How Do I Test an AI Developer Before Hiring? A Paid 3-Hour Assessment for 2026
September 12, 2026 · Borderless Recruit Team
The best way to test an AI developer before hiring is to combine three forms of evidence: a detailed review of one system the candidate actually helped ship, a paid three-hour work sample resembling your own project, and a live session in which the candidate explains and debugs the result. Give every candidate the same repository, synthetic data, time limit, AI-tool policy, acceptance tests, and weighted scorecard. Test whether the solution works, but also inspect evaluation design, failure handling, security, API cost awareness, documentation, and communication. For an AI agent, verify tool permissions and human escalation. For retrieval-augmented generation, or RAG, test retrieval quality and unsupported answers. For an n8n, Make, Zapier, CRM, or API specialist, test idempotency, retries, logging, credential handling, and maintainability. Permit controlled use of ChatGPT, Copilot, Cursor, or similar tools because tool use is part of modern development, but require disclosure and an unaided explanation before the candidate starts. A polished demo should not outweigh reproducible evidence. The strongest candidate can clarify an ambiguous requirement, make sensible trade-offs, verify AI-generated output, and leave another person able to operate the system.
What should you test before hiring an AI developer?
Test the work the person will own after hiring. A business integrating an LLM into an existing product needs different evidence from one building predictive models, a RAG knowledge assistant, an autonomous agent, or CRM automations. An impressive chatbot demonstration does not establish competence in data pipelines, model evaluation, Salesforce or HubSpot APIs, production monitoring, or n8n maintenance.
- Technical execution: Does the code or workflow run reliably and satisfy the written acceptance tests?
- Reasoning: Can the candidate explain assumptions, alternatives, and why the chosen design fits the business constraint?
- AI reliability: Are outputs evaluated against labeled examples, with explicit handling for uncertainty and unsupported responses?
- Production judgment: Are latency, API limits, token or execution costs, logging, retries, observability, and maintenance addressed?
- Security: Are secrets excluded from code, access minimized, sensitive data protected, and high-risk actions gated?
- Ownership: Can another developer understand, deploy, test, monitor, and modify the work?
- Remote collaboration: Does the candidate write clearly, raise blockers early, document decisions, and honor the agreed work schedule?
This emphasis on practical evidence reflects developer preferences. According to HackerRank's 2025 survey of 13,732 respondents across 102 countries, 66% of developers preferred practical coding challenges and 96% said problem-solving should count more than memorization. The same report found that 78% believed assessments did not align with real work, making a generic puzzle a weak substitute for the production responsibilities you need to evaluate. HackerRank 2025 Developer Skills Report
Define the AI role and business outcome first
Write a one-page role scorecard before selecting a test. Start with the business outcome, current systems, data sensitivity, expected deliverables, and who will maintain the result. For example, replacing manual lead routing is an automation problem until custom classification, retrieval, or decision logic makes an applied AI developer necessary. Training and deploying predictive models calls for a machine learning engineer. Building a support assistant over company documents calls for RAG and evaluation experience.
- Outcome: the measurable operational change, such as reducing unassigned leads or making approved knowledge easier to retrieve.
- Environment: languages, cloud services, databases, CRMs, APIs, n8n, Make, Zapier, and model providers already in use.
- Constraints: security rules, acceptable latency, expected volume, budget limits, regulated data, and required human review.
- Ownership: new builds, production support, incident response, documentation, vendor management, and ongoing optimization.
- Working arrangement: dedicated full-time responsibilities, required US-hour overlap, communication cadence, and escalation path.
Do not use the title AI developer as a catch-all. The US Bureau of Labor Statistics distinguishes software developers, data scientists, and computer and information research scientists by their underlying duties. Research-scientist work commonly requires deeper theoretical preparation than integrating an existing model API into a business application. BLS software-developer profile BLS data-scientist profile BLS computer-research-scientist profile
Choose a test that matches the type of AI developer
Job-relevant three-hour test templates
| Role | Representative assignment | Evidence to grade |
|---|---|---|
| Applied AI developer | Complete a small feature that calls an LLM API, returns structured output, and handles timeouts and invalid responses. | API design, schema validation, tests, fallback behavior, latency and usage-cost awareness, and maintainable code. |
| Machine learning engineer | Train or improve a small model using supplied tabular data with a fixed train-validation split and business error costs. | Data leakage checks, baseline selection, metric choice, reproducibility, error analysis, deployment thinking, and monitoring for drift. |
| RAG specialist | Repair a retrieval pipeline over a small document set containing duplicate, outdated, and irrelevant passages. | Chunking and metadata choices, retrieval evaluation, grounding, abstention, permissions, traceability to retrieved material, and failure analysis. |
| AI agent developer | Add one tool to a constrained agent and prevent it from executing an unauthorized or ambiguous action. | Tool schemas, state handling, permission boundaries, loop limits, audit logs, deterministic checks, and human approval. |
| Automation engineer | Repair an n8n, Make, or Zapier workflow that receives a webhook, updates a CRM, calls an AI classifier, and sends an exception to a review queue. | Idempotency, mapping, pagination, retries, rate limits, credential handling, logging, replay strategy, and readable handoff notes. |
A small test cannot prove every production skill. Use it to sample the highest-risk work, then verify adjacent experience through a live technical discussion. If the continuing role centers on CRM and workflow ownership, an automation assessment is more predictive than asking the candidate to build a model from scratch. The distinctions in AI automation specialist versus AI developer costs can help scope the position before testing.
Step 1: Verify identity, employment history, and claimed ownership
Begin with a short video call and compare the candidate's identity, résumé timeline, portfolio, and public professional history. With consent, verify relevant employment or client relationships through appropriate channels. International checks should be proportionate to the role and handled under applicable privacy and employment rules; the overseas employee background-check guide explains the broader process.
- Ask the candidate to open a project they claim to have built and navigate its architecture, commit history, tests, deployment configuration, and issue history.
- For GitHub work, inspect contribution dates and commit content rather than treating contribution counts as proof. Ask why specific decisions were made.
- Separate personal responsibility from team output: what did the candidate design, implement, review, deploy, monitor, and maintain?
- For confidential offshore projects, accept redacted diagrams or a reconstructed example, then probe technical decisions without requesting a former employer's proprietary material.
- Confirm that the person completing the assessment is the person being considered by keeping the live explanation tied to the submitted work.
Privacy restrictions may prevent a candidate from sharing production code. That is not automatically a red flag. Inability to describe a system's inputs, failure modes, evaluation method, operational incidents, and personal contribution is more concerning than the absence of screenshots.
Step 2: Review one real production system in depth
Ask the candidate to choose one deployed system and walk through it from business request to maintenance. A useful discussion covers the original baseline, users, data sources, architecture, model or automation choices, evaluation criteria, deployment, incidents, and later changes. Request measurable technical outcomes only when the candidate can define how they were measured; do not reward unsupported revenue or productivity claims.
- What failed after launch, and how was the failure detected?
- Which parts were deterministic software and which depended on a probabilistic model?
- How did you evaluate quality before and after changing the model, prompt, retrieval configuration, or workflow?
- How were credentials, personal information, tenant access, and production permissions handled?
- What did inference, automation runs, storage, observability, and third-party tools cost?
- Who responded when a provider changed an API or a workflow began failing?
- What would you redesign now, and why?
Step 3: Assign a paid three-hour test based on your work
Pay candidates for a substantial work sample and keep it small enough to finish in three hours. Use synthetic or de-identified data, a sandbox account, and a repository created for assessment. Do not ask candidates to solve an unpaid production backlog. Give identical core instructions and acceptance criteria to everyone while allowing written clarification questions.
Copy-and-use assessment brief
- Scenario: A webhook receives a new support request. The system must classify its topic, retrieve relevant approved guidance, create or update the CRM record, and route uncertain or sensitive cases to a human.
- Materials: A small multi-file repository or exported automation; API mocks; 20 synthetic support messages; a document set containing one outdated policy and one duplicate; 12 labeled evaluation cases; failing tests; and an environment-variable template with no live secrets.
- Deliverables: A working branch or workflow export, automated tests or reproducible test steps, evaluation results, a concise README, an architecture sketch, and a list of assumptions and unresolved risks.
- Time allocation: 15 minutes to read and clarify, 105 minutes to build or debug, 30 minutes for tests and evaluation, 20 minutes for documentation, and 10 minutes to prepare the handoff.
- AI-tool rule: Tools are permitted, but the candidate must list what was used, preserve relevant prompts or summaries when practical, and remain able to explain and modify every submitted component.
- Data rule: Use only the supplied synthetic data and sandbox credentials. No production customer data may be pasted into a public AI service.
- Submission rule: Stop after three hours and describe what remains. Finishing a smaller safe solution is preferable to hiding incomplete or unverified behavior.
Acceptance criteria for the sample assessment
| Area | Minimum acceptable evidence |
|---|---|
| Execution | The project starts from documented instructions, or the workflow imports cleanly, and the primary happy path completes. |
| Correctness | The supplied acceptance tests pass or failures are explicitly documented; structured outputs are validated before use. |
| Reliability | Retries are bounded, duplicate events do not create duplicate records, and model or API failures reach a visible exception path. |
| AI quality | The candidate runs the labeled cases, reports a relevant metric or categorized errors, and avoids presenting unsupported output as established fact. |
| Security | Secrets remain outside source code, logs avoid sensitive content, permissions are minimized, and consequential actions require appropriate approval. |
| Operations | Logs identify failed steps, a maintainer can replay or recover work safely, and usage-cost drivers are identified. |
| Handoff | The README states setup, assumptions, limitations, test procedure, and recommended next steps. |
The repository should contain at least one realistic defect rather than only greenfield coding. CoderPad's approximately 650-participant 2026 survey found that 38% of recruiters used real-world simulations, 38% assessed system design, and 30% used pair programming, while 43% still used algorithm exercises. In the same report, 69% used résumé review, but only 16% believed résumés predicted performance; live coding and technical discussion ranked highest as predictors. CoderPad State of Tech Hiring 2026
Step 4: Run a live debugging and system-design session
Schedule a 45- to 60-minute follow-up in which the candidate shares the submitted work. Begin by asking for an architecture explanation without AI assistance. Introduce one new requirement or failure, such as a CRM rate limit, an expired token, a malicious document, a duplicate webhook, or a model-provider outage. Observe how the candidate investigates before changing code.
Ask the candidate to explain logs, reproduce the issue, identify the likely fault boundary, propose the smallest safe change, and add a test. Then discuss how the system would behave at higher volume, across multiple customers, and during partial failure. The goal is not theatrical speed. It is evidence that the candidate can reason from symptoms, communicate uncertainty, verify a fix, and avoid making the incident worse.
A staged process has a documented industry precedent. CoderPad's 2026 report describes McAfee using an initial technical screen for baseline capability followed by live coding for further validation. This is an example reported by CoderPad, not a claim about the results your company will obtain. CoderPad State of Tech Hiring 2026
Step 5: Test evaluation, reliability, and hallucination controls
An AI feature is not finished when one demonstration looks convincing. Ask what dataset represents normal cases, edge cases, adversarial inputs, and costly mistakes. A good candidate selects metrics that reflect the business risk, records a baseline, inspects errors, and explains the threshold for human review. For generative systems, evaluation may combine deterministic checks, labeled human judgments, retrieval metrics, and carefully validated model-based grading.
- RAG: measure whether relevant passages are retrieved, whether the answer is supported by the retrieved material, and whether the system abstains when evidence is inadequate.
- Agents: test unauthorized tool requests, prompt injection, repeated tool calls, ambiguous instructions, partial completion, and escalation to a human.
- Structured extraction: validate schemas, required fields, ranges, and reconciliation totals before updating a CRM or accounting system.
- Classifiers: inspect false positives and false negatives separately because their business costs may differ.
- Automations: test duplicates, out-of-order events, pagination, rate limits, expired credentials, provider downtime, and replay after failure.
- Maintenance: require a regression set that can be rerun after changing prompts, models, workflows, connectors, or source documents.
A randomized study involving 95 professional developers asked participants to build the same JavaScript HTTP server and scored correctness with tests. Participants given GitHub Copilot completed the task 55.8% faster. The narrow experiment does not establish the same effect for every project, but it supports evaluating tool-assisted work against objective tests rather than judging typing speed alone. The Impact of AI on Developer Productivity
Step 6: Assess security, privacy, and production guardrails
Give the candidate a threat-oriented scenario. Ask what happens if a retrieved document contains malicious instructions, a user requests another customer's data, an agent tries to send an email without approval, or an API response is logged with personal information. Security competence appears in specific boundaries and tests, not in a generic promise to follow best practices.
- Use sandbox accounts, synthetic data, least-privilege roles, short-lived credentials where supported, and separate development from production.
- Keep secrets in an approved secret store or environment configuration, never in source code, workflow exports, prompts, screenshots, or chat history.
- Authorize access at the application and data layers; a prompt telling a model not to reveal information is not an access-control system.
- Require human confirmation for high-impact communications, purchases, deletions, financial actions, or irreversible CRM updates.
- Log decisions and tool calls without unnecessarily retaining sensitive inputs or model responses.
- Define a disable switch, incident owner, credential-rotation procedure, rollback path, retention rules, and vendor-data review.
Should candidates be allowed to use ChatGPT, Copilot, or Cursor?
For an AI-enabled development role, controlled use is usually more informative than a total ban. According to HackerRank's 2025 report, 97% of surveyed developers used at least one AI assistant, 61% used two or more, and respondents estimated that AI generated 29% of their code on average. A realistic test can therefore reveal whether a candidate verifies generated code, protects data, and recognizes a plausible but incorrect suggestion. HackerRank 2025 Developer Skills Report
According to CoderPad's approximately 650-participant 2026 survey, 34% of hiring leaders banned AI in technical assessments, 46% allowed it broadly or with constraints, and 20% decided case by case. CoderPad also reported that MNTN, Meta, and Yahoo permitted AI while assessing proficiency with it. MNTN's published approach asks for the initial reasoning before tool use, then permits assistance while preserving observable judgment. CoderPad State of Tech Hiring 2026
Recommended controlled-AI assessment policy
| Stage | AI policy | What it reveals |
|---|---|---|
| Initial explanation | No assistant for the first 10 to 15 minutes. | Independent problem framing, fundamentals, assumptions, and planned architecture. |
| Paid implementation | AI tools permitted with disclosure; no production data or undisclosed outside help. | Prompting, workflow integration, verification, productivity, and security judgment. |
| Live review | Candidate must explain and modify any submitted code or workflow. | Authorship, understanding, debugging ability, and responsibility for tool output. |
| Short fundamentals check | No AI assistance. | Core knowledge required to detect and repair incorrect generated output. |
Step 7: Evaluate communication and Philippine remote-work readiness
Remote technical work is already common: in 2025, 22.4% of US people at work teleworked for some or all hours, and the rate was 65.4% for computer and mathematical occupations. BLS telework annual averages The Census Bureau separately reported that more than 22 million US workers usually worked from home in 2023, representing 13.8% of workers compared with 5.7% in 2019. The measures use different definitions and should not be combined. US Census Bureau home-based worker analysis
For a developer in the Philippines supporting a US business, define overlap rather than asking whether the candidate is flexible. US-hour coverage may require an evening or night schedule in the Philippines. Confirm the exact recurring hours, daylight-saving changes, on-call expectations, local holidays, backup connectivity, equipment, and a sustainable rest schedule before an offer.
- Send one deliberately incomplete written requirement and score whether the candidate identifies the missing business rules.
- Require a concise end-of-test update covering completed work, evidence, blockers, risks, and the next action.
- Ask for an asynchronous handoff that another developer could follow without a meeting.
- Simulate a production issue and assess response discipline, escalation, status updates, and documented recovery.
- Verify primary internet, power, workspace, and backup arrangements without demanding unnecessary personal details.
- Evaluate spoken and written English in work situations rather than relying only on a conversational interview.
The Philippines has a substantial services-delivery base. According to IBPAP data reported by the Philippine News Agency in January 2025, Philippine IT-BPM export revenue grew 7% to $38 billion in 2024 while direct employment reached 1.82 million. Philippine News Agency report citing IBPAP IBPAP's original Roadmap 2028 identified the Philippines and India as the two largest IT-BPM delivery destinations, estimated a 16%–18% Philippine share of global delivery in 2022, and named intelligent automation, analytics, cloud migration, and cybersecurity as growth areas. More than 60% of surveyed Philippine IT-BPM organizations ranked workforce upskilling as their top priority. Philippine IT-BPM Industry Roadmap 2028
AI developer test scorecard: what a passing performance looks like
Score evidence, not confidence or presentation polish. A defensible starting point is 100 weighted points, a 75-point passing threshold, and mandatory minimums in correctness and security. Use the same rubric for all candidates in the same role. Record short evidence notes beside each score so interviewers can reconcile disagreements.
Weighted AI developer assessment scorecard
| Category | Weight | Full-credit evidence | Serious concern |
|---|---|---|---|
| Functional correctness | 20 | Core requirements work; tests are reproducible; edge cases are addressed or clearly documented. | Demo works only on one happy path, failures are hidden, or results cannot be reproduced. |
| Technical reasoning and architecture | 20 | Candidate clarifies requirements, compares alternatives, and explains boundaries and trade-offs. | Buzzwords replace reasoning, or the design is much more complex than the problem requires. |
| AI evaluation and reliability | 15 | Relevant baseline, labeled cases, error analysis, fallback behavior, and regression method are present. | Quality is judged from a few impressive outputs or model claims alone. |
| Security and privacy | 15 | Least privilege, safe secret handling, data boundaries, validation, approvals, and abuse cases are considered. | Secrets appear in code, untrusted model output directly triggers sensitive actions, or tenant boundaries are ignored. |
| Code or workflow maintainability | 10 | Clear structure, useful names, limited duplication, tests, versionable artifacts, and manageable dependencies. | Opaque generated code, tangled workflows, undocumented nodes, or manual steps only the author understands. |
| Operations and cost awareness | 10 | Logging, retries, rate limits, recovery, monitoring, latency, model usage, and software costs are addressed. | No operational plan, unlimited loops or retries, or no understanding of variable API costs. |
| Communication and documentation | 10 | Updates are concise, assumptions are explicit, blockers surface early, and the handoff is usable. | Candidate conceals incomplete work, cannot explain the submission, or leaves no reproducible setup. |
A suggested mandatory minimum is 10 of 20 for correctness and 8 of 15 for security. A candidate scoring 80 overall should not pass if credentials are committed or an autonomous agent can perform high-impact actions without authorization. Conversely, a candidate who stops safely, documents an unresolved defect, and proposes a sound fix may deserve more credit than someone who conceals it behind a polished interface.
Strong signals and red flags
Evidence to look for during screening
| Strong signal | Corresponding red flag |
|---|---|
| Asks who uses the output and what an error costs before selecting a model or tool. | Starts with a fashionable framework before understanding the business outcome. |
| Can separate personal contributions from team accomplishments and navigate implementation details. | Claims ownership of an entire system but cannot explain commits, failures, deployment, or maintenance. |
| Creates a baseline, tests edge cases, and discusses false positives, false negatives, or unsupported responses. | Uses a handful of favorable prompts as proof that the feature is reliable. |
| Validates AI-generated output and willingly replaces it when a simpler deterministic rule works better. | Treats model output as trusted or submits generated code that cannot be explained. |
| Plans for retries, idempotency, observability, API changes, and credential rotation. | Assumes a successful demonstration means the workflow will remain operational. |
| Explains variable software and API costs separately from compensation. | Promises savings or ROI without usage assumptions, baselines, or measurement. |
| Raises privacy and access questions before requesting real data. | Suggests pasting customer data or production credentials into an unapproved tool. |
How to test an AI developer when you are not technical
A nontechnical owner can still run a fair assessment by making the desired behavior observable. Prepare input examples, expected outcomes, unacceptable outcomes, data restrictions, and operational constraints. Ask an independent technical reviewer to create or review the repository, acceptance tests, and security rubric. Keep the reviewer separate from any candidate or agency that benefits from recommending a more complicated build.
- Require a one-page plain-English design before implementation.
- Use predefined test cases and make the candidate demonstrate failures as well as successes.
- Ask what happens when the model provider, CRM, database, or automation platform is unavailable.
- Ask the candidate to estimate monthly software and API usage under low, expected, and high volumes, listing every assumption.
- Require setup, rollback, monitoring, and ownership documentation before treating the work as complete.
- Have the reviewer score the work independently, then compare scores and evidence.
The reviewer should not select a candidate from résumé claims alone. CoderPad's 2026 findings indicate that recruiters themselves viewed live coding and technical discussions as more predictive than résumé review. The business owner should retain the final decision because technical quality is only one part of fit; ownership, communication, schedule, and judgment also matter. CoderPad State of Tech Hiring 2026
What does a dedicated AI or automation hire cost?
Budget with separate columns for worker salary, employer costs, recruiting or EOR fees, equipment, and variable software or API usage. Salary is what the worker earns. Total compensation adds employer-funded benefits and statutory costs. An EOR planning figure may also include 13th-month accrual, allowances, administration, and a provider fee. Model tokens, vector databases, cloud hosting, n8n or Make plans, Zapier tasks, CRM licenses, monitoring, and data tools remain operating expenses.
US, Philippine, and Latin American staffing benchmarks
| Role or benchmark | US employee benchmark | Remote or Philippine benchmark | Interpretation |
|---|---|---|---|
| Software developer — US versus Philippines | $6,872–$17,889 per month; median $11,332, using May 2025 BLS 10th–90th percentile and median wages. Applying the June 2026 broad full-time private-industry wage share gives an illustrative median total compensation cost of about $16,543 per month, excluding company overhead. | $2,000–$5,000+ monthly salary; $2,500–$5,950+ monthly all-in Philippine EOR planning cost in the July 9, 2026 Smart Outsourcing Solution dataset. | Salary-only difference from the US median is 55.9%–82.4%, calculated from these ranges. The Philippine provider separately reports 35%–70% all-in savings depending on seniority and benefits. |
| QA engineer — US versus Philippines | $5,120–$13,918 per month; median $8,692 for US software QA analysts and testers. The broad benefit ratio implies illustrative median total compensation of about $12,689 per month, excluding overhead. | $1,800–$3,500 monthly salary; $2,250–$4,200 monthly all-in Philippine EOR planning cost. | Salary-only difference from the US median is 59.7%–79.3%, calculated from the cited ranges. |
| Data scientist or AI analytics specialist | $5,603–$16,594 per month; median $10,019 for US data scientists. The broad benefit ratio implies illustrative median total compensation of about $14,627 per month, excluding overhead. | $1,500–$4,500 monthly salary; $1,900–$5,300 monthly all-in for a Philippine remote data analyst. | The Philippine figure is an adjacent data-analyst benchmark, not a verified data-scientist or machine-learning-engineer salary. |
| DevOps engineer — Philippines | Not supplied in this comparison. | $3,000–$6,000+ monthly salary; $3,550–$7,000+ monthly all-in EOR planning cost. | Useful when the AI role includes cloud deployment, observability, and production operations. |
| Data engineer — Philippines | Not supplied in this comparison. | $3,000–$6,000+ monthly salary; $3,550–$7,000+ monthly all-in EOR planning cost. | Useful when pipelines and data quality are a substantial part of the role. |
| Software developer — US versus Latin America | $11,332 monthly median US salary; illustrative total compensation of about $16,543 per month using the broad June 2026 benefit ratio. | $4,417–$5,250 per month, converted from Howdy's $53,000–$63,000 annual average based on verified 2025 payroll data covering more than 12,500 records in seven named Latin American countries. | Salary-only difference from the US median is 53.7%–61.0%. Howdy separately reports about 60%–65% savings under its fully loaded model. |
| Senior software, DevOps, AI, or data developer — Latin America | Not supplied in this comparison. | $4,583–$5,833 per month, converted from Howdy's verified 2025 senior-developer band of $55,000–$70,000 per year. | The band explicitly includes senior DevOps, AI, and data specializations. |
| Philippine domestic software-developer benchmark | Not applicable. | ₱66,180 per month in August 2024, approximately $1,157 using the BSP monthly average of ₱57.1935 per US dollar. | This Philippine Statistics Authority figure covers software developers in publishing activities at formal establishments with at least 10 workers. It is not a remote-for-US salary. |
According to the US Bureau of Labor Statistics, the May 2025 median software-developer salary was $135,980 per year, while benefits represented 31.5% of total compensation for full-time private-industry workers in June 2026. The 68.5% wage share makes total compensation roughly 46% above wages, but it is an economy-wide planning ratio rather than a software-specific benefit rate. BLS software-developer wages BLS Employer Costs for Employee Compensation, June 2026
The Philippine remote figures come from a provider dataset covering January 1, 2025 through June 30, 2026 and updated July 9, 2026. Its all-in estimates include salary, 13th-month accrual, estimated statutory contributions, selected benefits or allowances, and a separately identified $190 monthly EOR fee. They are planning ranges, not a universal market tariff. Philippines Salary Guide for Remote Teams
According to the Philippine Statistics Authority's August 2024 Occupational Wages Survey, software developers in publishing activities averaged ₱66,180 per month in formal establishments with at least 10 workers. The dollar conversion uses the Bangko Sentral ng Pilipinas August 2024 average exchange rate, but domestic formal-sector pay should not be substituted for compensation offered to a specialist working remotely for a US employer. Philippine Statistics Authority wage survey Bangko Sentral ng Pilipinas exchange-rate table
Philippines, Latin America, freelancers, or an agency?
Hiring-model trade-offs for a continuing AI workload
| Option | Best fit | Main trade-off |
|---|---|---|
| US employee | Roles requiring substantial local presence, local labor-market knowledge, or frequent in-person work. | Highest direct benchmark in the supplied comparison; salary, benefits, recruiting, equipment, and overhead should be budgeted separately. |
| Dedicated Philippine employee through an EOR or staffing arrangement | Continuing product, RAG, agent, API, CRM, n8n, Make, or Zapier ownership where documented processes and a stable team member matter. | Lower entry point in the supplied datasets, but US-hour overlap may require an evening or night schedule. Verify seniority rather than inferring it from location or rate. |
| Latin American nearshore employee | Continuing development requiring substantial synchronous overlap with US business hours. | Howdy's verified 2025 payroll range is higher than the Philippine provider's developer range, while time-zone overlap is generally easier. |
| Independent freelancer | A genuinely bounded project with clear acceptance criteria and limited continuing integration. | Invoice pricing may exclude benefits and administration. Long-term control and integration can create classification questions, so obtain jurisdiction-specific advice. |
| AI development agency | A defined project requiring multiple specialties or temporary delivery capacity. | The buyer may get broader coverage but less direct control over who performs and maintains the work. Confirm code, account, prompt, workflow, and documentation ownership. |
Howdy's dataset covers more than 12,500 developer payroll records and places its 2025 Latin American average at $53,000–$63,000 annually. It is first-party payroll evidence from one provider, not a complete regional census. Howdy Latin America developer salary data The Philippines versus Latin America AI developer comparison examines this buyer decision in more detail.
The broader markets also differ. The Inter-American Development Bank estimated in 2022 that nearshoring could add $78 billion annually to Latin American and Caribbean exports in the near to medium term, including $14 billion in services. Inter-American Development Bank UN Trade and Development reported that digitally deliverable services represented 56% of global services exports in 2024, while Africa and Latin America and the Caribbean together supplied 2.5% of global ICT-services exports. UNCTAD digital economy insights These regional statistics describe market context, not individual candidate quality.
Worker classification and social protection should not be treated as administrative details. An International Labour Organization survey published in April 2025 covered 1,153 web-platform workers across 21 Latin American and Caribbean countries; 40% lacked social-security coverage, and 52% used platform work for secondary income. It is evidence about surveyed platform workers, not dedicated employees or Philippine workers, but it illustrates why buyers should distinguish a freelancer marketplace from a locally employed full-time arrangement. ILO survey on web-based digital-platform workers
What to include in the paid-test and employment documents
The assessment agreement should state the time limit, payment, permitted tools, confidentiality, data restrictions, ownership of the test submission, and whether the company may use it. Avoid giving candidates live customer information or production access. If the exercise resembles commercially useful work, be explicit about compensation and rights rather than disguising production work as recruitment.
- Defined test scope, fixed maximum time, payment amount and method, submission deadline, and evaluation rubric.
- Confidentiality obligations applying to both supplied materials and candidate work.
- Intellectual-property terms identifying ownership or any limited evaluation license for code, prompts, workflow exports, diagrams, tests, and documentation.
- A prohibition on production credentials, real customer data, or unauthorized third-party material.
- An AI-tool and outside-assistance policy, including disclosure and responsibility for generated output.
- Deletion or return requirements for assessment data and temporary credentials.
- A statement that passing the test does not itself guarantee employment.
For the continuing relationship, address IP assignment, confidentiality, local employment terms, payroll, statutory obligations, acceptable use, device and account controls, data processing, incident reporting, and termination access procedures. Contractor and employee classification depends on facts and jurisdiction, not the label on an agreement. The legal guide to hiring an AI developer in the Philippines provides additional planning context, but businesses should obtain qualified advice for their circumstances.
Validate the hire with 30-, 60-, and 90-day milestones
A pre-hire assessment is a sample, not a guarantee. Continue validation during onboarding with written milestones tied to the real system. The first 30 days should emphasize access discipline, architecture comprehension, reproduction of the current environment, small fixes, and documentation. By day 60, the developer should own a bounded feature or workflow, add regression coverage, and demonstrate monitoring and recovery. By day 90, evaluate independent delivery, maintenance quality, estimation, stakeholder communication, and whether operating costs and incidents are visible.
- Day 30: complete security and system onboarding; map dependencies; reproduce builds or imports; resolve a low-risk issue; document unclear ownership.
- Day 60: deliver a bounded production change with tests, evaluation evidence, rollback instructions, and stakeholder handoff.
- Day 90: operate the assigned area through a normal release cycle; respond to a simulated or real incident; propose a prioritized maintenance plan.
- At every milestone: review code and workflow ownership, documentation quality, model and software usage costs, communication, schedule reliability, and unresolved risk.
Do not wait until day 90 to discuss a serious problem. Use weekly evidence reviews during the initial period and document coaching, access changes, and decisions. Any probation, termination, or replacement provisions should be written into the applicable employment or staffing agreement and reviewed for the relevant jurisdictions; they should not be assumed.
A practical next step for hiring a Philippine AI developer
Start by converting one real business problem into the role scorecard and three-hour brief above. If you want recruitment support, Borderless Recruit can source a dedicated full-time AI Developer or AI Automation Specialist in the Philippines and handle recruitment, local contracts, payroll, and HR. Candidates undergo an English interview, reliability assessment, and practical skills screening. Use the same client-approved work sample and weighted rubric for finalists, particularly when you do not have an in-house AI expert.
Live catalog pricing starts at $1,750 per month for an AI developer, $2,000 for a machine learning engineer, $1,500 for a chatbot developer, and $1,300 for a prompt engineer. Automation-specialist pricing starts at $1,300, with AI agent builders from $1,600 and API integration developers from $1,500. These are role-specific service prices, not market salary statistics or estimates of software and API usage. The useful next step is to document the outcome, systems, data restrictions, work schedule, and assessment rubric before requesting candidates.
Frequently Asked Questions
How do I evaluate an AI developer before hiring?
Review one production system in depth, assign the same paid job-relevant work sample to every finalist, and conduct a live explanation and debugging session. Score correctness, reasoning, AI evaluation, security, maintainability, operations, documentation, and communication with a written rubric.
What skills should I look for when hiring an AI developer?
Look for solid software fundamentals, API and data-pipeline competence, evaluation design, failure handling, security, testing, deployment, monitoring, cost awareness, and maintainable documentation. Add role-specific evidence for RAG, machine learning, AI agents, or n8n, Make, Zapier, and CRM integrations.
What questions should I ask an AI developer in an interview?
Ask what failed after a system launched, how quality was measured, how unsupported output was handled, what the candidate personally owned, and how the system was monitored and maintained. Also ask how credentials and sensitive data were protected, what drove operating costs, and what the candidate would redesign.
How do I verify an AI developer's past work?
Ask the candidate to navigate architecture, commits, tests, deployment configuration, and issue history while explaining specific decisions and personal contributions. When prior work is confidential, use redacted diagrams or a reconstructed example, then verify identity and relevant employment with consent through appropriate channels.
Should I let candidates use AI tools during a coding assessment?
Controlled use is usually appropriate for an AI-enabled role, provided candidates disclose their tools, protect assessment data, and explain and modify everything they submit. Pair the tool-assisted task with a short unaided planning or fundamentals segment so you can observe both independent reasoning and verification of AI output.
Is three hours enough to test an AI developer?
Three hours is enough to sample job-relevant execution when the repository, synthetic data, scope, and acceptance tests are prepared in advance. It is not enough to prove every skill, so combine it with portfolio verification, a live technical review, references where appropriate, and post-hire milestones.