Custom AI development cost in the Philippines: the honest number

A business owner in Metro Manila, or a foreign company looking to build here, sends an inquiry about a custom AI system. The follow-up message, almost every time: "We got an AI build price from another agency for ₱35,000. Why is your cost so different?" Here is why, and here is exactly what a real build costs and what each peso goes toward. Not to justify my prices, but because the ₱30,000 to ₱50,000 quotes flooding the market are setting clients up for failed projects.

Most of what gets sold as AI right now falls into one of three buckets.

• A ChatGPT wrapper. Someone takes an LLM API, wraps a basic UI around it, drops in your company name, and calls it your custom AI. No memory, no business logic, no real workflow integration. A competent developer can deliver it in a week.
• Zapier automations dressed up as AI. More deceptive because it actually does something. A chain fires when a form submits, sends data to an LLM, gets a response, posts it somewhere. It is if-then logic with a language model in the middle. If any single link breaks, the whole thing fails silently, and you find out three months later when a client complains.
• A freelancer who disappears after delivery. A delivered AI system needs ongoing tuning and monitoring. A solo dev who has moved on is not watching your system. Models update on schedules nobody tells you about. Your outputs drift and nobody notices because nobody is watching.

None of these are necessarily evil. For simple internal use cases, an FAQ bot or a basic email classifier, a cheap solution might be fine. The problem is when these get sold as full custom AI projects for revenue-critical workflows. That is when businesses lose money, time, and sometimes data.

Ontology design is the phase cheap builds skip entirely, and it is why those builds fail at month three. Before a line of code, we map your business objects: customers, orders, employees, documents, approvals. What are the relationships? What are the permission rules? Where does the AI need to act, and what inputs does it need at each decision point? This takes a senior architect 3 to 5 days. For a mid-size business with 4 to 6 core processes, you get a 15 to 25 page specification that every agent gets built against. Skip it and you are building on sand. Discovery, ontology, agents: three weeks minimum before a line of agent logic gets written, and that is the right order.

Agent architecture is real software work. A business AI system is not one model doing everything. It is specialized agents each responsible for a domain, passing context through a defined handoff protocol. A purchasing automation has a request-intake agent, a supplier-matching agent, an approval-routing agent, and a PO-generation agent, each with its own prompts, tools, and error handling. Designing how they communicate, in what format, with what fallback when one fails, requires someone who has done it before on a live system.

The permission layer and audit trail are not optional. Every AI action needs to be logged, not just for debugging but for RA 10173 compliance, for disputes, and for the one moment in 18 months when an agent makes a decision that affects a client or employee and that person pushes back. Without an audit trail you cannot prove what the system did, when, and on what basis. With the Data Privacy Act in play, you face personal liability if you process personal data without documented access controls. And once live, someone needs to watch the system: scheduled checks on output quality, flagging mishandled edge cases, feeding corrections back in. That retraining loop separates a system that improves over time from one that quietly degrades.

For a mid-size project of 3 to 5 agents handling 4 to 6 workflows for a company of 20 to 100 employees, you are looking at 3 to 4 weeks of senior engineering time, a project manager coordinating the build, and a delivery that includes a technical specification, a runbook for ongoing operations, and a 30-day post-launch support window. Real builds start in the ₱150,000 range and scale with scope.

An ongoing retainer for monitoring, prompt tuning, and model updates runs ₱15,000 to ₱25,000 per month depending on complexity. This is separate from the build and covers the ongoing work that keeps the system healthy after launch.

The same scope from a US agency, four weeks of senior engineering with proper architecture, audit trail, and monitoring, runs $25,000 to $45,000 USD, about ₱1.4M to ₱2.5M. We price the same scope at roughly one-fifth that, not because we cut corners or hire juniors, but because Philippine software development rates are genuinely different. That arbitrage only works if the firm is doing the same work. A ₱35,000 build is not one-fifth of a $45,000 build. It is a fundamentally different product.

Clients sometimes push back on line items to get the total down. Budgets are real. But there are three things we never move on.

Ontology design. Clients say "we know our business well, can we skip discovery?" No. Knowing your business is not the same as your business logic being documented in a form a software architect can build against. Skip formal ontology design because the client is confident, and the failure mode is predictable: six weeks in, a single permission edge case can invalidate a core architectural assumption, and the rework costs more than the discovery phase would have. That is why we do not skip it.

The audit trail. Some clients say it feels like overhead for something they will never need. You will need it, not for the happy path but for the one time in 18 months when an agent makes a consequential decision and someone disputes it. Without logs you cannot prove what happened, and with RA 10173 in force you face personal liability for processing personal data without documented access controls.

The retainer. This is where clients push back hardest. AI systems are not static products. Underlying models change in ways that shift output behavior without breaking the API. APIs your agents depend on deprecate old endpoints. Your business processes evolve. If a client truly will not accept a retainer, my advice is to hire an internal engineer who can read our documentation and maintain it. Few have that capacity. The retainer is almost always the right call.

Some clients come in asking for a one-time build, fixed fee, delivery date, done. They are thinking about it like a website: you build it, it runs forever. AI systems do not work like that. After delivery, LLM providers update their models on rolling schedules and your agents may behave differently against the new one, same prompts, different outputs, no obvious error. Third-party APIs deprecate endpoints. Your business processes evolve. Each requires someone to touch the system.

A well-maintained ₱80,000 build outperforms a neglected ₱200,000 build every time. If budget is genuinely constrained, the right move is to start smaller, fewer agents, narrower scope, and build in a maintenance budget from day one.

• No discovery or scoping phase. If the proposal jumps straight to "we'll build X, Y, and Z" without a defined discovery phase, they are building against assumptions that will be wrong in important places.
• No mention of RA 10173. Any project touching personal data, names and contact details, where the proposal says nothing about data privacy compliance, is a red flag. The vendor either does not know Philippine law or plans to ignore it.
• Working results promised in week one. Ontology design alone takes 1 to 2 weeks. Any vendor promising a live system in week one for anything beyond a toy project is delivering a wrapper or skipping the design phases that matter.
• No maintenance plan after delivery. If the proposal ends at "delivery and handover," ask what happens when the LLM provider updates their model. No clear answer means you are on your own the moment the project closes.

We quote what we quote because the work takes what it takes. If you want to understand what scope looks like for your situation, book a scoping call. We will tell you what category of build your needs fall into, the realistic cost range, and what we would cut if budget is genuinely constrained. No pitch, just an honest conversation.