Short answer
Start from work that is slow, expensive or often wrong, not from where to put AI: most AI projects that fail had good technology but started from the tool. AI suits high-volume work that tolerates an occasional error; legal or financial decisions still need a person to sign. Pilot one measurable job with stop criteria set on day one.
Key points
- Ask which work is slow, expensive or error-prone enough to fix; whether AI belongs follows, and often it doesn't.
- Before approving budget, answer four questions: what decision it drives, what being wrong costs, whether data is usable, and the non-AI fix.
- A practical line: if one wrong answer in twenty is acceptable, AI probably fits; if not, design in a human check.
- The build is usually the smallest cost; data preparation, quality evaluation and monthly usage are the ones people forget.
- Design it to be removable, so you can switch it off with the original process still running.
"Where should we use AI?" is a backwards question
Most AI projects that fail to pay off had perfectly good technology; they failed because they started from the tool instead of the problem. Once a company decides "we need AI", the next question becomes "where can we put it", and the team is now finding a home for a tool. Nobody is fixing anything yet.
The common shape: a team builds an impressive demo, leadership is pleased, and three months later nobody uses it. It didn't change anyone's decision. It showed something was possible, which is not the same as making work better.
The question that works is "which work in this business is slow, expensive or frequently wrong enough to be worth fixing?" Answer that and whether AI belongs follows on its own. Often it doesn't.
Starting from "we need AI"
- Asks where AI could go
- Finds a home for a tool
- Impressive demo that changes no decision
- Unused three months later
Starting from the problem
- Asks which work is slow, costly or error-prone
- Checks whether fixing it pays
- AI or not follows from the answer
- Often solved without AI
What should you answer before approving an AI budget?
Four things: what decision the work drives, what being wrong costs, whether the data is usable, and how you would solve it without AI.
First, what decision does this work drive? If the system produces an answer and nobody does anything differently, it isn't worth investing in yet. A system that gives good answers nobody acts on is a more expensive report.
Second, what does being wrong cost? AI does not get every answer right, and never will. A mistake that costs a few minutes and a mistake that pays the wrong supplier are entirely different propositions. This decides whether a person must check the output before it is used.
Third, does the data exist in usable shape? Data spread across differing spreadsheet versions, or typed by hand without a standard, needs tidying first, and the tidying is usually a bigger job than the AI.
Fourth, how would you solve this without AI? This matters most and gets skipped most. Much work is fixed by changing a process, designing a better form, or writing ordinary rules: cheaper, faster, and far easier to explain to an auditor.
- What decision?If nobody acts differently on the answer, don't invest yet.
- Cost of being wrong?Decides whether a person checks output before use.
- Usable data?Scattered or unstandardised data needs tidying first.
- Fix without AI?A process change, better form or plain rules is often cheaper.
What work does AI genuinely help with, and what shouldn't it be trusted with yet?
AI helps with work that is high in volume, loose in format, and tolerant of occasional error: summarising long documents, grouping large numbers of customer questions, drafting a first version for a person to edit, or pulling fields from documents that never arrive in the same layout. That work used to consume a lot of human time and never needed accounting-grade precision.
It shouldn't decide alone on work that must be right every time, must be explainable, or carries direct legal and financial consequence: tax calculation, credit approval, medical diagnosis, figures filed with a government agency. AI can assist, but a named person still signs.
The line that holds up in practice: if the system is wrong once in twenty, can you live with it? If yes, AI probably fits. If no, design a human check into the flow, or accept that it isn't AI work yet.
AI should not decide alone
- Tax calculation
- Credit approval
- Medical diagnosis
- Figures filed with a government agency
AI helps
- Summarising long documents
- Grouping large volumes of customer questions
- First drafts for a person to edit
- Pulling fields from documents in varying layouts
What costs do people forget in an AI project?
The forgotten costs are preparing the data, evaluating answer quality systematically (not trying it a few times and feeling good), and paying people to check output in the early period. Initial development is usually the smallest line.
The other invisible cost is per-call spend. A system a few people press a handful of times a day and one fired automatically on every transaction differ many times over in a month. Estimate real usage up front, not trial usage.
The most expensive cost is tearing it out later. AI wired into every step from day one is hard to remove once it turns out not to pay, because by then it is part of how work is done. Designing it to be removable is always cheaper.
How do you start with AI without betting the budget?
Start with one measurable job. Pick work with a clear time cost today, such as three staff spending two hours a day keying data off purchase orders. Numbers like that let you answer whether it paid, instead of guessing.
Write the success and abandonment criteria as numbers on day one. Teams without an abandonment criterion keep failing projects alive far too long, precisely because so much has been spent.
Keep a person in the decision loop early: the system proposes, a person confirms, and you track how often and where answers are corrected. That record is worth more than any demo, because it tells you whether the system is ready to work unattended.
And build it to come out. If month three shows it isn't paying, you should be able to switch it off with the original process still running. That is the difference between a controlled experiment and a bet.
- One measurable jobFor example, 3 staff spending 2 hours a day keying purchase orders.
- Success and stop criteriaWritten as numbers on day one.
- System proposes, person confirmsTrack how often and where answers get corrected.
- Build it to come outSwitch it off with the original process still running.
Business automation: n8n, Power Automate, AI agents or a custom build
Checklist before approving an AI project
- The work to be fixed is named, with a number for what it costs in time or money today
- You can say whose decision the output changes, and how
- You have assessed what a wrong answer costs, and whether a person must check before use
- You have confirmed the required data exists and is in usable shape
- You have genuinely asked how this would be solved without AI
- Success and abandonment criteria are written down in advance
- Budget covers data preparation, quality evaluation and monthly usage, not only the build
- It is designed to be removed with the original process still working
Frequently asked questions
Should a small business start using AI now?
Start with one job whose time cost you can measure, rather than a large plan. Small businesses usually see the fastest return on document work and repetitive question answering, because results show quickly and it is easy to stop if it doesn't pay. Work tied to money or legal obligations is better left until the process around it is stable.
Can we use customer data with AI?
Yes, but you need to know where the data travels and have a lawful basis. Under Thailand's PDPA, sending personal data to an external service is processing that needs a justification and notice to the data subject. Safer routes are removing identifying fields first, or processing inside your own system. Decide this at design time, before launch.
How much data do we need before we can start using AI?
Most work businesses want today does not need large amounts of your own data, because ready-made models already handle language well. What you do need is data that is correct and findable: current documents, policies or product information. Quality and organisation matter far more than volume.
How do we know when an AI project isn't paying off?
Measure it against the criteria you set on day one. Without those, the clearest signals are that people still correct nearly every answer, or that total time hasn't fallen once checking is counted. Another is the team calling the system interesting rather than saying how much work it saves.
What decisions should AI not make on its own?
Anything that must be right every time, must be explainable, or has direct legal or financial consequences: tax calculations, credit approval, medical diagnosis, figures filed with a government agency. AI can prepare the work, but a named person signs. If one wrong answer in twenty is unacceptable, build in a human check.