How to hire an AI developer without getting burned

Most AI hiring decisions are bad for one reason. The buyer talks to someone who knows AI, gets impressed, and signs. No one pressure-tested the actual build plan. Three months later, the shiny demo is a skeleton nobody uses.

This guide is what I wish every founder had read before booking their first AI consultant call.

The three kinds of "AI developer"

They all have the same job title. They do very different work.

1. The Prompt Engineer

Spends most of their time writing prompts for off-the-shelf LLMs. Great for marketing copy automation, content summarization, simple chatbots. Can absolutely help you. Usually cheap and fast.

Red flag: They pitch this as an "AI agent" or "autonomous system" when it is really a wrapper around ChatGPT.

2. The Integration Engineer

Builds glue between your data, your tools, and a foundation model. Workflows, RAG systems, internal dashboards. Most real business AI work lives here. Think n8n, LangChain, custom code.

Green flag: They ask to see your actual data and systems before quoting. They care about where data lives, who owns it, and how it flows.

3. The ML Engineer

Trains custom models. Works with fine-tuning, evals, model selection. You need this if you have a unique data moat or a regulated environment. You do not need this for most B2B automation.

Red flag: They default to "we will train a custom model" before understanding whether a foundation model already solves your problem.

Five questions that save you $20K

Ask these on the first call. Listen for the answer they give when they think you are not testing them.

1. "What is the minimum thing that would prove this idea works?"

If they go straight to a full build timeline, walk away. The right answer starts with a narrow test. Maybe one workflow. Maybe one customer segment. The AI world is full of projects that scoped themselves to death before shipping anything.

2. "What happens when the model gets it wrong?"

Every LLM hallucinates sometimes. Every classifier misclassifies sometimes. The question is not "will it fail" but "how do we catch failures before customers do."

Good answer: clear monitoring plan, confidence thresholds, human-in-the-loop for edge cases.

Bad answer: "The model is really accurate."

3. "Who maintains this after you are gone?"

The AI build is 20% of the total lifetime cost. Maintenance is 80%. If the consultant does not walk you through monitoring, retraining, cost control, and handoff, they are selling you a project instead of a system.

4. "Can you show me a project you shipped that is still running a year later?"

Not a demo. Not a case study PDF. Something currently in production. If every reference they have is less than 6 months old, they have never maintained an AI system. That matters.

5. "What would you tell me not to build?"

This is the most revealing question. A consultant worth their rate will have opinions about what is a waste of money. If every idea you propose sounds great to them, they are optimizing for your signature on the contract.

Green flags that beat every sales pitch

• They push back on scope. They tell you what to cut, not just what to add.
• They want to see real data before quoting. Not a sanitized sample. The actual mess.
• They quote maintenance alongside the build. Not as an afterthought.
• They name specific models and reasons. "Claude Sonnet for reasoning, GPT-4.1 for structured outputs, local Llama for high-volume routing." Generic answers mean generic builds.
• They show you something ugly that works. Pretty demos lie. Ugly internal tools that save someone 10 hours a week tell the truth.

The anti-pattern to watch for

The "AI transformation" pitch. You know the one. Six-figure retainer, strategy phase, capability audit, roadmap workshop. Three months in, no system has shipped.

Transformation is what consultants sell when they cannot build. Pick someone who ships in week one.

What I do on first calls

Half my calls end with me telling the prospect they do not need a custom build. A Zapier recipe works. A prompt in an existing tool works. Their problem is not AI-shaped.

This is not me being humble. It is protecting your money and my time. The builds that are worth doing are the ones where off-the-shelf genuinely falls short. The rest should not exist.

If you want that conversation, book a 25-minute call. If you want a 50-slide deck, I am not your guy.