Most owners I meet are asking the same thing right now: where should we use AI first?

I spent nine years on the other side of that question. I started Upstack Studio in 2017 and ran it until 2026. Companies came to me with software they wanted built. I wrote the proposals. I delivered the projects.

Almost nobody brought a number. Nobody could tell me what the problem was costing them that month. So nobody could tell me whether the thing they wanted was worth building.

That's the trouble with "where should we use AI?" It sounds like a technology question. It isn't. It's a money question, and you can answer it with figures you already have.

I closed Upstack Studio in 2026. The agency made money. It just wasn't the model I wanted to run anymore, because the overhead was fixed and I couldn't grow it without adding more people. I held on longer than I should have when our sales pipeline dried up, and that cost me about RM600,000 of my own money.

So I'm not neutral here. I've been the vendor writing the proposal. I've also been the owner who did the maths on his own business too late.

Here's the method I use now. You can run it yourself in about two hours.

  1. Ask where the business loses money
  2. Pull the real numbers
  3. Sort them: facts, fair guesses, unknowns
  4. Write down all five options
  5. Score every option, money against proof
  6. Choose, and write down why

    Nothing clears the bar?Stop. Say so out loud.

  7. Run one 90-day test with a stop rule
  8. Ask six questions before you approve spend
  1. Ask where the business loses money
  2. Pull the real numbers
  3. Sort them: facts, fair guesses, unknowns
  4. Write down all five options
  5. Score every option, money against proof
  6. Choose, and write down why

    Nothing clears the bar? Stop. Say so out loud.

  7. Run one 90-day test with a stop rule
  8. Ask six questions before you approve spend
Two hours of work, and a real exit. Most of the method happens before anyone writes a line of code.

Step 1: Stop asking "how can we use AI?"

Every business can use AI somewhere, so that question always produces a list. Twenty ideas, all possible, nothing to tell you which one to fund.

You can't act on a list like that. You have one budget and one team's attention this quarter. You need to know which single item deserves them.

Ask this instead:

Where is the business losing money, repeating work, or deciding without enough information?

Answer that first. Then ask whether AI changes anything about it.

Some of your best answers won't involve AI at all. That's the method working, not failing.

Step 2: Find the leak, and pull the actual numbers

Look in three places. Give yourself a week and pull real figures for each one. The rest of the method is worthless without them.

Money leaking after you already paid to get it. You spent the marketing budget. The lead arrived. Then something went wrong. Pull these:

  • Quotes you sent last quarter, and how many closed
  • Days between an enquiry arriving and your first human reply
  • Customers who ordered in the previous 12 months but not in the last 6
  • Your top 20 accounts by revenue, and when each one last bought

If your leak starts after the lead arrives, here's how to pull the four follow-up numbers and read what they say.

Work that repeats and needs no judgment. Pull these:

  • The same data typed into two systems. Count records a week, times minutes each
  • Reports rebuilt by hand every week. Count hours per month
  • Questions your team answers by email again and again. Count two weeks' worth
  • Multiply those hours by a loaded hourly cost. Loaded means salary plus tax, benefits and the desk they sit at, not salary divided by 2,080

Decisions you make blind. These are hard to price, so write the question instead of a number:

  • Which products or services actually make money after delivery cost?
  • Which salesperson is doing badly, and how would you know?
  • What does it cost you when stock is too high, or too low?

Pick the leak that's costing the most. Not the one that's most interesting to fix.

Step 3: Sort what you found into three piles

Do this before anyone proposes a fix. The figures below are invented to show the shape of it. They're not from a real company.

Pile What goes in it Example
Facts Things you can pull from a system today 412 quotes issued last quarter, 96 closed
Fair guesses What the facts point to, but don't prove Slow follow-up is part of why the rest went quiet
Unknowns Things you'd have to measure to know How many of the 316 were still open when we stopped replying

Most bad AI projects start when a guess gets treated as a fact. Someone says follow-up is the problem. Everyone nods. Nobody checks. Six months later you own a tool nobody uses.

If your biggest unknown is cheap to measure, measure it before you buy anything. A week of counting beats a quarter of building.

Step 4: Write down all five options

For the leak you picked, you have five honest options. AI is the right answer in some of them and the wrong answer in others. Two of them need no software at all.

  1. Fix the process. Change who does what, when, and in what order. No software.
  2. Buy something that exists. Someone has probably solved this. Pay them.
  3. Automate the part that repeats. Rules, links between your systems, and AI for steps that repeat and follow rules.
  4. Build something custom. Only when the job is unique to you and worth paying to maintain for years.
  5. Do nothing. The leak is real but small, and your attention is worth more elsewhere.

Write all five down every time, including the ones you don't like. A list with one item on it isn't a decision. If you're stuck between options 2 and 4, how I decide between building and buying software covers that choice in detail.

One warning on options 3 and 4. AI is good at work that repeats and follows rules. It can't replace the person who holds the customer relationship. It can't make the judgment call about what the business does next. Give the machine the repeats and keep your people on the judgment.

Step 5: Score all five options

Score each option from 0 to 5 on the five tests below, then add up the row. Zero means it changes nothing at all.

Test The question it answers
Money If this works, what changes in ringgit or dollars?
Proof How much of that number is fact rather than hope?
Speed Weeks or quarters before anything is different?
Fit Does this matter to where the business is going?
Ease Can your actual team run this once it exists?

Here's a worked example. The company's leak is quotes nobody chased. The scores are made up to show how the grid behaves.

Option Money Proof Speed Fit Ease Total
Fix the process 4 4 5 3 4 20
Buy a CRM add-on 4 3 4 3 3 17
Automate follow-up with AI 5 3 3 4 3 18
Build custom 5 2 1 3 1 12
Do nothing 0 5 5 1 5 16

The grid rarely picks the winner by itself. Its real job is to show you where you disagree. When two people score Money 5 and 2, you've found the argument worth having, and it's almost always about proof rather than about AI.

Watch any option that scores high on Money and low on Proof. That pair is how companies spend real money on a guess.

Real money on a guess

Money

0 5 0 5

Proof — how much of the number is fact, not hope

  • Fix the process: Money 4, Proof 4
  • Buy a CRM add-on: Money 4, Proof 3
  • Automate with AI: Money 5, Proof 3
  • Build custom: Money 5, Proof 2; in the shaded corner
  • Do nothing: Money 0, Proof 5

Money, 0 to 5

Real money on a guess

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5

Proof, 0 to 5 — how much of the number is fact, not hope

  1. 1Fix the process Money 4 · Proof 4
  2. 2Buy a CRM add-on Money 4 · Proof 3
  3. 3Automate with AI Money 5 · Proof 3
  4. 4Build custom Money 5 · Proof 2
  5. 5Do nothing Money 0 · Proof 5
The worked example from step 5, plotted. The shaded corner is the pair to watch: a big number nobody has checked.

Step 6: Choose, then write down why

  • The highest total with real proof behind it wins.
  • If the top two are close, take the faster one. You'll learn more from finishing.
  • If nothing clears the bar, do nothing and say so out loud. "Not yet" is a real answer.

Then write one paragraph. What you chose. What you turned down. What would have to change to bring a rejected option back. In six months, when someone asks why you're not doing the other thing, that paragraph is your answer.

Step 7: Turn the winner into a small, timed test

Give it edges. Fill in these five lines before anyone starts work:

  • The job: the leak you picked.
  • The owner: a named person, not a committee.
  • The number: what you expect to move, measured before you start.
  • The window: 90 days. Long enough to be real, short enough to stop.
  • The stop rule: what result makes you kill it, decided now while you're still calm.

Measure the starting number first. If you can't measure it, you can't judge the result. Go back to step 2 and pick something you can count.

Step 8: Six questions before you approve any spend

Make whoever is proposing the work answer these. Internal team or outside vendor, same list.

  1. Which business problem does this address, and what is it costing us now?
  2. Which parts of that number are measured, and which are guessed?
  3. What are the other four options, and why did this one win?
  4. What has to be true for this to work?
  5. How will we know in 90 days whether it worked?
  6. What happens to this if you disappear?

I wrote proposals for nine years. A proposal that can't answer those six questions isn't a plan for your business. It's a plan for someone else's revenue, and I say that as someone who used to sit on the other end of it.

Start here this week

  1. Pick one of the three leaks in step 2 and pull the numbers.
  2. Sort what you find into facts, fair guesses and unknowns.
  3. Score the five options and pick one.
  4. Write the five lines of the 90-day test.

If you'd rather have someone from outside run the diagnosis with you, that's what the AI Profit Opportunity Audit is. You get your opportunities ranked, the top three recommended, and a 90-day plan. Diagnosis and priorities only, no build work. Here's an example audit for missed B2B follow-ups.

Either way, start in the same place. Find the leak, put a number on it, and make AI earn its place against the options that don't need it.