RAG Developer vs AI Agent Developer: UK & EU Hiring Guide

Hire a RAG developer when the main problem is finding authorised information and producing answers grounded in company data. Hire an AI agent developer when the system must choose steps, call tools or coordinate a workflow. If it must consult policies before updating business systems, look for a developer who can build both retrieval and controlled execution. Define the required behaviour before choosing the job title.

RAG developer vs AI agent developer: the practical difference

Retrieval-augmented generation, or RAG, supplies retrieved evidence to a large language model. A typical pipeline prepares documents, indexes their content and selects relevant passages for an answer. A published financial-services implementation used structured and unstructured sources to help call-centre representatives find information NAACL Industry paper. Retrieval can also be a tool within an agent. Agents do not necessarily change records: some only research or recommend actions. The hiring distinction is how much control over the next step the model receives.

How Borderless Recruit helps

Borderless Recruit sources and vets dedicated Philippine AI developers who work only for your business, during your UK or European business hours. It handles local employment contracts, payroll and HR through a flat monthly fee. The standard AI Developer catalogue price is from £1,330 / €1,540 per month ($1,750), covering recruitment for LLM applications, agents and RAG.

What each specialist should own

RAG developer responsibilities

• Map source systems, document owners, permissions, update frequency and the questions users need answered.
• Build repeatable ingestion, including parsing, metadata, versioning and deletion.
• Choose document chunking, keyword or semantic search and result ranking based on representative content.
• Evaluate retrieval separately from answer generation: a plausible answer can conceal a failure to find the right source.
• Monitor failed ingestion, stale content, unsupported answers, response times and usage costs.
• Document how sources are added, updated, removed and recovered.

AI agent developer responsibilities

• Separate fixed process rules from decisions the model may make, with explicit human approval points.
• Design narrowly scoped tools for APIs and business systems; enforce permissions outside the model.
• Manage execution state, time-outs, retries and duplicate prevention.
• Test ambiguous instructions, malicious content, rejected approvals and unavailable dependencies.
• Define how to recover from partial completion, including actions that cannot simply be rolled back.
• Maintain integrations, tests and incident instructions after launch.

A hiring decision tree for employers

• Choose a RAG developer when employees need reliable access to answers across policies, manuals, contracts or product documentation.
• Choose an AI agent developer when language interpretation or variable routing must guide tool use.
• Choose a hybrid developer when the system must retrieve authoritative evidence before proposing or executing an action.
• Choose ordinary automation when the rules and required steps are predictable; updating a CRM does not itself require an agent.
• Start with data and process work when records conflict, access is unclear or nobody can define the correct outcome.

Illustration: a support assistant that explains a cancellation policy needs retrieval and answer evaluation. If it also proposes an account change, the developer must validate the customer's identity, check eligibility and request approval before execution. If every step follows fixed rules, a conventional workflow may be sufficient. Write these boundaries into the vacancy brief so candidates can challenge unnecessary complexity.

A hybrid engineer may suit a focused internal application with accessible APIs and a named product owner. Broader systems may need data engineering, backend or platform support. Ask candidates to identify dependencies they cannot own alone, especially authentication, source-system permissions and business approval rules.

Skills matrix: evidence to request

A realistic paid assessment

Ask for a small working system using synthetic documents and sandbox credentials. Define the expected behaviour in advance, pay for the agreed scope and require a short handover. Use the same brief for comparable candidates. The practical AI developer testing guide provides a broader assessment process; the task below distinguishes retrieval skills from execution skills.

Illustrative brief: provide a fictional returns policy, an obsolete version, restricted customer notes and a sandbox ticketing API. The RAG candidate should answer policy questions with traceable evidence. The agent candidate should use the policy to propose a ticket update, obtain approval and submit it. A hybrid candidate should demonstrate the complete path.

During handover, ask the candidate to explain a failed test, change a permission and trace a request through the logs. Reject solutions that expose credentials, bypass approvals or cannot demonstrate access isolation, even if the normal demonstration works. For consequential systems, use an experienced technical reviewer to inspect the implementation.

Employer cost: compare the same scope

Do not assume the title alone determines cost. A document assistant with straightforward access rules and an agent spanning sensitive systems require different engineering effort. Compare the total monthly staffing fee with local employer costs, then allow for infrastructure and management on both sides.

For a worked UK illustration, the supplied catalogue models annual gross pay of £62,000–£100,000 plus a 16% employer-cost allowance, giving approximately £5,993–£9,667 per month. Its Netherlands model uses €58,500–€93,500 plus 17%, giving approximately €5,704–€9,116 per month. These are calculated catalogue planning scenarios, rounded to whole currency units catalogue cost assumptions. The calculation adds the stated allowance to annual salary, then divides the result across the months of the year.

The catalogue identifies ITJobsWatch AI Engineer data and CBS occupational wage data as salary references. Treat the catalogue ranges as modelling inputs, not current quotations from those pages or separate RAG and agent salary surveys. The live UK data changes, and the Netherlands range requires assumptions about occupation and annual working time.

The allowances are simplified budgeting assumptions, not statutory contribution rates or complete payroll calculations. UK employer National Insurance is generally 15% above the applicable secondary threshold GOV.UK employer rates. The usual auto-enrolment minimum employer pension contribution is 3% on the relevant earnings basis GOV.UK workplace pensions. Dutch contributions depend on the employer and employment circumstances Belastingdienst payroll tables.

To produce a fully loaded local budget, replace those allowances with your payroll calculation and add applicable benefits, recruitment, equipment and management costs. The catalogue staffing fee is the client's cost, not the developer's salary. Budget separately for model usage, storage, cloud hosting, monitoring and software licences. No savings percentage is implied: scope, seniority and operating costs must be comparable.

A contractor may suit a bounded build. For UK engagements through an intermediary, check whether off-payroll working rules apply and who must determine status; responsibility depends partly on the client's sector and size GOV.UK off-payroll guidance.

Ownership, data access and maintenance

Put source code, cloud resources, model accounts, indexes and monitoring in client-controlled environments. Give the developer individual access with only the permissions needed. Contracts should address ownership of code and documentation; the operating plan should name who approves releases, handles incidents and accepts business risk.

Remote access from the Philippines can engage international-transfer requirements even when personal data remains hosted in the UK or Europe. Map the organisations involved and their controller or processor roles before deciding which safeguards apply. Use the ICO international-transfer guidance for UK GDPR and the European Commission's international data-protection guidance for EU GDPR. Hosting location alone does not settle the question.

• Use approved test data, separate development and production environments, managed access and multifactor authentication.
• Apply document permissions before retrieved content reaches the model; do not rely on a prompt to hide restricted material.
• Keep credentials outside prompts and code, and validate tool requests against the user's authority.
• Version prompts, retrieval settings, tool definitions and evaluation datasets alongside application code.
• Maintain a manual fallback and test recovery before expanding permissions.

A focused onboarding sequence

• Discovery: agree the users, source owners, prohibited actions and acceptance tests.
• Prototype: build against controlled data and demonstrate the most consequential failure cases.
• Pilot: restrict users and permissions, compare outputs with a manual baseline and require approval for consequential actions.
• Production: establish monitoring, incident ownership, cost alerts and handover documentation.
• Expansion: add sources or tools only after evaluating the changed permissions and behaviour.

The specialist works your UK or European business hours through a Philippine shift aligned with your team's working day. The Philippines does not observe daylight saving, so agree how the shift changes when UK or European clocks change. Give the hire a named product owner and access to colleagues who understand the process.

Before approving the hire

Make the hiring decision on demonstrated behaviour: evidence retrieval, permission enforcement, controlled tool use and failure recovery. Confirm who owns the data, approvals and ongoing maintenance. If the candidate cannot explain how the system refuses an unsupported answer or stops an unauthorised action, the specification and assessment need more work before production access is granted.