Shipping Small Internal Tools in Two Weeks

Most of the internal tools a Philippine business needs are small. A form that routes a request to the right person. A screen that shows which orders are stuck. A page that turns a pile of customer messages into a summary someone can act on before lunch.

None of these justify a six-month development project, and that is exactly why they never get built. They sit on a list, the list gets longer, and the work keeps happening in spreadsheets and group chats.

Bubble paired with the OpenAI API is one of the few combinations that actually closes this gap. Not because it replaces engineering, but because it moves a specific class of small tool from "we should build that someday" to "we shipped it on Thursday."

The Backlog Problem Philippine Businesses Actually Have

The first two are familiar in any market. The third is the one that changed recently.

Until fairly recently, "read this text and decide what it is about" required either a person or a machine learning project with a labelled dataset. That was well outside the budget for a tool that saves one supervisor four hours a week. Now it is an API call, which means the economics of small internal tools have shifted underneath a lot of teams who have not noticed yet.

I see this most often in operations and admin roles, where the work is unglamorous, high volume, and full of judgement calls that are easy to describe and tedious to perform.

Related: How AI Reality Checks Help Philippine SMEs Avoid Costly Bubble Mistakes explains this in detail.

Why the Three Usual Approaches Stall

Custom development is not the enemy here. For anything touching money, inventory or regulated data, I still recommend it. The mismatch is one of scale. A properly scoped development engagement has a discovery phase, a specification, a build and a handover, and all of that is correct and necessary — and completely disproportionate for a form that routes requests.

Off-the-shelf software fails in a subtler way. The product covers eighty percent of what you need, so you configure around the missing twenty, and the configuration becomes its own undocumented system. Two years later nobody remembers why a field is named the way it is.

Spreadsheets are where most of this work actually lives, and they work until the moment someone leaves. Then you discover that the process existed only as a habit.

Related: How OpenAI and Anthropic APIs Help Philippine Businesses Build Custom AI Agents explains this in detail.

What Bubble Plus OpenAI Changes

Bubble is a visual application platform. You lay out pages, define a database, set up user roles and build workflows without writing application code. Its API Connector lets you call an external service, which is how the OpenAI API enters the picture.

The pairing works because the two halves cover each other's weaknesses. Bubble gives you the boring but essential parts — logins, permissions, a real database, an audit trail — that a spreadsheet cannot. The language model gives you the judgement layer that would otherwise require a person reading every record.

A concrete shape I keep coming back to: a form where staff or customers submit a request in free text, a workflow that sends that text to the API and asks for a category and a one-line summary, a database record storing both the original text and the result, and a queue screen where the assigned person confirms or corrects the category before acting. The correction is the important part, and I will come back to it.

A Six-Step Build You Can Run in Two Weeks

Step 2 is where projects succeed or fail, and it has nothing to do with technology. If you cannot write down how a person currently decides which category a request belongs to, the model cannot do it either. I have watched teams spend a week on the interface and ten minutes on the rules, then conclude that the AI does not work.

Step 5 is the one people want to skip. Do not skip it. The confirmation action costs the user two seconds and gives you a record of every case where the model was wrong, which is the only data that will let you improve the prompt. It also keeps responsibility where it belongs, with the person acting on the result.

Step 6 is your safety net. Running both methods in parallel for two weeks feels wasteful and is not. It gives you a real comparison rather than an impression, and it means you can stop at any point without disruption.

Related: How Advanced AI Automation Helps Philippine SMEs Go Beyond No-Code Limitations explains this in detail.

How to Work Out Whether It Pays

I am deliberately not quoting figures for the first two costs, because both platforms revise their pricing and plan structures, and a number written today would mislead someone reading in six months. Check the live pages and put the date next to whatever you write down.

The one number you must capture yourself is the baseline. Suppose a supervisor spends six hours a week on request triage. Over a month that is roughly 24 hours. If your fully loaded cost for that role is, say, ₱250 per hour — use your own figure, not mine — that is around ₱6,000 a month of capacity. Whether the tool pays depends on how much of those 24 hours it actually removes, which is not the same as how much it could theoretically remove.

This is the same discipline I would apply to any automation. Record the before-figure on the day you decide to build, in writing, with the date. Teams that skip this cannot answer the only question head office will ask, which is whether it worked.

There is a second return that is harder to price. Once a process exists as an application with a database and permissions rather than as a spreadsheet and a habit, it survives staff turnover. In an operations team that hires and promotes regularly, that continuity is often worth more than the hours saved.

FAQ

Q: Do I need a developer to build this?

A: For a first internal tool, no — Bubble is designed to be usable by someone comfortable with logic and data structures rather than code. What you do need is someone who can describe the process precisely, and that person is usually in operations, not IT. I would bring in a developer at two points: when the tool starts holding data you would be uncomfortable losing, and when you need to connect it to an existing system such as your accounting or inventory software.

Q: Is it safe to send customer data to an external API?

A: Treat it as a personal data question from the start, because it is one. Under the Philippine Data Privacy Act of 2012, processing personal information requires a declared, legitimate purpose and a lawful basis, and the implementing rules require notification of a reportable breach within seventy-two hours of knowledge. Before you send anything, decide what fields actually need to go out — often names and contact details can be stripped and only the message body sent — and confirm your approach with counsel or your data protection officer.

Q: What happens when the model gets something wrong?

A: It will, which is why the confirmation step is not optional. The correct design assumes a percentage of wrong answers and puts a person between the model and any consequential action. Track your correction rate from the first day. If it is high, the fix is almost always a clearer description of the decision rules, not a different model.

Q: When should we move off Bubble to a custom build?

A: When one of three things happens: transaction volume grows past what the platform handles comfortably, the tool starts touching money or regulated data, or you need integrations that the platform cannot support cleanly. None of these are failures. A working Bubble application is the best possible specification document for the custom build that replaces it, because it shows exactly what people actually use.

Q: Can we build this in English and Tagalog?

A: Yes, and the language model handles mixed-language input better than rule-based systems ever did, which matters in a market where a single message often contains both. Test with real messages from your own inbox rather than clean examples. The failure cases you care about are the ones your customers actually send.

Start With One Process, Not a Platform Strategy

The temptation with no-code tools is to plan a portfolio. Resist it. Pick the single most annoying repetitive task in your operation, write down how long it currently takes, and build one tool that does part of it with a human confirming every result.

Two weeks later you will know three things you cannot know now: whether the decision rules were as clear as you thought, what your real correction rate is, and whether the hours actually came back. Those three answers are worth more than any roadmap you could write today.

If you want a second opinion on which process to start with, or on where the human confirmation step belongs in your workflow, PH AI Works runs a 30-minute consultation for Japanese-affiliated companies operating in the Philippines. Bring the process, not a technology preference.

References

• Bubble — Pricing
• Bubble — API Connector documentation
• OpenAI — API pricing
• National Privacy Commission — Data Privacy Act of 2012
• National Privacy Commission — Implementing Rules and Regulations of the Data Privacy Act of 2012