Which business processes are worth automating with AI?
Not every process deserves AI. Here is how to spot the ones that do, score them in an afternoon and start with a small, safe first project.
Key takeaways
- Repetitive, document-heavy work with rules plus a little judgement is where AI pays back first.
- Score before you build: frequency × time per task × data available × risk if wrong gives you a ranked shortlist.
- Keep a human in the loop wherever a mistake costs money or trust: the AI drafts, a person approves.
- Start small and measure. One process, one owner, a few weeks of real use, then expand.
Most owners we talk to are not short of ideas for AI. They are short of a way to choose. Everything looks possible, a vendor says everything is easy, and the first project ends up being whatever was most exciting rather than whatever was most useful.
This article gives you a plain method for picking: which processes suit AI, which do not, how to score them in an afternoon, and how to start without putting the business at risk.
Which business processes suit AI automation?
AI tools today are good at work that has a pattern but not a rigid script. In practice, the processes that pay back first share four traits: they repeat often, they involve text or documents, there are rules to follow plus some judgement, and the cost of a draft being slightly off is low because someone checks it.
Common examples in Thai SMEs:
- Quoting and proposals. Assembling a first draft from a brief and from similar past jobs.
- Data entry from documents. Reading invoices, purchase orders or forms and filling in your system.
- Answering the same customer questions. Opening hours, delivery terms, product details, booking rules.
- Reporting. Turning raw sales, stock or booking data into a weekly summary people actually read.
- Searching old records. Finding what a job cost last time, or who supplied what, without opening twenty files.
- Drafting listings and content. Product names, descriptions, categories, English versions of Thai text.
Which processes should you not automate with AI?
Be cautious with work that happens rarely, where one mistake is expensive, and where nobody would review the output. A once-a-year contract negotiation, a hiring decision or a legal commitment gives you little data to learn from and a high price for being wrong. AI can still help as a research assistant in those cases, but it should not be the decision-maker.
Also skip processes that are simply unclear. If two managers describe the same task differently, the first job is to agree how it works, not to buy software.
A simple scoring method: frequency × time × data × risk
You do not need a consultant to rank candidates. List your processes and score each one from 1 to 5 on four questions:
- Frequency. How often does it happen? (1 = a few times a year, 5 = many times a day)
- Time per task. How long does one round take a person? (1 = minutes, 5 = hours or days)
- Data available. Do you already have the documents, examples or records the AI would learn from or draw on? (1 = none, 5 = plenty and reasonably tidy)
- Safety if wrong. If the AI gets it wrong, how bad is it? (1 = costly or hard to undo, 5 = harmless and caught in review)
Multiply the four scores and rank by the result. The highest numbers are your shortlist. The table below is purely illustrative: the processes are generic and the scores are made up to show how the method works, not taken from any client.
| Process (illustrative) | Frequency | Time per task | Data available | Safety if wrong | Score |
|---|---|---|---|---|---|
| Drafting quotes from past jobs | 3 | 5 | 4 | 4 | 240 |
| Answering repeat customer questions | 5 | 2 | 4 | 4 | 160 |
| Typing supplier invoices into the system | 4 | 3 | 5 | 3 | 180 |
| Weekly sales report | 2 | 3 | 5 | 5 | 150 |
| Writing product descriptions | 4 | 2 | 3 | 5 | 120 |
| Annual contract renewal terms | 1 | 4 | 2 | 1 | 8 |
The ranking tells you something the gut feeling often misses. Quoting looks less frequent than customer questions, but each quote is long enough that it comes out on top. Contract renewals sit at the bottom however clever the tool is.
Treat the scores as a conversation starter. Ask the person who does the task to score it, not only the manager.
Keep a human in the loop
The safest pattern is that the AI prepares and a person approves. Our own product TellSell works this way: the owner sends a product photo in LINE and says the price, the AI writes the name, description, category and English text, and nothing goes live until the owner taps Publish. If a price looks unusual, the AI asks first rather than guessing. The result is fast, but the owner stays in charge. You can read how that was designed in the TellSell case study.
Decide for every process where the review step sits, who does it, and what they are looking for. If you cannot name the reviewer, the process is not ready.
Start small, measure, then expand
A sensible first project is narrow: one process, one team, one owner, a few weeks of real use. Before you start, write down how long the task takes today and how often it needs correcting. After a month, measure the same two things.
Some projects are larger. For a production company with ten years of invoices and quotes, we built a master price list and a proposal generator on top of its own archive, so proposals now start from a priced draft instead of a blank page; the Bangkok Productions case study shows how. Your first step does not have to be that large.
If the first project works, expand to the next-highest score on your list, reusing the data and habits you have already built.
Common mistakes when automating with AI
- Automating a broken process. AI makes a messy process faster, not better. Fix the steps first.
- No owner. Someone must be responsible for the result, for the review step and for deciding when the tool needs adjusting. Shared ownership usually means none.
- Ignoring data quality. Duplicates, inconsistent names and missing fields lead to confident but wrong answers. Budget time for cleaning; it is often the largest part of the work. Our guide on turning an archive of invoices and quotes into a knowledge base explains how.
- Forgetting privacy. Supplier prices, margins and customer records need care once an AI tool touches them. See using AI on company data privately and in line with PDPA.
- Starting with the biggest idea. Large first projects stall. Small ones teach you what your team will actually use.
Next step: shortlist three processes
This week, list ten processes, score them with the four questions above and pick the top three. Then talk to the people who do them. If you want a second opinion, or want to know what building a tool around your shortlist could involve, our overview of what a custom business app costs in Thailand explains what drives the price, and the web and application projects on our homepage show the kind of work we do. You can also send us your shortlist and we will tell you honestly which item looks worth building first.
Frequently asked questions
Which business processes should a small company automate with AI first?
Start with work that is repetitive, takes real staff time, and already leaves a trail of text or documents: answering the same customer questions, preparing quotes, typing data from documents into a system, or producing routine reports. Avoid rare, high-stakes decisions that nobody would review.
Do we need a lot of data before AI can help?
It depends on the task. Drafting product descriptions or answering from a document set needs little. Estimating prices or finding similar past jobs needs a decent archive, and the quality of that archive matters more than its size.
Will automating a process with AI replace my staff?
In the projects we describe here the aim is to remove the repetitive drafting and searching so people can spend time on checking, deciding and talking to customers. The sensible design keeps a person responsible for the final result.