AI Solution
AI Solution
AI Set Up on Your Own Data and Rules
AI that answers your customers in WhatsApp and reads your invoices, using your own price list and return policy, with a person checking anything about money or health.
AI tools, agents and automation for a client's work.
If the AI built into WhatsApp Business or your marketplace app already does the job, start there. Our Consulting Services can help you set it up. Set Up a Product Catalogue and Quick Replies in WhatsApp Business
At a Glance
- You Get
- one AI task on your own data
- Starts With
- one task you choose, as a pilot if you want one
- Your Data
- processed in Indonesia if you choose
- Start
- name the work that takes longest
Start with One Task, Then Decide
The way to start is agreed in the first conversation. One way is a pilot: one task for an agreed period, so you see the AI's work on your own data before you decide on more.
When you choose a pilot, it typically takes 2 to 4 weeks.
- 01
Choose the Task
We go through a week of your chats and documents with you and pick the one task where AI saves the most time with the least risk.
You ReceiveA short note on which task to start with, and why.
- 02
Check the Data
We agree what the AI may read, what it must never see, and where it is processed. You choose before we build.
The AI May ReadThe AI Never Sees
You ReceiveA data check and the written pilot plan (see Sample Pilot Plan).
- 03
Build the Pilot
We set the AI up on that task with your rules and test it on real past questions or documents.
You ReceiveThe assistant, reader or automation set up on your data; its rules (guardrails); a hand-over point to a person, with the conversation so far.
- 04
Run the Pilot
Your team uses it for the agreed period while a person checks its work. You see what it got right and wrong.
You ReceiveA screen to see and correct what the AI uses; a running-cost estimate with a spending limit.
- 05
Decide Together
You stop, adjust or roll it out further, with the running cost known.
You ReceiveA results report and a decision meeting.
- 06
Keep It Right
When it runs daily, we check its answers and costs every month.
You ReceiveA monthly check of answers and costs while you keep a support plan.
- Pilotone task for an agreed period.
- Rolloutmore tasks or more staff after a successful pilot.
- Ongoing Supportwe run the AI and check its work each month.
Any of these can also be part of the Turnkey Solution.
Work AI Can Take On
Where Most Businesses Are Today
For most UMKM, AI arrives free inside chat and platform apps they already use, and is used by the owner to make content.
- Most businesses that use AI use it at a basic level.
- AWS and Strand Partners name a lack of digital skills as the main barrier.
- IBM names infrastructure as the most cited barrier, at 84%.
Sources
- AWS and Strand Partners, AI adoption in Indonesia, 7 August 2025.
- IBM, Indonesia AI study, 4 June 2025.
In a 2025 UKM Indonesia survey of 204 small businesses, mostly very small, about two in three already used AI; writing a good prompt was the biggest barrier. The sample came from UKM Indonesia's own channels, so it is not a national figure.
The Step to Deeper Use
AI solutions answer the step from basic to deeper use.
- The same questions all day.Price, stock, opening hours and shipping cost, answered in chat hundreds of times.
- Replies stop when the shop closes.Messages that arrive at night wait until morning, and some buyers go elsewhere.
- Documents typed in by hand.Invoices, receipts and delivery notes are retyped into spreadsheets, with errors.
- Answers differ by who replies.Each staff member explains prices and returns a little differently.
- Staff use AI with no rules.Customer data goes into free AI apps with no guidance on what is allowed.
AI We Set Up
Examples of a First AI Task
Customer Service Assistants Answer common questions in WhatsApp or on your website in Bahasa Indonesia, and pass the rest to your staff.
How a Chat Assistant Answers Customers
Example: An Online Skincare Seller in Medan
- The Situation
- Two admins answer about 200 WhatsApp chats a day, most about price, stock and how to order.
- What We Would Set Up
- An assistant that answers from the current price and stock list and hands complaints and health questions to a person.
- What Would Change
- The admins spend their time on the chats that need a person.
Document Readers Read invoices, receipts and delivery notes, and put the figures into your spreadsheet or app for a person to check.
Large Language ModelsHow AI Uses Your Own Business Data
Example: A Shipping Agent in Makassar
- The Situation
- Staff retype delivery notes and invoices from photos every evening.
- What We Would Set Up
- A document reader that turns each photo into a row for a person to approve.
- What Would Change
- Evening data entry becomes a quick check.
Forecasts from Your Sales Suggested reorder quantities and slow-moving items, from your own sales history.
Example: A Minimarket in Yogyakarta
- The Situation
- Supplier orders are written from memory every Monday morning.
- What We Would Set Up
- A weekly reorder suggestion from the cashier app's sales history, sent to the owner on WhatsApp for approval.
- What Would Change
- The owner approves a list instead of writing one.
Automation Agents Routine steps done for you, such as order confirmations, reminders for unpaid invoices and a weekly report, with a person approving anything that moves money.
Example: A Printing Shop in Solo
- The Situation
- The owner sends order confirmations by hand, chases unpaid invoices every Friday and writes the weekly report on Sunday night.
- What We Would Set Up
- An agent that sends order confirmations and reminders for unpaid invoices and drafts the weekly report, with the owner approving every message about money.
- What Would Change
- Friday afternoons go to customers instead of chasing payments.
Search over Your Own Documents Staff ask a question and get the answer from your price lists, procedures and contracts, with the page it came from.
How AI Uses Your Own Business Data
Example: A Clinic Group in Semarang
- The Situation
- New front desk staff ask senior staff the same questions about procedures, prices and insurance rules.
- What We Would Set Up
- An internal assistant that answers from the clinic's own procedure documents and shows the source page; patient records stay out of it.
- What Would Change
- New staff find answers themselves, and no patient data reaches the AI.
Examples to show what is possible. They are not Gravita clients.
Sample: Pilot Plan
- Task
Answer WhatsApp questions about price, stock and how to order.
- Period
Four weeks, agreed in writing.
- The AI May Read
The current price list, the stock list, the shipping table, the return policy.
- The AI Must Never See
Customers' addresses, payment proofs, and any question about skin conditions or health.
- Where It Is Processed
Chosen by the client: an open model in Indonesia, or a global AI service. Written in the contract.
- Rules
Never promise a discount or a delivery date. Hand complaints and health questions to an admin, with the conversation so far.
- How Success Is Judged
An admin marks a sample of answers each week; the pilot passes if the agreed share is correct.
- Cost
The pilot fee as quoted, plus the AI service cost by use, with a monthly spending limit.
- At the End
A short report and a decision meeting: stop, adjust or roll out.
Examples to show what is possible. They are not Gravita clients.
You Choose Where Your Data Is Processed
We use an open model running in Indonesia: in a cloud region in Jakarta that keeps requests inside that region, or on a server in Indonesia. Choice of models is narrower, and we test quality on your own messages.
Requests go to a global model, such as Google's Gemini or Anthropic's Claude. On the provider pages we checked on 25 October 2026, neither offered a way to keep processing inside Indonesia, so your data may leave the country. More model choice.
- The Law
- UU PDP Article 56 sets conditions for sending personal data abroad. We keep personal data out of AI requests where the task allows, and tell you in writing when it cannot be avoided.
- Conversations
- Conversations are stored in your own system for the period you choose, so named staff can review and correct the AI. Our own usage records hold cost and errors, not message content.
- Training
- Before we build, we show you the clause in the AI service's terms that says whether your data is used to train its models, and we choose a service and settings where it is not.
- Keys
- AI service keys stay on the server, never inside an app on a phone.
How the data path works: AI and Your Data in Indonesia.
A person's rights: Your Rights over Personal Data under Indonesia's UU PDP.
This is a summary, not legal advice.
Questions about AI
Will the AI make mistakes?
Sometimes, yes. Every AI can give a wrong answer, which is why a pilot is a good way to start, and why we set rules for what it may answer and pass unclear cases to a person. You see its mistakes before you decide.
Why AI Sounds Sure and Can Still Be WrongChecking an AI Answer before You Use It
How much does it cost to run?
Two parts: the setup, in your itemised quote, and the AI service, charged by use. You get an estimate before work starts, and a monthly spending limit is set.
Does it understand Bahasa Indonesia and regional languages?
We test the AI on your own customer messages, including slang and regional words, during the pilot, and you see the results before you decide.
Will AI replace my staff?
It takes over repeated work, such as the price and stock questions in the skincare example's 200 chats a day, so staff spend more time on customers and decisions. A person stays in charge of anything the AI cannot answer with confidence.
WhatsApp Business already has AI. Why pay for this?
Use it if it fits. An assistant we set up is worth it when it must use your own stock, prices and rules, or connect to your other systems.
What does AI need from us?
Your current price lists, procedures or past documents, and one person who can check the AI's work during the pilot.
What can AI not do?
It cannot take responsibility for a decision, it knows nothing you have not given it, and it should not give customers medical, legal or financial advice without a person checking.
Can our office use AI safely without a project?
Yes. Consulting Services include a short AI use policy for your team.
Applications and AI Explained
Kinds of AI1 / 9
The AI Family
Artificial intelligence is a family of four nested layers: AI is the broad field, machine learning learns from examples, generative AI creates drafts, and agents use tools to act.
- Ring 1 (AI) lights as the outer ring: any program that handles tasks needing human-like understanding, recognition or prediction.
- Ring 2 (Machine Learning) lights inside AI: instead of hand-written rules, the computer learns patterns from thousands of past examples.
- Ring 3 (Generative AI and LLMs) lights inside ML: models trained on vast collections of text and media that create new drafts.
- Ring 4 (AI Agents) lights at the centre: a language model connected to tools, rules and human approval to complete a multi-step job.
Ring 1 (AI) lights as the outer ring: any program that handles tasks needing human-like understanding, recognition or prediction.
Next in This PathDefinition
Artificial intelligence is a family of four nested layers: AI is the broad field, machine learning learns from examples, generative AI creates drafts, and agents use tools to act.
Everyday Comparison
Think of a kitchen team's skills. The whole kitchen is AI, where machines help with thinking tasks. Machine learning is the fruit buyer who learned to spot ripe mangoes from thousands of past crates. Generative AI is the assistant who drafts the menu cards and daily notices. An AI agent is the trusted helper who checks the shelves, drafts the supplier order, and brings it to you to sign.
What It Means for Your Business
Knowing which layer fits your problem stops you from buying a complex tool when a simpler one works better. Forecasting stock needs machine learning on past sales, drafting product descriptions uses generative AI, and capturing free-text chat orders into your sheet uses an agent with your approval. Always keep a person checking money, prices and promises to customers.
Examples
- Spam filters and sales forecasts built with machine learning on past records
- Drafting Indonesian and English product descriptions or summarising supplier contracts with generative AI
- Reading incoming WhatsApp order messages and preparing draft rows in your order sheet with an AI agent
Kinds of AI2 / 9
What Is AI
- People collect many examples of the task, called data.
- A program studies the examples and finds patterns in them.
- The result is a trained model.
- You give the model a new input, such as text, a photo or numbers.
- The model gives an output: an answer, a label or a prediction.
- A person checks the output and uses it.
People collect many examples of the task, called data.
Next in This PathDefinition
AI, short for artificial intelligence, is computer technology that does tasks which normally need human thinking, such as understanding language, recognising pictures or making predictions.
Everyday Comparison
AI is like a trainee who learned from thousands of past examples. The trainee is fast with familiar patterns, can still make mistakes, and needs checking.
What It Means for Your Business
For most small businesses, AI already arrives inside chat and platform apps they use. It helps with drafting texts, sorting, translating and answering common questions. It can be wrong, so keep a person in charge of the final answer, and keep customers' personal details out of public AI tools.
Examples
- Spam filters and voice typing on a phone
- Translation tools
- Chat assistants such as Gemini and Microsoft Copilot
Sources and Related Topics
Kinds of AI3 / 9
Machine Learning
Machine learning is a way of building AI in which a computer finds patterns in data and uses them to make predictions, with no step-by-step rules written by hand.
- Gather data where the right answer is already known, such as photos marked ripe or unripe.
- Feed the data to a learning program.
- Training adjusts the model's settings until its predictions are good.
- Test the model on new data it has never seen.
- Use the trained model on real inputs, which is called inference.
- Retrain with fresh data when the results start to slip.
Gather data where the right answer is already known, such as photos marked ripe or unripe.
Next in This PathDefinition
Machine learning is a way of building AI in which a computer finds patterns in data and uses them to make predictions, with no step-by-step rules written by hand.
Everyday Comparison
Think of a fruit seller who learns to spot a ripe mango after handling thousands of them. Nobody handed the seller a rulebook; the experience did the teaching. A machine learning model learns from examples in a similar way.
What It Means for Your Business
Machine learning suits decisions you repeat often and for which you hold plenty of past records, such as which products sell out, which customers return, or which transactions look unusual. It needs enough clean data. With little data, a simple report or rule may serve you better.
Examples
- Product recommendations on shopping apps
- Fraud checks on card and wallet payments
- Sales forecasts based on past orders
Sources and Related Topics
Kinds of AI4 / 9
Generative AI
- You type a request, called a prompt.
- The model breaks your request into small pieces it can process.
- Using patterns learned from very large collections of data, it builds new content piece by piece.
- The draft appears on your screen.
- You check it and ask follow-up requests to improve it.
- You use the final version.
You type a request, called a prompt.
Next in This PathDefinition
Generative AI creates new content, such as text, images, audio or computer code, in response to a request you type.
Everyday Comparison
It is like a well-read assistant who drafts the first version of your poster text or product description when you ask. You still read it, correct it and decide what goes out.
What It Means for Your Business
Generative AI is handy for first drafts: captions, product descriptions, replies to customers, translations and promotional images. Its output can contain mistakes or invented facts, so check prices, specifications and legal claims before you publish. Keep customers' personal details out of public chat tools.
Examples
- Text assistants that draft messages and descriptions
- Image generators for promotional pictures
- Coding assistants that suggest program code
Sources and Related Topics
Kinds of AI5 / 9
Large Language Models
A large language model, or LLM, is an AI model trained on an immense amount of text so that it can understand and write human language.
- The model is trained beforehand on very large amounts of text.
- Your message is split into small pieces called tokens, which can be words, parts of words or single characters.
- The model works out which token fits best next, again and again.
- The tokens are joined into a reply.
- You read the reply. It can sound convincing and still contain false statements, which is called a hallucination.
The model is trained beforehand on very large amounts of text.
Next in This PathDefinition
A large language model, or LLM, is an AI model trained on an immense amount of text so that it can understand and write human language.
Everyday Comparison
Picture someone who has read a whole library and is very good at finishing your sentences. The person sounds sure of every answer, even on the days the answer is wrong.
What It Means for Your Business
LLMs power chat assistants that can answer customer questions, summarise documents and draft replies. They can invent facts, can reflect bias in the text they learned from, and cost real computing power to run. Give them your own facts to work from, and keep a person checking anything that touches prices, promises or legal matters.
Examples
- General chat assistants such as Gemini, Microsoft Copilot and ChatGPT
- Customer-service chat on a shop's website
- Tools that summarise long documents
Sources and Related Topics
Kinds of AI6 / 9
How AI Uses Your Own Business Data
Retrieval-augmented generation, or RAG, lets an AI answer from your own price lists and policies by looking up the right page before it writes a reply.
- N1 Question ("Can I return an opened item?") enters and lights the Open-Book Shelf (N2 Your Documents) first.
- N2 Your Documents searches your current policies and price lists, pulling out the matching passages (N3 Found Passages).
- Arrow N3 -> N4 lights as the question and the found passages travel together to N4 Language Model.
- N4 Language Model writes N5 Grounded Reply using those passages and attaches N6 Source Reference to the exact clause.
- Toggle Memory Only (No Shelf): the model guesses without your documents; Toggle Fine-Tuning: teaches style, not daily prices.
N1 Question ("Can I return an opened item?") enters and lights the Open-Book Shelf (N2 Your Documents) first.
Next in This PathDefinition
Retrieval-augmented generation, or RAG, lets an AI answer from your own price lists and policies by looking up the right page before it writes a reply.
Everyday Comparison
Asking a general AI about your shop is like quizzing a stranger from memory: the stranger may guess and sound confident. RAG is like giving your clerk an open-book shelf of your current price list, return policy and stock sheet, and requiring every answer to point to the exact page it came from.
What It Means for Your Business
When your prices, stock rules or SOPs change every week, open-book lookup (RAG) keeps the AI current without expensive retraining, because updating a document on the shelf updates the next answer. It still needs clean, up-to-date documents and guardrails so it declines questions outside your files instead of guessing. If you want to test this on one task first, see the pilot in [AI Solution](#ai-solution-pilot).
Examples
- A customer-service assistant that answers shipping, warranty and return questions from your shop's written policy
- An internal search tool where new staff ask how to handle a wholesale discount and get the exact SOP paragraph
- A product catalogue helper that quotes current wholesale tiers from your latest approved price sheet
Kinds of AI7 / 9
AI Agents and Automation
An AI agent is a program that reaches a goal by using tools like search, databases or apps, while automation runs repeated steps without manual effort.
- You give the agent a goal and its rules.
- The agent breaks the goal into smaller tasks.
- It uses tools, such as a database, a search or another app, to get what it lacks.
- It acts, for example sends a message or updates a record.
- It reviews the result and adjusts its plan.
- It reports back, or asks for your approval when the rules say so.
You give the agent a goal and its rules.
Next in This PathDefinition
An AI agent is a program that is given a goal and carries out the steps to reach it, using tools such as search, a database or other apps; automation means a repeated task runs without a person doing each step.
Everyday Comparison
Think of a capable shop assistant. You say "restock what is almost gone." The assistant checks the shelves, looks at the supplier list, sends the order and tells you when it is done.
What It Means for Your Business
Plain automation follows fixed rules, such as "when an order arrives, send the receipt", and is predictable. An AI agent helps where inputs vary, for example reading orders typed in free text in a chat. Start with one repetitive task, and keep a person's approval for money and for promises to customers. Gravita builds AI tools, agents and automation for a client's work.
Examples
- Capturing orders from chat messages into a sheet
- Reading invoices and filing the figures
- A weekly sales summary sent to the owner
Sources and Related Topics
Kinds of AI8 / 9
How a Chat Assistant Answers Customers
A business chat assistant is a program connected to your messaging channel that answers routine customer questions from your approved documents and hands complex cases to a person.
- N1 Customer Message arrives from your chat channel and enters N2 Guardrail Check.
- N2 Guardrail Check tests whether the question is safe and inside the topics the assistant is allowed to handle.
- Arrow N2 -> N3 lights as N3 Document Lookup searches your approved price list, FAQ and shop rules.
- N4 Language Model drafts N5 Reply from those found facts, staying inside its written boundaries.
- If the question is confident and routine, Path A (Send Reply) delivers N5 Reply to the customer.
- If the question touches money, complaints or unknown facts, Path B (Hand to a Person) passes the chat to staff.
N1 Customer Message arrives from your chat channel and enters N2 Guardrail Check.
Next in This PathDefinition
A business chat assistant is a program connected to your messaging channel that answers routine customer questions from your approved documents and hands complex cases to a person.
Everyday Comparison
Think of a front-desk receptionist with a clear rulebook. When a visitor asks for your opening hours, catalogue prices or return rules, the receptionist checks the counter binder and replies right away. The moment a visitor asks for a special discount, reports a damaged parcel or wants to change a payment, the receptionist calls you over with the visitor's note already written down.
What It Means for Your Business
Most Indonesian businesses already talk to buyers in chat, so an assistant that handles repetitive questions about stock, sizes and opening hours saves evening typing. An assistant without boundaries can make promises you did not approve, so set strict guardrails and require human approval for discounts, refunds and disputes. If you want to see how a first assistant is scoped, see the [sample pilot plan](#ai-solution-sample).
Examples
- Answering late-night buyer questions about stock availability, size charts and delivery cut-off times
- Collecting a buyer's name, order items and delivery address into a clean summary before staff confirm the total
- Handing a warranty complaint straight to the shop owner with the customer's order number and photos attached
Kinds of AI9 / 9
AI and Your Data in Indonesia
Using AI sends the text or files you include to the server running the model, which may sit inside Indonesia or in an overseas data centre.
- N1 Your Prompt and Attached Document leave your app to be processed by an AI model.
- N2 Classification Check marks whether the request carries customer names, health records, children's data or personal finances (UU PDP Article 4).
- Route A (Model in Indonesia) sends the request to a server in an Indonesian data centre, keeping processing inside the country.
- Route B (Global AI Provider) sends the request across the border to a global model server when you choose wider model options.
- N5 Returned Answer arrives back in your app, noting that transfers abroad must meet the statutory conditions of Article 56.
N1 Your Prompt and Attached Document leave your app to be processed by an AI model.
Next in This PathDefinition
Using AI sends the text or files you include to the server running the model, which may sit inside Indonesia or in an overseas data centre.
Everyday Comparison
Think of sending a ledger folder to a typing bureau. In Route A, the bureau sits in a Jakarta building you rent, and your pages stay inside the country. In Route B, the courier flies the folder to an overseas office where a specialist reads it, writes the summary, and sends the reply back.
What It Means for Your Business
For full statutory compliance dates under UU PDP, see [Personal Data Duties for Small Businesses](#tip-life-personal-data-duties). Never paste customer names, phone numbers, ID numbers or financial records into public consumer chat tools, and decide before building whether your AI workload stays on servers in Indonesia or uses a global provider API. If you want to choose where your data is processed, see [AI Solution, You Choose Where Your Data Is Processed](#ai-solution-data).
Caption: This is a summary, not legal advice.
Examples
- Running an open-weights model on a Jakarta cloud server so patient or financial records stay in Indonesia
- Stripping customer names and phone numbers before sending product questions to a global language model API
- Recording customer consent and transfer safeguards when a business chooses an overseas AI provider
This is a summary, not legal advice.