Before you pick what AI should do,
find your business's bottleneck first

Picture a building-materials business. A contractor messages on LINE asking for a quote. The assistant already has the file open and every line item filled in. One cell is left, the price, and only the owner can fill it. The owner is on a job site all day and gets back at 8 pm. The quote goes out the next morning, and the contractor replies that the order has already gone to someone else.

The assistant finished their part by the afternoon. The work stopped at the owner's desk. If you own a business like this, think about what makes your phone buzz most. It might be a staff member calling to ask a price, a customer asking for a quote, or a delivery nobody has signed for. All of these can be the same thing: work waiting for you.

We made this business up, and every other business and number in this post too. They exist to teach, and none comes from a real business. Problems shaped like this can happen in real ones, though.

This post condenses a workshop, SME Business Design 101: What Should People Do, What Can AI Do, and Should You Build or Buy? (the Thai title is ออกแบบธุรกิจ SME 101), at AI Expo Thailand 2026, Business Stage, Challenger Hall Muang Thong Thani, Friday 6 November 2026, 11:00 to 12:05. We wrote it from the slides, worksheet and speaker script prepared for that workshop. It is not a record of what happened in the room.

The first question that comes to mind when you think about AI is which of our tasks AI can take over. At Productize, we think the first question is a different one: where in the business does work pile up? A pile shows where the business is stuck. That spot is the bottleneck of an SME, and it is where effort should go first, whether the effort is a person, AI, or a system.

Part 1The thing stuck in your head

The worksheet opens with questions you can answer in half a minute. The first is the 7-day test: if you did not come to work for 7 full days starting tomorrow, what would happen to the business? Pick one: it runs as usual, some things wait for you, or everything waits for you. The second is the one thing that waits for you most, or the one thing stuck in your head right now.

The worksheet has one rule. Use one real case of yours from start to finish. Do not pick the case that looks best. Pick the one that makes your phone buzz most. Where you do not know a number yet, write "not known". Do not guess.

The first question gives the fastest hint of where the bottleneck is. If you answer that everything waits for you, the bottleneck of the business may be you. We come back to that in the next part.

Part 2Where work piles up, and what a bottleneck is

This part of the worksheet asks which pile of yours grows every day, why it grows most, and whether you know the weekly numbers for work coming in, going out and coming back as rework. Before answering, you need to know what a bottleneck is. Take a made-up food business with three steps working in a row.

The whole business sells 40 dishes an hour, and every order queues in front of the kitchen. If the kitchen sits idle for one hour, the business loses the money left after ingredients on 40 dishes, and it cannot make that up anywhere else. If the order takers sit idle for an hour, nothing is lost, because only 40 can move on anyway. An hour lost at the bottleneck is an hour lost for the whole business. An hour saved anywhere else is an illusion.

The kitchen in this business is what the book The Goal, by Eliyahu M. Goldratt and Jeff Cox, calls a bottleneck. The book tells the story of running a factory as a novel. Its picture is a bottle: a bottle can be as big as you like, but the water still leaves at the speed of the neck. Work piles up in front of that step, while the step after it sits idle and waits.

What should this business do first if it wants to sell more? A common first answer is to expand the kitchen. That is half right, since it has to be the kitchen. But before spending, look at what eats the kitchen's time. Does the cook have to chop vegetables between orders? Could that be done ahead? When the cook takes a meal break, does the whole kitchen stop? Could someone else take over? Hiring another order taker would only add a wage every month. The kitchen still sends out 40 dishes, and the money left after variable cost does not move.

Now go back to the building-materials business. The quote that waited until 8 pm: the owner is that business's kitchen. If most work has to pass through your desk, the bottleneck of your business may be you. Not because you are slow, but because there is one of you. When you spend half a day in a meeting, the whole business waits half a day, like a kitchen that has lost power.

The pile: a bathtub

A clothing business sells through LINE. In a corner there is a pile of orders waiting to be packed. It is taller this evening than it was this morning, and tomorrow it will be taller again. Ask the team who is busiest and everyone says they are about equally busy. The pile answers more honestly. A pile that grows every day is the place that truly cannot keep up. Businesses that sell no goods have piles too: production orders not yet finished in a factory, repair jobs taken in but not yet returned, customers standing in line for a table.

A helpful picture of a pile is a bathtub. It comes from Thinking in Systems by Donella Meadows. Meadows calls whatever has built up a stock, which here we call a pile: something you can count right now. What changes the pile is two things, work coming in and work going out. In a tub that is the tap and the drain. In a business it is new orders and orders packed and shipped.

The tub has one rule. If more comes in than goes out, the pile rises. If more goes out, it falls. If they are equal, it stays level, even though work keeps passing through. Meadows goes on to say that a pile changes only through its inflow and outflow, so to lower it you cut what comes in or raise what goes out, and you cannot press the pile down directly. A pile also changes slowly: the flows can change in one day, but a pile built up over weeks takes time to drain. So a tall pile today often reflects weeks in which work came in faster than it could be done.

Pushing everyone to clear the pile may waste your time. The pile is what you see, not the cause. If work still comes in and goes out the same way, the pile is back next week. The fix is at the inflow and the outflow.

Name the pile, see why it grows, then count

On the worksheet, you name one pile, then pick why it grows most. There are 4 answers: more comes in than can be done, it waits for someone to decide, it comes back for rework, or not sure. Then you say whether you know the weekly numbers in, out and back as rework, with three options: know, roughly know, or not known yet, count it in the homework.

The small factory shows that the 4 rework jobs matter more than they look. Cutting the jobs that need a second question may do more than pushing people to work faster. The next example runs the same arithmetic in more detail.

A made-up example: a pile growing by 50 orders a week

The clothing business that sells through LINE, and the numbers in this section, are all made up to teach, and are not data from a real business. It sells through chat, with no order form. Customers message to order. Some do not give a phone number, and some do not give a postal code.

250 orders come in through LINE a week. The admin who takes orders can open an order to handle it about 240 times a week (this counts the times an order is opened, not the number of orders). Of all the orders, 40 need another message to the customer, so the admin handles these orders twice. The first time, the admin opens the order and finds details missing. The second time is when the customer replies. Work that has to be redone like this is rework. Those 40 openings go to orders already seen once, so the orders really finished are 240 minus 40, which is 200 a week. 250 come in and 200 go out, so the pile grows by 50 a week, exactly as the tub rule says.

When you see 50, hiring is the first thing that comes to mind. Hold on. This admin is not slow. If every order arrived complete the first time, those 240 openings would all go to new orders. That is almost exactly the 250 that come in. The admin's time is eaten by missing details from the start, and those 40 openings are time lost at the bottleneck, like an hour the kitchen sits idle. It cannot be bought back.

Now think about this: if rework orders were cut in half, from 40 to 20, what happens to the pile? The arithmetic is direct.

If this made-up business cut rework orders in half, its pile would grow by about 30 a week instead of 50, with nobody hired and no AI. The 10 left in the last row is the gap between work in and work out, 250 minus 240. Closing that remaining gap need not mean a full-time hire, and the real numbers should be checked before deciding.

Money piles up in the pile of work too

Go back to the food business whose kitchen is the bottleneck. Orders pile up in front of the kitchen. Ingredients chopped ahead pile up there too. These things are bought but not yet sold, so the money paid has not come back. A pile of work is also a pile of money. Goldratt tells you to watch three money numbers at once.

1. Money left after variable cost enters the business only as fast as the kitchen sends dishes out.
2. Money stuck in the pile things already bought, work half done and waiting. The money has gone out but has not come back as cash. It takes up space, and some of it can still spoil or go out of date.
3. Monthly running costs such as wages, rent and electricity.

If money left after variable cost grows faster than running costs, profit rises. If money stuck in the pile goes down, cash gets easier. This is why hiring another order taker, or buying a machine that is not the bottleneck, gets you only higher running costs.

But the pile in front of the bottleneck does not have to be zero. A few order tickets waiting in front of the kitchen are a buffer. They keep the kitchen busy so it never has to stop and wait. A buffer is a small pile that stays level and does not grow every day. Keep that. The money wasted is the part above the buffer, plus piles sitting at other steps.

And if you are the one who releases work in, such as ordering stock or making things ahead, Goldratt says to release no faster than the bottleneck can take it. Release faster, and not one more dish gets sold. The only result is money stuck in the pile.

Part 3Which fix first

Once you know where the pile is and why it grows, the worksheet asks you to pick one fix to try first on the pile you named, and to write one line of reasoning: the symptom you see. The method is a ladder that runs from cheap and easy to undo, up to expensive and hard to undo. Pick the fix by the cause, not by the tool you feel like using, and climb from the low numbers first. Steps 3 to 6 are ones where a free chat AI can help you. Letting AI do the work, with a person checking, is step 9.

1. Step 1 Leave it for now Park what is not the bottleneck. For example, if the kitchen cannot keep up, you do not need to hire another order taker yet. Gain: costs nothing and saves effort for the bottleneck. Watch for: if you misjudged, the problem grows quietly, so set a date to look again.
2. Step 2 Cut it First ask what problem the task used to prevent, then stop doing work nobody uses. For example, a third copy of the delivery note that nobody opens. Gain: time back at once, nothing to buy. Watch for: if the task did prevent a problem, it returns, but this is easy to undo.
3. Step 3 Make the steps simpler Cut hand-offs and approval layers. For example, let the order taker send orders straight to the kitchen. Gain: work flows faster and less gets dropped. Watch for: easy to undo, but talk to the team before changing. AI can help: write out the current steps and find the repeated hand-offs
4. Step 4 Set rules and standards Agree that everyone does it the same way. For example, write down how much discount staff can give on their own. Gain: anyone gets the same result, and it is ready to grow into a system. Watch for: it takes time to write, and someone must check that people follow it. AI can help: draft the SOP or the approval rule
5. Step 5 Fix the data at the source Fix it where the data is born. For example, one product code used by both the sales screen and the stock list. Gain: no more fixing the same wrong numbers every month. Watch for: the people entering data must change habits, and it is slower at first. AI can help: find the rows in the stock sheet that do not match
6. Step 6 Get the numbers to the decider Let the person who decides see the numbers in time. For example, yesterday's sales reach the owner every morning. Gain: faster, better decisions at low cost. Watch for: if the source data is still wrong, you only send wrong numbers faster. AI can help: explain numbers that have already been sent
7. Step 7 If it does not sell, change the terms of sale Find out why customers do not buy, then change the terms. For example, lower the minimum order. Gain: sell more without adding capacity. Watch for: use it only when you can make it in time but cannot sell it. A lowered price is hard to raise again.
8. Step 8 Automation by rules Let a program follow a rule you set. For example, when stock falls below 10 pieces, send a LINE reminder. Gain: repeats without tiring or forgetting, with the same result every time. Watch for: the work must be standard first, or it repeats mistakes faster. It has upkeep cost and is hard to undo.
9. Step 9 AI does the work, with a person checking Let AI do the work. For example, read a bank transfer slip and enter the amount, with a person spot-checking every day. Gain: it can read text or pictures, which rules alone cannot. Watch for: the work must be standard first, someone must check all the time, it has ongoing cost and is hard to undo.
10. Step 10 A person decides, with a check before approval Work that needs judgment, such as approving credit. A person decides, and AI only prepares the data. Gain: risk is under control and someone clearly owns the answer. Watch for: it is slower and takes people's time, and the checker's name must be written down.

If you are not sure which step to start on, the worksheet lets you pick by the symptom you see.

• Done but nobody uses the result: go to step 2, cut it
• Work passes through many hands: go to step 3, make the steps simpler
• Everything waits for the owner: go to step 4, set rules so the team decides
• Numbers in the system do not match reality: go to step 5, fix the data at the source
• The decider cannot see the numbers: go to step 6, get the numbers to the decider
• Documents arrive in different shapes: go to step 9, let AI read them with a person checking

Hire, fix, buy or build

On hiring: hire after you have fixed the cause, when the gap that remains is still growing and the new person adds capacity at the real bottleneck. For example, a second admin for the clothing business on the day every order arrives complete but new orders still cannot be handled in time. Hire someone for a different step, and you only get a higher wage bill.

If fixing what you have is not enough, then think about buying or building a system. Ask 2 questions: does this work make your business different from others, and can someone on your team maintain the software? If every business does the work the same way, such as payroll and tax, buy. If the work makes you different and someone can maintain it, build. The other case is work that makes you different but that nobody can maintain in code. Do not build that. Buy software that lets you set your own rules. A building-materials business with its own pricing formula can put that formula into bought software. The difference stays, and nobody has to maintain the software.

Before deciding, compare 3 years of cost for every route. Purchase or build cost and upkeep count as money. The team's time counts as hours, kept separate and not converted to money. Then ask whether the hours you get back will go to the bottleneck. If they do not, that is not a reason to buy. The worksheet has cells for the 3-year cost of fixing what you have, buying ready-made and building new. Do it as homework H5. The full way to compare is bundled into an assessment on the Fix, build or buy page.

Part 4Can AI help?

AI comes later: you need to see the bottleneck first, not start from AI. So the worksheet asks 4 questions before using AI about the case you wrote in Part 1. All 4 must pass. Where one does not, do that first.

1. Are the steps clear? Does everyone do it the same way, or are there examples of correct work that AI can compare against? If not yet, set rules.
2. Can you trust the data? Is it complete and correct, and are you allowed to use the customer data? If not yet, fix the data or check your permission.
3. How bad is it if it is wrong? If AI gets it wrong, can you live with the damage? You answer: can live with it, cannot, or not sure. If you cannot, let a person decide, and AI only prepares the data.
4. Who checks, and who takes over when it is down? Who checks the work, and if the system goes down, who picks the work up? Put it in cell 3 of the one-page plan. You answer know, do not know, or not sure. If not yet, name the people.

Question 2 in this example hides a trap. Suppose you have AI draft a message asking for the phone number and postal code on every order. It does help, and it is quick. But once asking gets easier, nobody sets the rule about what to ask for when taking an order. Customers still give incomplete details, and AI has to draft the question every day, forever. The burden has only moved to AI. The real fix is the single rule from Part 3. While the rule is not steady yet, you can have AI ask in the meantime, but set an end date for it, or the shortcut becomes the permanent way of working without anyone deciding it should be.

Try deciding 3 cases yourself

Which cases should use AI, and which should not? The worksheet has 3 cases to try first. Answer in your head, then check the answer.

1. The owner has to approve every discount, even small ones that happen daily. Answer: rules. Set a limit the team can approve on its own, then let a system approve the normal cases. AI adds nothing here, because the answer is the same every time and a rule already covers it.
2. Orders come in on LINE, and customers type in different ways. Staff have to read them and key them into the system. Answer: AI. Let AI read the messages that are typed differently and fill them into one order form, with a person checking the unclear items and checking before saving. One customer types black shirt M 2 pieces, another sends a screenshot, another types the same as yesterday. A fixed rule cannot read those.
3. Stock in the system does not match the real stock, so staff confirm orders wrongly. Answer: data. Fix the steps for receiving in, taking out and counting first. Adding AI would only give you the wrong answer faster.

Who decides, and can they see the numbers?

This part closes with 3 more questions. What decision repeats every day? Who makes it? And can that person see the numbers needed? The options are: see it right away, see it 1 to 7 days late, see it more than 7 days late, or cannot see it at all.

If the decider cannot see the numbers, get a report from your system to them first, for example the amount customers still owe, with the report date on it. It can be set to send on a schedule, with no AI. You can have AI help explain it and point at what to check, but compare the figures with the source report before deciding. For work that AI does on its own inside the business, and for deciding when it can act alone and when it must wait for a person, we wrote a separate post, Not every action needs to wait for a human.

Part 5Check yourself and write a one-page plan

In this part the worksheet has a 6-question Quick Check. Pick one answer per question, going by how you feel today. Questions 1, 2, 4 and 6 are the core. Questions 3 and 5 can be skipped and done later.

• Question 1 What is wearing you out most right now? For example, repeated work that takes a lot of time and cannot be finished, or everything waits for the owner or a key person.
• Question 2 If one thing could get better in the next 6 months, what do you want to see? For example, work done faster with the same people, or the team deciding without waiting for the owner.
• Question 3 What level is your way of working at now? From relies on the owner or on people's memory, up to measured and improved all the time.
• Question 4 What is the biggest limit you feel? For example, data that is incomplete, wrong or scattered, or everything waits for a few people to decide.
• Question 5 How much do you want to change? From fix a few urgent things first, up to change how the business works in a big way.
• Question 6 Which process holds the main problem? For example, taking orders through to getting paid, or stock, buying and sourcing.

Your answers join into one sentence that the worksheet calls your hypothesis today, and it tells you what to do first. For example, if the limit is that everything waits for a few people to decide, set a rule for what the team can decide on its own, such as discounts up to a limit, without waiting for you.

This result is still only how you feel today, not a verdict. The homework tests it against real numbers. And one warning: if in the next 30 days you might not be able to pay wages or rent, or might have to stop your main work, talk to an expert on that first. Do not start with AI.

The one-page plan

The worksheet ends with a one-page plan of 5 cells that you take home. Cells 1 and 2 pull your answers from Parts 1 and 3. Cell 3 is names of people. In cell 4 you choose fix, buy or build (not known is fine, do it in the homework). Cell 5 is the first thing you will do, one thing you can start right away, with no extra budget.

Part 6One week of homework

Real numbers sometimes point somewhere other than where the problem feels like it is. So the homework takes the case from Part 1 and collects numbers, then checks whether the fix you picked is still right. There are 6 items. They are for you, and the answers stay on your phone. Where you cannot count, write not known.

• H1 Count your pile for a full week Write down each day how many came in, went out, came back as rework, and are left at day end. You can count from customer LINE chats, a sales notebook or a POS report. Remove customer names and phone numbers before you give it to AI. A one-day example (made up): left from yesterday 16, in 36, out 30, so the end of day is 16 plus 36 minus 30, which is 22. Also note how long the longest-waiting item waited.
• H2 Three decisions that happen often Write what is decided, who decides, whether they see the numbers, and how many days late. See which row has a decider who cannot see the numbers.
• H3 Things you decide yourself that the team could decide Note every time it happens during the week, then see which kind happens most. That one is the first candidate for a rule you set for the team.
• H4 Something you fixed that came back Write what it was, how you fixed it, and why you think it came back. An example (made up): the clothing business fixed it by reminding the admin to ask for complete details, but it came back, because there was no single rule everyone follows the same way.
• H5 Fix, buy or build, compared over 3 years Do this when you picked step 8 or 9. Fix what you have first, and compare buying and building only if that is not enough. Then ask the seller 4 questions before choosing.
• H6 Plan a small 30-day experiment Write what you will try, one number that shows whether it starts to improve, one guard number (check that quality does not drop), and the date you expect to see a real result. An example (made up): a food business, where the number to watch is the time from the customer ordering to the food reaching the table, and the guard number is how many dishes were served wrongly. Some things show results within a week and some take months.

If you want to follow the full hour slide by slide, the workshop's public deck is at productize.life/workshop/ai-expo-sme-101/slides. For choosing between fix, build and buy, the ready-to-use form is on the Fix, build or buy page.

If the first question of Part 1 shows that some things really do wait for you, start with one rule that lets part of the work skip your desk, before you count what to hire or buy. The building-materials business and the clothing business in this post both started that way.

Related: Not every action needs to wait for a human on what AI can do on its own and what must wait for a person / Fix, build or buy: a verdict tool with 3-year numbers