AI agents / Knowledge & action
AI agents and RAG chatbots grounded in business data
Friday Works builds AI agents and RAG chatbots that retrieve authoritative business knowledge, cite sources and perform carefully bounded actions. The system is designed around access control, answer quality, cost and a route to human help when confidence is insufficient.
- Indicative budget
- 15–30 million VNDFirst-phase document chatbot; integrated agents are quoted by scope
- Timeline
- 3–5 weeksFocused chatbot; agents integrated with several systems often take 6–12 weeks
- Answers
- With sourcesSays clearly when documents do not support an answer
- Human handover
- When unsureWith the full conversation context
Common problems
What businesses usually ask us to solve
Situation 01
New staff ask the same questions about policies and procedures again and again.
How Friday Works responds
An internal chatbot answers from approved documents, respecting each group's permissions, with a link to the source.
Situation 02
Customers ask after hours on the website or Zalo and nobody replies.
How Friday Works responds
The chatbot explains approved information, records the request and hands over to staff with context when needed.
Situation 03
You tried an AI chatbot and it invented answers.
How Friday Works responds
RAG grounds answers in your documents; when evidence is missing, the system says the information is not available.
Situation 04
You want AI to perform some tasks but fear it will get them wrong.
How Friday Works responds
The agent uses only an allow-list of tools, important actions require approval and every step is logged.
Project types
Projects we typically take on
Internal knowledge chatbot
Answer from policy, technical and process documents according to each user's access.
Customer assistance
Explain approved information, collect context and transfer cases that need a person.
Task-support agent
Prepare drafts, create tickets or call APIs within explicit permissions and approval points.
Chatbot or AI agent: where to start?
A good fit when
- You have reliable documents such as policies, guides and product information, and someone who updates them.
- Questions repeat often and answer quality can be measured on a sample set.
- You accept starting with an answering chatbot before letting an agent take limited actions.
Not a fit when
- Documents are contradictory or outdated and nobody owns their upkeep.
- You want an agent to execute transactions, payments or important decisions without approval.
Work in production
Real products serving real users
How we work
Step by step: who does what and what you receive
Example schedule for a website chatbot in one language with up to 100 pages of machine-readable documents (package A02, 3–5 weeks). Agents integrated with a CRM, helpdesk or Zalo OA often take 6–12 weeks and are quoted by scope.
- 01Before signing
Conversation & job definition
Decide whom the assistant serves, what it answers and what it may do.
Friday Works does
- Understand the users, question types and channel: website, Zalo OA or internal.
- Review existing documents and how current they are.
- Propose the scope: knowledge sources, channel, permitted actions and evaluation.
Your part
Share frequent questions and the documents used to answer them today.
You receive
- A scope proposal with evaluation criteria
- Schedule and quotation
CheckpointSign the proposal and pay the first instalment (40%).↗︎
- 02Week 1
Knowledge & evaluation set
Chatbot quality depends most on the source documents.
Friday Works does
- Select authoritative documents and remove duplicates or outdated versions.
- Set permissions: which group may query which documents.
- Write an evaluation set with correct answers and a list of questions to refuse.
Your part
Provide documents and confirm the reference answers.
You receive
- A knowledge-source inventory
- An evaluation set and answering rules
CheckpointAgree the evaluation set.↗︎
- 03Weeks 1–3
Prototype & evaluation
Measure quality in numbers before users see it.
Friday Works does
- Build the RAG flow: split documents, retrieve relevant passages and answer with citations.
- Choose a model by Vietnamese-language quality, latency and cost.
- Measure correct, grounded and refused answers and the cost per question.
Your part
Ask test questions and flag answers that misread the business.
You receive
- A chatbot prototype
- An evaluation report on the question set
CheckpointDemo and the second instalment (40%).↗︎
- 04Weeks 3–4
Interface & channel integration
Put the assistant where users already ask questions.
Friday Works does
- Embed the chat on the website or another in-scope channel.
- Set up handover to staff with context.
- Build the document-update screen and conversation logs.
Your part
Name the people who receive handed-over conversations.
You receive
- The chatbot on a test environment
- A document administration screen
- 05Weeks 4–5
Trial & launch
Internal users test it before customers do.
Friday Works does
- Let an internal group try it and adjust from feedback.
- Test questions that deliberately try to make the bot break its rules (prompt injection).
- Launch with a tracker for unanswered questions.
Your part
Organise the trial group and confirm the launch date.
You receive
- The chatbot live
- A quality and cost tracker
CheckpointLaunch.↗︎
- 06After launch
Handover & improvement
Your team updates the knowledge and knows when to expand.
Friday Works does
- Show how to update documents and read conversation logs.
- Hand over source code, configuration and accounts.
- Widen agent permissions only when evaluation data supports it.
Your part
Accept the work against the criteria and name a document owner.
You receive
- Operating documentation
- Three months of defect fixes for Friday Works' implementation
CheckpointAcceptance and the final instalment (20%).↗︎
Deliverables
What you receive
Knowledge & evaluation
- Knowledge-source inventoryAuthoritative documents, versions and access rights.
- Evaluation setAnswers confirmed by the business, used to measure before and after every change.
- Answering rulesWhat may be said, what must be refused and when to hand over.
Chatbot & agent
- Answers with sourcesA RAG flow retrieves the right passages before answering.
- Chat widget and staff handoverOn the website or another in-scope channel.
- Document-update screenAdd, edit or remove documents without a developer.
- Bounded actionsFor agents: allowed tools only, approval steps and logs.
Quality monitoring
- Evaluation reportAccuracy, unsupported-answer rate, latency and cost.
- Conversation logsRetained under the agreed policy for improvement.
- Unanswered questionsPoint to the documents that should be added.
Handover
- Source code and configurationWith AI service accounts registered in the company's name.
- Operating guideUpdating knowledge, reading logs and monitoring cost.
- Three months of fixesDefects in Friday Works' implementation are fixed after acceptance.
Technology and tools we commonly use
The exact stack is chosen for each project and recorded in the proposal rather than forcing one platform onto every problem.
- AI modelsOpenAI, Anthropic Claude, Google Gemini or another model suited to Vietnamese, latency and budget.
- RAG and vector databasesRetrieve the right passages and answer with citations.
- Channel and system integrationWebsite, Zalo OA, CRM and helpdesk through APIs where the platform allows.
- Guardrails and monitoringPermissions, topic limits, logs and cost tracking.
Collaboration
What both sides prepare
What you prepare
- Machine-readable source documents (text-layer PDF, Word, web pages) and someone who updates them.
- Frequent questions with correct answers.
- Rules: what may be said, what must be refused and when to hand over.
- People who receive handed-over conversations.
- Access to the channels (website, Zalo OA) and related systems.
What we check before handover
- The agreed quality threshold is met on the evaluation set.
- Answers cite sources; out-of-scope questions are declined politely.
- Users can retrieve only the documents their role allows.
- Handover to staff works and keeps the full context.
- Common rule-breaking prompts (prompt injection) have been tested.
- Cost and latency stay within the expected limits.
Pricing
Indicative packages and how the budget is set
Related starting packages
Version 20 October 2026 · Unit: million VNDDocument-grounded AI chatbot
15–30million VND / first phase
Estimated: 3–5 weeks
Baseline scope. One website chatbot in one language; up to 100 machine-readable document pages; cited answers, missing-evidence notices and a document update screen.
Not included: Scanned-document OCR, Zalo OA, acting on behalf of users, model fine-tuning or AI usage fees.
These are indicative budgets for new engagements, not fixed quotations or prices from completed contracts. Scope, schedule and cost are confirmed for each project before work begins.
Typical timelines
- Document chatbot (A02)
- Often 3–5 weeks when documents and permissions are clear.
- Integrated agent
- Often 6–12 weeks depending on tools, actions and safety requirements.
- Expanded autonomy
- Only considered when evaluation data shows suitable quality and controls.
The budget depends on
- The amount, format, update frequency and permissions of knowledge sources.
- Model choice, conversation volume, latency and retention requirements.
- The number of tools, APIs and actions an agent may use.
- Evaluation, guardrails, monitoring and operational support needs.
Proposed payment schedule
- 40%Project start
- 40%Demo & core features
- 20%Acceptance & handover
Before a quote is final
- Agree the page, feature, data and acceptance-criteria list before implementation.
- The schedule starts after scope approval, receipt of required inputs and the first payment.
- The proposal includes three months of defect correction for Friday Works' implementation after acceptance.
- Domain, hosting, email, AI/API, data, licences, third-party services and applicable tax are shown separately.
After handover
Warranty, maintenance and next steps
Three months of defect fixes
Defects in the work Friday Works delivered are fixed for three months after acceptance and are not charged as maintenance.
Monthly maintenance
2–5 million VND / month
Up to 6 hours/month · One application; log review, compatibility updates and technical support.
Support is provided during business hours, not as a 24/7 service. Infrastructure and work above the included allowance are quoted separately. Warranty defects are not charged as maintenance.Frequently asked questions
Before we begin
01Can a RAG chatbot still give a wrong answer?
A RAG chatbot can still give a wrong answer. RAG helps ground an answer in sources but cannot eliminate every error. The system needs citations, an evaluation set, fallback thresholds and a route to a person.
02What determines AI agent or RAG chatbot cost?
AI agent or RAG chatbot cost depends on the use case, data sources, document quality, permissions, request volume, AI models, system integrations, evaluation and human approval requirements. A focused proof of concept or pilot measures quality, latency and cost per task before production rollout.
03How is an AI agent different from a chatbot?
A chatbot mainly exchanges information; an agent can plan steps and use tools to act. That is why agents require stronger permission limits, approvals and logging.
04Can internal data remain permissioned?
The architecture should preserve least-privilege access. Users should retrieve only the documents and actions their role permits; the exact controls are defined during discovery.
05Does the chatbot answer well in Vietnamese?
Current large models understand and write Vietnamese well, but real quality depends on the source documents. That is why quality is measured on the business's own Vietnamese question set before launch; package A02 covers one language.
06Can scanned image documents be used?
A document chatbot needs machine-readable text. Scanned documents must first go through OCR or Document AI to become text; that work is outside package A02 and quoted separately.
Talk to Friday Works
Bring business knowledge into the right workflow
Share the users, document sources and actions the assistant should support. We will define a RAG or agent scope that can be evaluated.
Describe your projectAfter you send a request
- 01Send a short brief
Your goal, current situation and preferred timing; attachments are welcome.
- 02A reply within one business day
Friday Works sends clarifying questions or proposes an online call.
- 03Discuss the scope
Review the goal, risks and an approach that fits the budget.
- 04Receive a written proposal
Scope, milestone schedule, cost and acceptance criteria.