- The first job AI can take in every department has the same shape: read a document, type it into a system, pass it to someone to check.
- Every example here has four fields: what the AI reads, what it produces, who approves, and what to do in the first week.
- Pick your first three uses from different departments that share the same parts.
- Leave for later any task nobody checks, any mistake you cannot undo, and any data still spread across several places.
Every department hides the same move
Walk past the desks in your office. Accounting types figures from invoices into the accounting program. Sales copies customer messages from LINE into a spreadsheet. Admin renames files that arrive in chat. Different documents, same move: read, retype, hand over for checking.
So every AI use case below is written in the same four fields. This is chapter 2 of AI for business: a guide from benefits to building, and it follows the measuring method in chapter 1. How to pick the first department is in If you want to cut staff costs, which department should start using AI first?
Sales: capturing leads and drafting quotes
Leads: the AI reads customer messages from LINE OA, the website form and email, and opens a record in the CRM with the product of interest and how urgent it looks. No approval is needed; salespeople only decide whom to call first. First week: the sales manager compares the AI's urgency ratings on last month's messages with their own.
Quotes: the AI reads the request, the standard prices and the customer's purchase history, and drafts a quote for the salesperson to check and send. Discounts above the limit wait for the sales manager's approval. First week: write the discount rules on one sheet of paper. See Sales and CRM workflow.
Customer service: repeat questions and order status
Answers from documents: the AI searches product manuals, warranty terms and past replies using RAG, then answers in LINE OA or drafts a reply for staff to send. The team lead spot-checks replies every week. First week: pull the most-asked questions from chat history and have the team write the correct answers beside them.
Order status: the AI reads the order number in a message, looks it up in the ERP or the courier's system, and replies with the delivery date and tracking number. It only reads, so no approval is needed. Complaints go straight to a person. First week: find out how the courier's system lets you read status. See Order and status tracking agent.
Accounting: invoices, reconciliation and collections
Invoices: the AI reads Thai-language invoices and tax invoices with OCR and matches them to the purchase order and goods receipt. Matches are posted as payables in the accounting program you already use, such as Express, PEAK, FlowAccount, SAP Business One, Xero or Odoo. Mismatches go to the exception queue for an accountant to decide, and every payment waits for an authorized approver. First week: collect last month's real invoices, handwritten ones included.
Reconciliation and collections: the AI matches the bank statement against entries in the system, sends only the unmatched pairs to an accountant, and drafts payment reminders for a person to send. First week: write down the reminder steps you actually follow. See Accounting and finance workflow and Thai document OCR with AI.
Purchasing: requests, quote comparison and goods receipt
Purchase requests: the AI reads the request and a photo of the supplier's quote, checks it against the remaining budget, past prices and the vendor list, and sends a three-line summary in LINE to the approver for that amount. First week: write down the table of amounts and approvers.
Quote comparison and goods receipt: the AI turns quotes in different formats into one table for purchasing to choose from. When goods arrive, it compares a photo of the delivery note with the purchase order and alerts the warehouse lead only to lines that do not match. First week: keep the latest round of quotes from your main suppliers. See Purchase request and approval workflow and Purchasing and goods-receipt workflow.
HR: staff questions, onboarding and applications
Staff questions: the AI reads the company rules, the benefits handbook and announcements, and answers staff in LINE about leave, claims and social security. Personal matters such as pay go straight to HR. First week: HR picks the current version of each document the AI will answer from. See HR and IT helpdesk agent.
New hires and applications: once the contract is signed, the AI requests accounts from IT, sends the tax and social security forms, and follows up until everything is in. HR confirms before the start date. First week: list what each position needs. See Employee onboarding workflow. For applications, the AI ranks CVs against the posted requirements with its reasons, and the interviewer decides on every candidate. First week: set, under Thailand's PDPA, how long applicant data is kept and who can see it.
Admin and documents: filing, drafting and data entry
Filing: the AI reads files that arrive by LINE and email, tells a tax invoice from a contract or a government letter, names each file to your standard and files it. Admin staff see only the files the system is unsure about. First week: agree the naming standard and folder structure. See Client document workflow.
Drafting and data entry: the AI fills letter and contract templates with data from your systems, and an authorized person checks and signs. For websites with no API, such as bank portals or government filing sites, RPA types the data in, and a person checks the final screen before submitting. First week: pick the template you use most. The bigger picture of office automation is in our AI office article.
Operations and logistics: closing jobs and planning trucks
Closing jobs: drivers photograph the delivery note the customer signed and send it to LINE. The AI reads the job number and the signature, closes the job and adds it to the billing run. The operations lead checks only unclear photos. First week: collect the photos drivers actually take, tilted and dark ones included. See Job sheet and billing workflow.
Truck planning: the AI reads confirmed orders, delivery areas and available trucks, and proposes the next day's runs for the planner to confirm or change. First week: write down the rules the planner keeps in their head.
Management reporting: summaries from the source systems
Weekly summary: the AI reads the pipeline from the CRM, receivables from the accounting program and stock from the ERP, and writes a one-page summary of what changed since last week. Each department head confirms their own numbers before it goes out. First week: agree which system each number comes from.
Questions on the numbers: an executive asks which customer has owed the longest, and the system answers from the database, read-only. First week: check whether the numbers still live in spreadsheets passed around in chat. If they do, this use has to wait.
Hypothetical case
An electrical equipment importer and wholesaler in Bangkok (name and details invented) receives supplier invoices, LINE OA leads and import documents every day. The owner wanted to start in every department at once, and in the end chose three uses that share the same parts.
The three were invoice reading in accounting, filing import documents in admin, and lead capture in sales. The first two use the same document reader. The third reuses the LINE approval step built for accounting, this time for discounts above the limit. Month one opened accounting only. Month two opened admin without building a new reader. The mistake they hit was handwritten invoices from small suppliers, where the system misread a total by one digit, so the team now sends every low-confidence invoice to a person whole.
Choosing the first three uses across departments
A good first use shows three signs: it repeats, it comes with a document, and someone already checks it. Use this card in one meeting with all department heads.
The first-use table: repeats, has a document, has a checker
- Step 1: list. Each head writes up to five repeated tasks in the four fields used here.
- Step 2: score. Repeats, has a document, has a checker: 0 to 3 each.
- Step 3: find shared parts. Circle high scorers from different departments that use the same document reader, document search or approval step.
- Step 4: pick three. Three uses from three departments. Get the first one stable before opening the second.
Not yet: Tasks nobody checks, mistakes you cannot undo, or data still in several places.
| Department | First use | AI reads | AI produces | Who approves |
|---|---|---|---|---|
| Sales | Leads | LINE OA, web form, email | CRM record | Salesperson |
| Customer service | Answers from documents | Manuals, past replies | Reply in LINE OA | Team lead spot-checks |
| Accounting | Invoices | Invoices, purchase orders, goods receipts | Payables entry | Accountant, payment approver |
| Purchasing | Purchase requests | Request, budget, past prices | Summary in LINE | Approver by amount |
| HR | Staff questions | Rules, benefits handbook | Answer in LINE | HR takes personal matters |
| Admin | Filing | Files from LINE and email | File name and folder | Admin staff |
| Operations | Closing jobs | Delivery note photos | Closed job, billing run | Operations lead |
| Management | Weekly summary | CRM, accounting, ERP | One-page summary | Each department head |
Which uses can start today, and which must wait for data
A use that passes every item below can start today. When an item fails, fix that first; it is usually the data or the checker.
- The task repeats every week, and the person who does it can explain it in five minutes.
- Input arrives as files, messages or photos a system can open.
- The output goes into a system with an API or a file import.
- Someone checks this work today and can say what they check.
- A mistake can be fixed before it reaches a customer, a bank or a government office.
- You have last month's real documents and questions to test with.
- An owner in the department will watch the demo every week.
For the development team · Four shared parts
Every example is built from four parts. The document reader is OCR plus an LLM extracting fields to a schema per document type; a document with a low confidence score goes to the exception queue whole. Document search is RAG over documents that have owners and versions. The approval step is a policy by amount, sent to LINE, with a timeout and a backup approver; every write to a system goes through it. The fourth part is an audit log of input, output and approver. Build the four as shared services, so each new department adds only its own schema and policies. Writes to the ERP must be idempotent, keyed on the document number.
We start with one department, then reuse the parts
Which invoices can be paid is a rule your staff know, so DNA Maker starts by sitting with the first department you choose. We follow real documents to the person who approves them, write each task in the four fields used here, and agree which fields the AI takes and which stay with people.
Then we build the shared parts for that department and connect them to the ERP, CRM or accounting program you already run. We test with your real documents and questions, show a demo on real data every week, and fix the scope and price before work starts. The next departments reuse the same parts and add their own rules. We do this kind of work for banks, hospitals and large Thai groups, and we offer in-house AI training for your team. See AI Workflow, or describe the work to Dream on the chat page.
SOFTWARE ENGINEERING GLOSSARY
Software engineering glossary
| Term | What it is | A simple example | What executives should ask the development team |
|---|---|---|---|
| OCR | Reading text in an image or PDF into data | Reading a tax invoice photographed and sent by LINE | Have you tested it on our Thai and handwritten documents? |
| RAG | Making the AI search company documents before it answers | Answering warranty terms from the latest manual | Who updates the documents the system answers from? |
| Exception queue | Items the system is unsure about, waiting for a person | An invoice whose total does not match the purchase order | Who watches this queue, and how often? |
| Human approval | A point where a person must approve before the system does anything tied to money | Approving a payment run in LINE | Can we change the limits and approvers ourselves? |
| Confidence score | A score for how sure the system is | A handwritten total scores low, so it goes to a person | Who can change the threshold for the queue? |