Before You Hire Agentic AI Developers, Read This Teardown

How to hire agentic AI developers, through a teardown of a WhatsApp bot that told leads they were booked while the calendar stayed empty, plus what to ask.

By Carlos Familia, CRM & AI automation expert. Updated October 16, 2026. 5 min.

"You're booked," the bilingual WhatsApp bot told leads of an insurance agency in Puerto Rico. The calendar stayed empty, and the client complained to me about it twice. The agency also said its cancer-policy leads "weren't coming in."

I opened the account and worked through five problems. Only the first was about the AI, and each one became a question you can put to a developer before you hire them.

To hire an agentic AI developer, make them prove the agent's actions instead of describing them. Ask for the tool-call log of one real booking, the test they ran against a real calendar and what happens when the agent is wrong. Anthropic's guide to building effective agents draws the line: workflows run LLMs and tools through predefined code paths, while agents are systems where the LLM directs its own process and tool use. That freedom is the risk you are hiring someone to manage.

What did the account hold?

Problem 1: The bot said "booked" and the calendar was empty

A language model writes the sentence that fits the conversation. Unless a tool call succeeded, nothing happened, and the model can't tell the difference unless the code tells it.

Agents also act on the wrong record when state leaks between conversations. On a credit repair firm's booking bot, a reused session ID attached a second booking to the first person. That is the agentic version of a wrong answer: a real action, on the wrong contact.

Here, I gave the bot a booking action pointed at the correct calendar, plus a prompt rule: no confirmation until the reservation is actually there. On my other bots I go further, so that a booking counts only when the tool reports success.

Problem 2: Two Facebook forms nobody had mapped

A lead form that is not mapped to fields drops its leads. The cancer-policy leads were stuck on the Facebook side, because nothing connected those two forms to fields in the CRM.

Mapping them to new fields backfilled 39 stuck leads in about 3 minutes. Ask a builder where each lead source lands, and ask them to submit a test lead from every form while you watch. My guide to Facebook lead forms in GoHighLevel walks through the mapping.

Problem 3: Backfilled leads started nothing

The 39 recovered leads arrived in the CRM and fired no workflows, so none of them got a follow-up. A recovered lead is not the same as a worked lead.

I built two workflows. One sends a cancer-policy lead to the supplemental pipeline and raises an alert. The other, on a confirmed appointment, moves the auto-insurance opportunity forward, never backward. Both were proven in the execution logs, not assumed from a successful publish.

Problem 4: My own QA booked a real appointment

The trial chat endpoint for GoHighLevel's Conversation AI books real appointments if the scripted lead accepts a slot. During my QA, one landed on the owner's own calendar.

That is the general trap with agents: a test that exercises a real tool is a real action. Ask for a separate test calendar before any testing begins, and ask the developer how they prove a test can't reach yours. The setup side is in GoHighLevel Conversation AI setup.

Problem 5: The system refused a second opportunity

This location rejects a second opportunity in the same pipeline, even when the first one is abandoned. A returning customer's new opportunity is refused, and any automation that expects one breaks for people who come back.

So the confirmed-appointment workflow moves the existing opportunity forward and never backward. Ask your builder what happens to a customer who returns a year later.

What does the market say about agentic AI projects?

Gartner predicted in a June 25, 2025 press release that more than 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. The same release estimated that only about 130 of the thousands of agentic AI vendors are real, and called the relabeling of existing tools as agentic "agent washing."

Anthropic's guide adds that the autonomy of agents means higher costs and the potential for compounding errors. Expect a good developer to ask whether you need an agent at all, or whether a fixed workflow does the job.

What makes a good paid test task?

Before the full project, pay for the smallest agent that can fail in ways you care about. Ask for an agent that books onto a test calendar, then break the booking tool on purpose. A sound build says it could not book and offers a next step. A weak one says "you're booked" anyway. Add a returning customer who already has an opportunity, ask for the tool-call log, and ask for the alert that fires when the agent stops.

A day or two of paid work shows more than any portfolio. It also shows how the developer reports a failure, which is most of the job.

What should you ask whoever builds yours?

1. "Walk me through the tool calls behind one booking a real customer made." A developer who can't may have an agent that narrates actions it never performed.
2. "Which rules are enforced in code, and which only in the prompt?" On the credit repair bot, Georgia residents are blocked twice, in the prompt and inside the booking tool. A Georgia test produced zero appointments, and a New Jersey test produced one.
3. "Where does each lead source land, and did you submit a test lead from each?"
4. "Which calendar do your tests write to, and how do you know it isn't mine?"
5. "What happens when a lead arrives by backfill, or a customer comes back?"
6. "What alerts me when the agent stops, and who receives it?"

For the wider buyer's view, read hiring an AI automation consultant. The chat-specific checks are in my AI chatbot developer scorecard, and n8n builds are covered in n8n AI agent mistakes. To put the scope in writing first, use the AI automation scope of work template, and the insurance side is in insurance agent CRM. To have me tear down your own agent, book a call.

About the author

Carlos Familia, CRM & AI automation expert. GoHighLevel, n8n, AI agents, servers. Full-stack developer.

I build CRM systems in GoHighLevel, n8n automations on my own servers and AI voice and chat agents for service businesses. Full-stack developer: I write the workflow, the API integration, the database and the server it runs on, and I test every build with real journeys before handoff.