AI in Hiring · 9 min read

What Are AI Agents? A Business Guide With Examples and Pitfalls

What are AI agents and when should a business use them? How they differ from chatbots and workflows, building blocks, hiring examples, pitfalls, a checklist.

Door Ingmar van Maurik · Founder & CEO, Making Moves


Short answer

An AI agent is a language model that plans its own steps and calls tools, such as a calendar, a database or an API, in a loop: plan, act, look at the result, choose the next step, until the goal is reached or it has to stop. That makes agents flexible, but also less predictable than a fixed workflow. For a business, the real question is not "can we build an agent?" but "where does that flexibility add value, and where does a human need to check?"

What exactly is an AI agent?

Most definitions boil down to the same three parts. An agent has:

  • A goal, for example "schedule an interview with this candidate and two interviewers next week".
  • Tools it can call: read calendars, draft an email, update a record in the ATS.
  • A loop: the model picks an action, gets the result back and decides the next step based on it.
  • The difference from ordinary software lies in that last part. In a classic script, the developer fixes the order in advance. With an agent, the model decides the order while it runs.

    Chatbot, workflow or agent: the difference

    ChatbotFixed workflow / automationAI agent

    |---|---|---|---|

    Who decides the steps?Nobody: one question, one answerThe builder, in advanceThe model, at run time Takes actions in systems?Usually notYes, always the same onesYes, chosen per situation PredictabilityHigh for form, variable for contentVery highLower, needs monitoring Good forAnswering questions from known informationRepeatable processes with fixed rulesTasks with variation, exceptions and several systems Typical riskMade-up answersBreaks on an unexpected caseWrong action, loops, rising costs

    A workflow can still contain a language model, for instance to classify an email. It remains a workflow as long as the steps are fixed.

    When a fixed workflow is the better choice

    Choose a deterministic workflow when:

  • The steps are always the same. New application in, send confirmation, create record.
  • Mistakes are expensive or irreversible and you need to explain every step to an auditor.
  • Volume is high and margins are thin. An agent often makes several model calls per task. A fixed workflow is cheaper and faster.
  • The rules are already known. If you can draw the process as a flowchart, you usually do not need an agent.
  • A useful rule of thumb: start with the simplest solution that works, and add autonomy only where the variation in the work calls for it.

    The building blocks of a reliable agent

    1. Model. The language model that plans and decides which tool to call.

    2. Tools and APIs. Each with a clear description, defined inputs and the fewest permissions possible.

    3. Memory and state. What has been done, which data has been fetched and where the task stands.

    4. Instructions. The goal, the boundaries and what the agent must do when something fails.

    5. Guardrails. Checks on input and output, limits on steps, time and cost, and an allowlist of permitted actions.

    6. Human approval. A required checkpoint for actions that are visible to outsiders or cannot be undone.

    7. Logging and tracing. Every step, tool call and result recorded, so you can trace an error back.

    Examples in recruitment and operations

    Scheduling interviews. The agent reads interviewer availability, proposes three slots and prepares a draft invitation. *Human in the loop:* the recruiter approves the invitation before it reaches the candidate, or the candidate picks from the proposed slots.

    Enriching candidate files. The agent gathers missing details from the ATS and submitted forms, summarises them and flags gaps. *Human in the loop:* the recruiter reviews the summary. The agent does not judge suitability and rejects no one. Under GDPR Article 22, a decision with a significant effect on a person, such as a rejection, may not be based solely on automated processing, apart from narrow exceptions that still require safeguards. So make sure a human genuinely decides, rather than just signing off.

    Answering candidate questions. The agent answers questions about the process, timeline and practicalities from approved information. *Human in the loop:* questions about pay, rejections or personal circumstances go straight to a recruiter. Candidates know they are talking to an AI assistant.

    Operations. An agent that sorts incoming requests, pulls data together from three systems and prepares a proposal. An employee carries out the proposal.

    In every example, the agent prepares and a human decides on anything that affects the candidate.

    Common failure modes

  • Endless retries. An API returns an error and the agent keeps trying. Fix: a maximum number of attempts with backoff, then stop and report.
  • Invented results when a tool fails. The calendar API times out, yet the agent reports the interview as booked. Fix: only report success based on a real success response from the system.
  • Permissions that are too broad. An agent that reads CVs does not need a tool to send email externally. See how far this can go in [prompt injection explained](/en/articles/prompt-injection-explained).
  • Cost runaway. Every step is a model call. Without a limit per task and per day, a loop can get expensive. More in [what an AI agent costs](/en/articles/ai-agent-cost).
  • No audit trail. If you cannot see what the agent did, you cannot explain or fix errors.
  • A checklist to get started

  • Pick one well-scoped process with a measurable goal, such as time to schedule.
  • Draw the process as a workflow first. Use an agent only where the flowchart falls short.
  • Give the agent only the tools and permissions it needs, ideally read-only at first.
  • Decide which actions need human approval.
  • Set limits on steps, time and cost per task.
  • Have logging and tracing in place from day one.
  • Build a test set of real cases, including awkward ones, and run it on every change.
  • Check GDPR and Article 22 whenever candidates or employees are affected.
  • How to tell whether someone can build reliable agents

    A demo tells you little. Ask about what goes wrong. Someone with real experience can explain straight away how the agent handles a tool that does not respond, why that tool has those permissions, where a human approves, how you see what the agent did and what it costs per month. Also ask when they would not build an agent. Anyone who always proposes an agent is not looking at your process.

    Want to test this in a structured way? The AI Skills Assessment combines scenario questions and a practical case covering agents, permissions and evaluation, among other topics. Read more about measuring these skills in how to assess AI skills. Looking to use an agent in your own hiring? See our AI hiring system.

    Frequently asked questions

    What are AI agents in plain language?

    Software in which a language model chooses its own next step and uses tools, such as a calendar or database, to do it, until a goal is reached.

    Is a chatbot an AI agent?

    Usually not. A chatbot answers questions. An agent also takes actions in systems and decides the order of those actions itself.

    Can an AI agent reject candidates?

    Not on its own. GDPR Article 22 restricts decisions based solely on automated processing that have a significant effect. Let the agent prepare and a human decide.

    How do I start with AI agents in my business?

    With one well-scoped process, minimal permissions, human approval for external actions, cost limits and a log of every step. Measure the result before you expand.

    Summary

  • An AI agent plans and calls tools in a loop towards a goal.
  • Use a fixed workflow where the steps are known. Use an agent only where variation requires it.
  • Build with minimal permissions, guardrails, human approval and tracing.
  • In hiring, the agent prepares and a human decides.
  • Test agent builders on what goes wrong, not on a demo.
  • Want to know whether your team or vendor can build reliable agents? See the AI Skills Assessment or get in touch.


    Book intake call · View our AI Hiring System