Hiring an AI Freelancer: How to Evaluate and Protect Your Business

Hiring an AI Freelancer: How to Evaluate and Protect Your Business

When you are hiring an AI freelancer, how to evaluate them comes down to four steps. Define the process and the expected result first. Ask for proof of similar work. Limit access to the minimum. Then put confidentiality, ownership of what gets built and documentation in the contract. The biggest risk is rarely technical. It is handing data and keys to someone who has no method for testing, logging and handing the work over.

Key takeaways

• Define the process, the baseline and the success metric before you talk to candidates. Without them, every proposal looks good.
• Good screening asks how the person tests, logs errors and stops the system, not only which tools they know.
• A deliverable is not just a working agent. It includes documentation, tests, decision logs and a handover plan.
• Access should be minimal, temporary and on separate accounts, with keys your company controls and can revoke.
• The contract must cover confidentiality, personal data, ownership, subcontracting and what happens when the freelancer leaves.

What should you define before looking for someone?

Hiring usually goes wrong when the request is vague: I want an AI agent for customer service. A competent freelancer will ask questions. A less careful one will fill the gap with whatever they already know how to build, and you find out later that the scope was something else.

Before you open the conversation, write one page covering:

• The current process, step by step, with who does each step today.
• The exact pain point: lost time, a recurring error, a queue, rework.
• The data involved: what is personal, what is confidential, what can leave the company and what cannot.
• The actions the AI may take on its own, the ones that need human approval and the ones it must never take.
• The pilot's success metric and who in the company decides whether it continues.

If you cannot answer half of this yet, the first job is not hiring a developer. It is a short diagnosis of the process. To set boundaries on what should not be handed to an automated system, read what not to delegate to AI in your business.

What screening questions should you ask an AI freelancer?

A resume and a list of tools say little, because almost everyone uses the same models and platforms. What separates a reliable professional is operational discipline. Use open questions and notice whether the answers are specific or generic.

• Describe a similar project without naming the client. What was the process, what did you limit and what went wrong?
• How do you know the agent is getting things right? What tests or evaluations do you run before going live?
• What happens when the model makes a mistake or invents something? Is there a fallback, a block or a handoff to a human?
• How do you stop the system from looping, overspending or taking an improper action?
• What data do you really need to see, and what can be replaced with anonymized examples?
• Where will company data be processed? Which providers and which regions are in the flow?
• Do you subcontract anyone or use other providers during the work?
• What documentation do you leave behind? Could someone without you operate and fix the system?
• What would you refuse to do on this project, and why?

Warning signs: a promise of guaranteed results, a refusal to do a small test, a request for full access on day one, no interest in records and logs, and answers that talk only about prompts and never about failures.

Positive signs: the person questions the scope, proposes starting with a single flow, talks about action limits, accepts that the company owns the accounts and keys, and describes what they will document. For the conversation about agent evaluations, Anthropic publishes useful material on how to test agents that can help you steer the discussion.

How do you run a proof of work without exposing the company?

Ask for a paid, small test using fictional or anonymized data. It can reproduce a slice of the flow: receive a small set of sample messages, classify them, reply within set rules and record the reason for each decision. The goal is not to see a polished system. It is to see how the person organizes the work, what they test and whether they explain failures honestly.

Compare a few candidates using the same script. The difference shows up more in how they handle hard cases than easy ones.

Which deliverables should the contract require?

An agent that works in the demo but that only its author can change is dependency disguised as delivery. List the deliverables in writing, with an acceptance criterion for each one.

• Flow description: what the system does, what it does not do and when it hands off to a human.
• Versioned instructions and settings, in a repository or folder the company controls.
• A set of test cases with expected results, including error and refusal cases.
• Decision log: what the agent answered or did, with date, input and reason, for auditing.
• Configured limits: maximum steps, time, spend and permitted actions.
• A list of integrations, permissions used and keys in use, with rotation instructions.
• Operations guide: how to pause, roll back and fix, and who to call in an incident.
• A recorded or in-person handover session with the team that will operate it.

If the work includes writing code, also require review and tests. On how far to trust tests written by models, see how to validate AI-generated tests.

Tie payment milestones to accepted deliverables, not just hours worked. This aligns incentives and keeps documentation on the schedule instead of a task left for the end.

Which access should you give a freelancer, and which not?

The principle is least privilege: access only to what the task requires, for as long as it requires. The OWASP guidance on authorization sums up the logic well: deny by default and grant explicitly.

A simple decision guide:

• Company-wide administrator account: do not share it. Create a separate user for the freelancer with a restricted role.
• Real customer data: start with an anonymized sample. Release real data only in the pilot, limited to what is needed.
• API keys for models and services: create keys dedicated to the project, on a company account, with a spending limit, and revoke them at the end.
• WhatsApp and social media: use a test number and account first. Before any outbound sending, check the opt-in and template rules, as explained in WhatsApp outbound messages.
• Databases and internal systems: read access before write access, a staging environment before production.
• AI coding tools: configure permissions and isolation, following Claude Code security for teams.
• The freelancer's personal accounts: do not accept company data passing through their personal AI accounts. This should be written into the contract.

Everything created should live in company accounts: repository, cloud project, keys, monitoring dashboard. That way, when the contract ends, revoking access is something you do, not something you ask them to do. An internal policy on what can be pasted into AI tools also helps. See a company policy for employee ChatGPT use.

What should the contract say about confidentiality, data and documentation?

This article does not replace a lawyer. But some clauses are often forgotten in contracts with independent contractors and are worth taking to legal review.

• Confidentiality: a definition of confidential information (processes, customer lists, agent instructions, keys), how long the obligation lasts, and return or deletion at the end of the work.
• Personal data: when the freelancer processes personal data on the company's behalf, they act as a processor under LGPD (Brazil's data protection law) and must follow the company's instructions. Record the purpose, types of data, retention and deletion. The national data protection authority, the ANPD, publishes guidance.
• Providers and international transfer: list which AI and cloud services may be used and forbid changes without notice. If any provider processes data outside Brazil, see international data transfer under LGPD when using AI providers.
• Subcontracting: forbidden without prior written authorization, with the same obligations passed down.
• Ownership: who owns the code, the instructions, the flows and the materials produced. Provide in writing for assignment to the company, and for what the freelancer may reuse in a generic way without exposing your information.
• Documentation as an obligation: the operations guide, logs and test cases as part of acceptance, not as a courtesy.
• Incidents: a deadline for the freelancer to report a leak, improper access or unexpected agent behavior.
• Use of data to train models: forbid company data from being used to train or fine-tune third-party models unless expressly authorized.
• Termination: return of access, key rotation, delivery of everything to a company repository and a support period for questions.
• Liability: who is responsible for failures, how they are fixed and within what timeframe, without promising a business result.

If the system makes decisions that affect people, also include the human review that LGPD provides for, discussed in LGPD automated decision-making and the right to review.

How do you run the first project, step by step?

An illustrative scenario: a services company wants an agent to answer frequently asked questions on WhatsApp and pass quote requests to the sales team. It hires a freelancer and follows this sequence.

• Map: the company and the freelancer describe the current flow, note message volume by type and define how response time will be measured before anything changes.
• Limit: the agent only answers questions from an approved list, does not promise deadlines or prices, has no access to the finance system and hands off to a person when there is a complaint or sensitive data.
• Pilot: a single flow, first on a test number, then with a small group of contacts who have already consented. Each answer is logged with its reason. The team reviews the logs every week.
• Hand over: at the end of the pilot, the freelancer delivers documentation, test cases and a training session. The company decides whether to expand, adjust or stop, based on the metric agreed at the start.

This is, in essence, the order of work Juan Carlo uses in his projects: map the process, limit data and actions, pilot a real flow with logs and hand the operation over. To define metrics and decision criteria, how to measure an AI automation pilot offers a framework.

One last note: hiring an independent contractor works best when the project is well bounded. If the need includes ongoing monitoring, on-call support and maintenance, ask whether a single professional can cover absences, vacations and incidents, or whether you need an operation with more than one person.

Frequently asked questions

When does it make sense to hire a freelancer instead of an agency?

It makes sense when the scope is clear, the flow is single and your team can review and operate the result afterward. If the project requires on-call support, continuous monitoring or several specialties at once, compare it with a team, because depending on one person becomes an operational risk.

Can the freelancer use my customers' real data during development?

Ideally you start with anonymized or fictional data and release real data only in the pilot, limited to what is needed and with a recorded purpose. Because they will handle personal data on the company's behalf, the contract should define instructions, retention, deletion and which AI providers may receive that information.

Who owns the code and the agent they build?

That depends on what the contract says, and it should not be left implicit. Ask for an express provision on ownership or assignment to the company, keep everything in repositories and accounts you control, and define what the professional may reuse in a generic way without exposing your information. Take the text to a lawyer.

How do I know the work was done well if I do not understand AI?

Ask for objective acceptance criteria: test cases with expected results, readable decision logs, configured action limits and a demonstration that includes error situations. Someone on your team should be able to pause the system and understand a log entry without the author's help. If that is not possible, the delivery is not complete yet.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• OWASP: guia de autorização (Authorization Cheat Sheet)
• Anthropic: Demystifying evals for AI agents
• NIST AI Risk Management Framework