Data labelling, RLHF rating, model evaluation and content moderation can all be done from the Philippines. For most buyers in the US, Australia, the UK and the EU, a BPO contract for a defined queue fits best. An employer of record (EOR) fits a team you want to direct yourself. Before you pick a model, read the outsourcing to the Philippines hub, which compares the three setups side by side.
It does not walk through incorporation, which the software company registration page covers.
Which tasks can be done from the Philippines?
Remote annotation of text, images, audio and video is the core work. So are pairwise preference ranking for RLHF, rubric-based evaluation of model outputs, and review of user content for policy violations. Work that needs a licensed professional’s judgment is different. Where your country or the Philippines requires a licence for that judgment, a licensed person has to do it. Labelling and moderation have no general licence rule of their own, so check the rules for your specific data with counsel before you move the work offshore.
Which setup fits: contractor, BPO, EOR or your own entity?
| Model | Who employs the annotators | Start-up speed | Fits when |
|---|---|---|---|
| Freelancer or contractor | The annotator, as a contractor | Fastest | A small pilot with a clear deliverable. Weak once you set hours, guidelines and daily output. |
| BPO contract | The vendor, which also supervises | Depends on the vendor’s onboarding | You buy a defined queue with quality targets and do not want to manage the workforce yourself. |
| EOR | The EOR, on its Philippine payroll | Set up through the EOR | You want your own leads, tools and guidelines, but no Philippine entity yet. |
| Your own entity | Your Philippine company | Slowest, because incorporation and registrations come first | A large, long-term team with direct control of data and tooling. |
AI labelling work changes often, and that churn is why contractor labels break down. When you set the hours, the guidelines and the daily output, the arrangement looks like employment whatever the contract calls it. The independent contractor vs employee guide explains the test. For most teams, a BPO contract or an EOR is the safer route. The EOR vs BPO guide compares who employs and who manages in more detail.
Does the Data Privacy Act cover your training data?
Not always. It depends on where the data was collected. Section 4(g) of Republic Act 10173, the Data Privacy Act, excludes “personal information originally collected from residents of foreign jurisdictions,” processed under those jurisdictions’ laws, when that information “is being processed in the Philippines.” A buyer’s customer, patient or user data that was collected abroad and is annotated by Filipino staff will often fall within that exclusion. The buyer’s own country’s law follows that data.
The Act does apply to personal information collected in the Philippines, such as Filipino staff records or the data of Philippine customers. Whether any security duties still attach to data that falls within the exclusion is a question for Philippine counsel. This guide does not decide it.
Training data often contains names, voices, faces, messages and account details. Where the Act does apply, four provisions matter most for annotation work.
Sensitive personal information. Section 3(l) defines the term. It covers personal information about race, ethnic origin, age, and religious, philosophical or political affiliations. It also covers health, education, genetic or sexual life, and certain legal proceedings. It includes government-issued identifiers and records, such as social security numbers and tax returns.
Processing is restricted. Section 13 starts from a prohibition on processing sensitive personal information, then lists the cases where processing is allowed. Those rules apply to data inside the Act’s scope, so map each dataset’s origin before you decide which rules govern it.
Subcontracting. For data the Act covers, Section 14 lets a controller subcontract processing, but the controller must ensure proper safeguards are in place. Those safeguards must protect confidentiality, prevent use for unauthorized purposes, and generally comply with the Act.
Accountability. Section 21 makes each personal information controller responsible for personal information under its control or custody. That includes information transferred to a third party for processing, domestically or internationally. The controller must use contractual or other reasonable means to provide a comparable level of protection while a third party processes the data. For data the Act covers, the vendor’s assurance does not move your duty to the data subjects.
For data the Act covers, Section 20(f) requires the controller to notify the National Privacy Commission and affected data subjects promptly when sensitive personal information, or information that could enable identity fraud, is reasonably believed to have been acquired by an unauthorized person.
In practice, do four things before the first batch goes out. Map where each dataset was collected and under which law, because that decides whether the Act applies. Remove or redact personal information the annotation task does not need. Write access limits, security controls, breach notice timing and data deletion at contract end into the vendor agreement. Name the person on your side who answers for compliance. The Data Privacy Act and NPC registration guide covers your own registration and officer duties. The data protection officer guide covers appointing that officer.
How should content moderators’ wellbeing be protected?
Moderators see abusive, violent and sexual material as part of the job. Two legal layers apply.
The first is the general safety law. Republic Act 11058 is titled “An Act Strengthening Compliance with Occupational Safety and Health Standards and Providing Penalties for Violations Thereof.” Its declaration of policy says that “a safe and healthy workforce is an integral aspect of nation building.” The text of the act does not mention mental or psychosocial hazards, so do not cite it for that point.
The second is the mental health rule. According to Laborlaw.ph, DOLE Department Order No. 208, series of 2020, says it is mandatory for all workplaces and establishments to formulate a mental health workplace policy and program. Check the official text before you write the policy into a contract.
The order is the floor, not the whole job. Good practice for moderation teams includes limits on exposure per shift, rotation away from the most harmful queues, access to counselling, supervisor training, and a clear route for reporting harm. Write these into the vendor contract, and ask the vendor to show its written policy before go-live. Treat these as operating standards you set, not as rules quoted from statute.
Moderation work also tends to contain sensitive personal information. Where the Act covers a moderation queue, apply the Data Privacy Act points above to it. Where a queue holds foreign users’ data processed in the Philippines, Section 4(g) may take it outside the Act, so ask Philippine counsel what security duties still apply.
What do EU buyers need to know about the AI Act and GDPR?
Regulation (EU) 2024/1689, the AI Act, has an Article 10 on data and data governance. It applies to high-risk AI systems that are trained with data, and it sets quality and governance requirements for the training, validation and testing data sets used to build them.
Do not overstate the scope. Article 10 is aimed at high-risk systems. It does not cover every model you train, and whether your model is high risk is a classification question for counsel. A Philippine labelling vendor is not normally the party the Act regulates. The practical step is to pass your data governance steps to the vendor in the contract, so the same controls apply to the labels it produces.
GDPR transfer rules can also apply. Article 44 of the GDPR sets the general principle for transfers of personal data to a third country, and Chapter V sets out the conditions. If EU residents’ personal data goes to a Philippine vendor, ask counsel which transfer mechanism applies to that flow. This guide does not state which mechanism is available for a given vendor.
Who owns the labels and outputs?
Labels, preference rankings, rubrics and written guidelines are work product, and the ownership rules matter for every model you train on them. Under Section 178.4 of the Intellectual Property Code, Republic Act 8293, a work commissioned by someone other than the author’s employer is owned by the person who commissioned it. The copyright in the work stays with the creator, unless the parties agree otherwise in writing.
That default is the reason for a written IP assignment. Without one, the vendor or the individual annotator can keep the copyright in the output. Put the assignment in the vendor contract, and make sure it reaches the people who do the work. A vendor’s own employment contracts with its annotators should assign their work product to the vendor. The vendor’s agreement with you should then assign those rights on to you. Cover labels, rankings, rubrics, guidelines and any dataset derived from them. Have a Philippine lawyer check the wording, because the chain of title has two links.
How do you protect confidential model and product information?
This is general advice rather than a statutory rule. Sign NDAs with the vendor, and make sure the NDA binds the staff who see your material. Limit each annotator to the guidelines and data slices the task needs. Use named accounts, not shared logins, and log who opened what. Remove access on the day someone leaves a project. Where you run evaluations, keep the held-out test set and the model being judged away from the people doing the rating, if the design allows it.
What does it cost?
This guide gives no salary figures. Cost depends on the role and skill level, the shift, the mandatory benefits, and the vendor or EOR fee. Shift is the factor buyers miss most.
Article 86 of the Labor Code, Presidential Decree No. 442, provides that every employee shall be paid a night shift differential of not less than ten percent (10%) of the regular wage. The premium applies to each hour worked in the late-night period, which runs from 10 pm to 6 am. US Eastern business hours from 9 am to 5 pm in standard time fall inside that window when translated to Manila time. A daytime queue for a US buyer is therefore mostly a night shift for the agents.
Quote by shift, not by a flat hourly rate. The night shift differential guide explains how the premium is computed. If you need one person for a short pilot, the guide to hiring a Filipino virtual assistant legally covers the lawful routes.
Common mistakes
- Treating “we use a BPO” as the employment answer. The contract label does not settle who really directs the work. If you set the shifts and guidelines, the arrangement can look like your own employment.
- Sending raw personal data when redacted data would do. For data the Act covers, every extra field the annotator sees is one you must protect under Section 21.
- Assuming the vendor’s wellbeing policy exists. Ask for it in writing before the first moderation shift starts.
- Using contractors for a shift team. When you set hours, guidelines and daily output, a contractor label will not hold.
- Skipping the written IP assignment. Under Section 178.4, the copyright in the labels can stay with the creator unless you have a written agreement.
- Quoting an hourly rate without the night shift. Hours inside the 10 pm to 6 am window carry the differential.
- Forgetting EU transfer and AI Act questions. Ask counsel before the first EU dataset moves.
When should you move to your own Philippine entity?
Move when the team is large and stable enough that you want direct control of staff, data and tooling. A Philippine company gives you your own employment and data controls. It also means incorporation, registrations and permits before the first annotator starts. The software company registration page covers a tooling or data-products company, and the BPO company registration page covers the export test that applies to a services operation. The PEZA vs BOI comparison helps you choose between the two incentive routes. If you need a Philippine team employed by a third party while your entity is being built, see PEO services.
Next step
Write down three facts before you talk to any vendor: the data types the annotators will see, including any sensitive personal information; the volume and severity of moderation content each shift will handle; and the hours you need covered. Those facts decide the model, the processing terms, the wellbeing measures and the shift cost. Book a consultation and Chamberlain will review the contract scope and the compliance points before you sign.
This guide is general information, not legal advice on your facts. The rules for personal data, labour, copyright and EU law change, so confirm the current text before you rely on it.
Frequently asked questions
Do I need a Philippine company to hire data annotators?
No. A BPO contract lets the vendor employ and supervise the annotators, and an employer of record (EOR) lets the EOR employ them. You need your own Philippine company only when you want your own employees and registrations.
Am I still accountable for personal information my vendor processes?
For data the Data Privacy Act covers, yes. Section 21 makes the personal information controller responsible for information under its control, including information transferred to a third party for processing. Section 4(g) excludes personal information collected from foreign residents under their own laws and processed in the Philippines, so map each dataset's origin first.
Does Philippine law require a mental health policy for content moderators?
According to Laborlaw.ph, DOLE Department Order No. 208, series of 2020, says it is mandatory for all workplaces and establishments to formulate a mental health workplace policy and program. Check the official DOLE text before you rely on the wording.
Who owns the labels that a vendor's annotators create?
Under Section 178.4 of the Intellectual Property Code, the person who commissions the work owns it, but copyright stays with the creator unless the parties agree otherwise in writing. A written IP assignment closes that gap.
Does the EU AI Act apply to the training data I send to the Philippines?
Article 10 of Regulation (EU) 2024/1689 covers data and data governance for high-risk AI systems. It does not cover every model you train, so classify your system with counsel before you assume it applies.
Official sources
Primary references this guide is checked against.
- Lawphil: Republic Act No. 10173, Data Privacy Act of 2012
- Lawphil: Republic Act No. 11058, An Act Strengthening Compliance with Occupational Safety and Health Standards
- Laborlaw.ph: Mental health workplace policy (summarizes DOLE Department Order No. 208, s. 2020)
- Lawphil: Republic Act No. 8293, Intellectual Property Code of the Philippines
- Lawphil: Presidential Decree No. 442, Labor Code of the Philippines
- AI Act Article 10 (unofficial reproduction of Regulation (EU) 2024/1689)
- GDPR Article 44 (gdpr-info.eu mirror of Regulation (EU) 2016/679)
Related guides
Insurance Outsourcing Philippines: Legal Setup Guide
Outsource insurance and mortgage back office to the Philippines: which tasks stay onshore, the FTC, APRA and FCA vendor rules, and Data Privacy Act duties.
Legal Process Outsourcing Philippines: Roles, Limits, Setup
Legal process outsourcing from the Philippines: which paralegal, research and intake tasks fit, the Philippine bar limit and your supervision duties.
Real Estate Virtual Assistant Philippines: Legal Setup Guide
Hire a real estate virtual assistant in the Philippines. Which tasks need a licence, what Data Privacy Act and US call rules require, and which model fits.
