Plenty of managers have been handed the same one-line brief: “do something with AI”. Nobody said where. If you can’t yet see where to begin, that’s normal. In Deloitte’s 2026 survey of Thai organisations, 40% said they couldn’t find the right use case.
Starting doesn’t solve it either. Deloitte found 81% still at the pilot stage, and NECTEC’s AI Readiness Survey 2026 found that 86.4% can’t say what AI has achieved for them; 49.4% have never measured it at all.
The cause is rarely the wrong tool. It is usually the wrong first task. The idea that sounds most exciting in the meeting room tends to be the hardest to measure and the one whose data is least ready.
Write down five to ten tasks your team repeats every day, then put each one through these seven questions.
1. How often does it happen?
AI pays off on work done dozens or hundreds of times a week: answering staff questions about benefits, sorting incoming documents, summarising field reports. A task that happens twice a year will never produce enough numbers to show a difference, however well AI does it.
2. What does it cost today?
If you can’t answer, there will be nothing to compare against afterwards. It doesn’t need to be research. “About 20 minutes each, roughly 150 a week” is enough. If you have no figure at all, count for one real week before you start.
In Deloitte’s survey, what Thai organisations got from AI was mostly lower cost (50%) and greater efficiency (48%); only 9% said it unlocked new revenue. So measure your first task in time, cost or errors, not sales.
3. Where is the data, and how ready is it?
AI answers only as well as the data it can reach. Ask which system holds it, who owns it, and whether it’s current. NECTEC found only 40.8% of Thai organisations have data of AI-ready quality. If this task’s data lives in everyone’s own spreadsheet, or in one person’s head, it isn’t your first task yet.
4. Who owns the work, and will they use the result?
You need someone who does this work every day and wants it faster, to help test and to say whether the output is good enough. AI built by IT for another department, without the people who own the work, tends to end as a demo nobody opens.
5. What happens when the AI is wrong?
Start where a person can easily check the output before it’s used: drafting replies for staff to send, or sorting documents for someone to confirm. Not a task where AI decides on its own and a customer, money or someone’s rights are affected straight away. It should hand over to a person when it isn’t sure, and show which document an answer came from.
- 1 AI drafts, a person checks Replies for staff to send, or documents sorted for someone to confirm.
- 2 It shows its source Every answer says which document it came from, so checking is quick.
- 3 Unsure? It hands over When it isn't confident, a person takes the case instead of the AI guessing.
- 4 Not a first task: AI decides alone A customer, money or someone's rights affected straight away.
6. Does it need personal or confidential data?
It still can, but before you start you must be able to say where the data goes, how long it’s kept and who can see it. PDPA applies to everything AI processes. A first task that needs no sensitive data moves faster, because there are fewer sign-offs to collect from legal and the DPO.
7. If it works, can it spread?
A good first task is both a result and a lesson. If an assistant that answers benefits questions works, purchasing or sales can use the same pattern on their own documents. A one-off that can’t be repeated makes the next task no easier.
Rank them on one page
Score each task from 1 to 3 on two axes: value (questions 1, 2 and 7) and readiness (questions 3, 4, 5 and 6). The task that scores high on both is your first. One with high value but unready data becomes your second, and the work on its data starts now.
- 1 Your first task High on both. Write down its measures and today's figures before you start.
- 2 Your next task High value, data not ready. Start the data work now.
- 3 Not yet Low value: however ready it is, it won't show a difference.
Pick one task, then write down three things before you begin: one or two measures, today’s figure for each, and how you’ll collect the figure afterwards. At the end you’ll have one page that says what changed, which is what most organisations in NECTEC’s survey still don’t have.
- 1 One or two measures Time, cost or errors: for example, minutes per request.
- 2 Today's figure About 20 minutes each, roughly 150 a week. Count for a week if you have to.
- 3 How you'll collect it afterwards Decided now, so the after figure is counted the same way.
Signs a task shouldn’t be your first
- Nobody can say which number would show it worked.
- Three or more systems need connecting before you can start.
- No one on the team does this work every day.
- The AI has to make decisions for people from day one.
- You’re choosing an AI tool or vendor before you know which task it’s for.
Want help choosing?
AI Start is a shorter Blueprint for exactly this. One or two workshops with your team to list your candidate tasks and rank them on the seven questions above, by value and by data readiness. It ends with one first task, the one or two measures for a before-and-after report, and a fixed quote to build a working prototype on your own data. Never just a slide deck.