How to Measure ROI on AI in a Philippine Business

Why AI Projects Are Hard to Judge After the Fact

Six months after a project finishes, someone in management asks whether it was worth it. In my experience with Philippine businesses, this question is usually unanswerable — not because the project failed, but because nothing was measured before it started.

That is a measurement problem, not an AI problem, and it is fixable in about two weeks of light effort. What follows is how I would set it up.

I should be clear about what this article does not cover. It is not about whether AI improves returns in general. It is about how a business owner decides, for a specific project, whether the money came back.

Where the Usual Way of Calculating It Falls Down

The hours-times-wage calculation is the one I see most, and it is the most misleading. Suppose an automation removes 90 minutes a day of routine replying. At a PHP 25,000 monthly salary, that looks like a clear saving. But if the staff member is still employed full-time, the company's costs did not change at all. The 90 minutes went somewhere else.

That does not mean the project was worthless. It means the benefit was not a cost reduction, and calling it one will not survive scrutiny from anyone reading the accounts.

The second approach — using the supplier's own numbers — has an obvious problem. I write proposals myself, and any projection I make rests on assumptions about how your team currently works. Those assumptions are usually reasonable and occasionally wrong.

The third is the most tempting and the least reliable. Revenue rose after the project, so the project caused it. In a real business, the season changed, the peso moved, a competitor closed, and somebody ran a promotion. Attribution needs more care than a before-and-after glance.

Related: What Are AI Agents? A Five-Minute Guide for Philippine Business Owners, With Real Project Examples explains this in detail.

The Four Things Actually Worth Counting

The first line matters most and is the one usually recorded wrongly. If a manager gets 90 minutes back and spends it visiting customers, the benefit is real, but it belongs in the revenue column, not the cost column. Write down what the person actually does with the time. If the honest answer is "general admin", the project produced comfort rather than return, which is worth knowing.

The second line is where small and medium Philippine businesses see the clearest financial case. If inquiry volume grew 40% and you handled it without adding a person, the benefit is the salary you did not have to add — and that is a real, defensible number.

The third matters in businesses where customers shop around. If your first reply used to take four hours and now takes four minutes, that is measurable and it is plausibly linked to conversion.

The fourth applies where mistakes have a price: wrong quotes, missed bookings, incorrect stock answers. Count the corrections before and after.

Notice that three of these four are not cost savings. That is deliberate. Most AI projects in smaller businesses do not reduce headcount, and pretending otherwise produces a number nobody believes.

How to Set Up the Measurement: Five Steps

Step 1 costs almost nothing and is skipped almost every time. Have the person doing the task note roughly how long it takes each day for two weeks. Rough is fine — "about an hour" is enough. Without this, everything afterwards is guesswork.

Step 2 stops arguments later. Decide in advance which kind of benefit you expect, because different benefits are judged differently. If you expect to avoid a hire, say so before you start.

Step 3 is where the cost side goes wrong. The invoice is not the cost. Add the hours your own staff spend explaining the work, testing, and correcting the output during the first weeks. For a small project this can add 20–30% to the true cost, and leaving it out flatters the result.

Step 4 must repeat step 1 exactly. Same person, same rough method. Do not upgrade to a stopwatch at this stage — a change in measurement method will produce a change in numbers that has nothing to do with the project.

Step 5 is a single sentence in a document you can find later: what it cost, what changed, how long until it paid back. That sentence is what makes the second project easier to approve.

Related: How to Scale AI in a Philippine Business: The Roadmap From One Working Pilot to Company-Wide Use explains this in detail.

What the Numbers Tend to Look Like

I want to ground this rather than talk in the abstract, so here are two of my own engagements.

For a small mobile phone shop near Little Tokyo in Makati, run by a Filipino owner, we set up automatic and semi-automatic replies to routine Messenger questions — "how much is this model?", "do you have stock?" — switching between Tagalog, English and Japanese, plus a recommendation flow and AI-produced text and thumbnails for Facebook posts. The cost was about PHP 30,000 and it took roughly two to four weeks. Inquiry handling became far easier and the shop could post consistently on social media.

For a Japanese restaurant of around 80 seats in the same area, the work was wider: rebuilding an outdated website on WordPress, targeting small keywords with current SEO and GEO practice, using AI in Facebook-centred social marketing, adding a chatbot for simple inquiries, and making the menu editable from a normal PC. Consultation took about a week, the build one to two months, and full operation began two weeks after staff training — around two to three months in total. The fee is not public. The small keywords still hold positions one to five, staff almost never handle simple inquiries now, and both visibility and customer numbers rose.

Those two projects need different measurement. The first is a volume-and-time case. The second is a visibility case, and judging it on hours saved would miss the point entirely.

Related: How Philippine Businesses Can Find Their First AI Use Case When They Have No Idea Where to Start explains this in detail.

FAQ

Q: We did not measure anything before the project. Can we still work out the ROI?

A: Not reliably, but you can start now. Measure the task as it currently runs, then treat that as your baseline for the next change. It is also worth asking the person who does the work what it was like before — an honest recollection is weak evidence, but it is better than a number you invented afterwards.

Q: Should we count time saved if nobody was let go?

A: Count it, but not as a cost saving. Record where the time went. If it went into work that brings in revenue, that is the benefit and you should describe it that way. If it went into general admin, be honest that the project bought capacity rather than money.

Q: How long should we wait before judging a project?

A: One month for a single automated task, two to three months for anything involving search visibility or customer behaviour. Judging too early is the more common mistake — search position in particular does not move in a fortnight.

Q: The supplier promised specific savings. Should we hold them to it?

A: Ask what assumptions the number rests on, and check whether those assumptions match how your team actually works. A supplier's projection is a starting point for a conversation, not a commitment you should build a budget on. I write these proposals myself, and I would rather a client questioned mine than budgeted around it.

Q: What if the honest answer is that it did not pay back?

A: Write that down too. A recorded failure is what stops the same mistake being repeated in the next project, and it is far more useful than a vague sense that AI "did not really work for us". Usually the problem turns out to be the task that was chosen, not the technology.

Start Measuring Two Weeks Before You Start Building

The reason most Philippine businesses cannot say whether their AI project paid back is that the measurement was supposed to begin before the project did, and it did not.

If you have a project planned, the useful thing to do this week is not to compare suppliers. It is to pick the task, ask whoever does it to note roughly how long it takes each day, and decide which of the four benefit types you expect. That is perhaps twenty minutes of work, and it is what makes the answer available later.

At PH AI Works we support Philippine businesses from a first PHP 30,000 automation through to full custom systems, and we are happy to help set up the baseline measurement before any build begins. If you would like a second opinion on what to count for your project, get in touch.

References

All project costs and timelines in this article (the Makati mobile phone shop and the 80-seat Japanese restaurant) are from my own client engagements. No external statistics are cited; where a figure is not disclosed, it is withheld under the client agreement rather than estimated.