quotas to large-scale customized projects.

The main shortcomings are the complexity of the products and billing systems; most enterprise solutions require quotes, and the publicly available performance metrics cannot be directly applied to customer data. Sensitive information such as identity details, voice prints, and call-related data necessitate detailed data protocols, accuracy tests, and manual verification.

Frequently Asked Questions

What does FPT AI do?

It offers enterprises services such as AI agents, text and voice robots, speech synthesis and recognition, document OCR, eKYC, and employee training.

Is FPT AI free?

Some services come with a free quota: Conversation allows 10,000 requests per month, text-to-speech provides 100,000 characters per month, and speech recognition offers 60 minutes per year; most enterprise products require a quote.

How is there a charge when the chatbot exceeds its limits?

After exceeding 10,000 requests in the Free tier, the cost is $1.5 per 1,000 requests. The base price for the Enterprise tier requires a quote; the publicly available rate is $0.9 per 1,000 requests.

What is the free quota for text-to-speech synthesis?

The free plan provides 100,000 characters per month, but the conversion speed is low, requests may be queued, there are daily and usage time limits, and no technical support is available.

What is the free quota for speech recognition?

The trial version provides 60 minutes per year without technical support. The business package starts at 2,000 minutes; for amounts exceeding 10,000 minutes, it is necessary to contact sales.

Can Vietnamese audio be created?

Yes. Voice Maker explicitly supports male and female voices, as well as accents from the northern, central, and southern regions of Vietnam.

Is Chinese supported?

Chinese is not listed as a primary supported language. The platform focuses on languages such as Vietnamese, English, Japanese, and Indonesian; Chinese-related projects need to be tested first or an inquiry should be made.

Is it possible to create AI agents?

Yes. FPT AI Agents enable the creation of custom intelligent agents by leveraging knowledge, business processes, tools, Guardrails, and Personas, and it supports multiple languages as well as enterprise data.

Which chat channels are supported?

The public product page lists websites, Facebook Messenger, Zalo, Instagram, WhatsApp, and Viber, among others. The specific features and costs associated with each channel need to be checked separately.

Can FPT AI Read recognize handwritten text and tables?

The product description may include documents such as handwritten text, tables, signatures, and red stamps, but the accuracy depends on the quality and formatting of these elements. It is necessary to use actual samples to verify each field individually.

What functions does eKYC include?

This includes the surface of the document, QR codes, NFC, facial recognition and comparison, liveness detection, and anti-fraud features; it can be integrated via Web or mobile SDKs.

How much does document recognition cost?

When the number of transactions is less than 10,000 per month, the public price for image recognition of Vietnamese ID cards is 1,000 Vietnamese dong per transaction; for higher volumes, a quote is required.

Are APIs provided?

Available. Interface documentation is available for Conversation, Speech, Vision, and AI Marketplace; it is necessary to create an account and obtain keys in the console before using them.

Is there an official SDK?

Yes. Chatbot provides Java and Python source code along with examples, while eKYC offers Web and mobile SDKs; however, the licenses for each repository and delivery package need to be checked separately.

Is FPT AI open source?

The platform itself is not open source. What is made available on GitHub are mainly the SDKs, examples, and API documentation; some repositories do not have clear licensing terms, and the model weights as well as the commercial backend components are not made public.

Will customer data be used to train models?

The privacy policy states that personal data will not be used to train AI models without proper consent or without adequate anonymization. Corporate contracts should still clarify the boundaries between training for specific purposes and training for general use.

How long will the data be stored?

The privacy policy stipulates that data shall be retained only for as long as it is necessary to achieve the intended processing purposes or to meet legal requirements; no fixed number of days is specified for each type of product. Companies should define these periods in the contracts, on a case-by-case basis.

Is it suitable for Chinese companies?

When the business involves clients in Vietnam or Southeast Asia, as well as aspects related to language, documentation, and distribution channels, FPT AI proves to be a very suitable choice. When targeting only Chinese speakers, it is necessary to first verify the language capabilities and the integration with local systems.

Summary

FPT AI is suitable for organizations that wish to integrate chat, telephone, voice, document, identity, and agent functionalities within a single corporate AI ecosystem, especially those with operations in Vietnam and Southeast Asia. Developers on a small scale can test Conversation and Speech features using the available free quotas.

For formal projects, the requirements and invoices should be broken down by specific products; tests should be conducted using real-language inputs, documents, and accents as well as fraud samples. Additionally, costs related to channels, models, SLAs, data retention, cross-border operations, training, SDK licenses, and manual intervention must be specified in the technical and procurement documents.