Joan Comadran
Field notes

Engineering

Hire an AI engineer: how to pick one who ships

A buyer's guide to hiring an AI engineer for a project, not a job opening: when you need one, what to ask, the red flags, and how to scope the first engagement.

November 18, 20269 min read

Hire an AI engineer when a model has to do a job inside your company and keep doing it after the demo. That job is an agent, a feature or a small app: it answers a customer, books something, or writes into software you already run. Pick the person by what they have kept in production, how they measure it, and what you own when they leave.

How do you hire an AI engineer?

Write the outcome before the job title. Name the task, the system it has to touch, and the signal that would show it worked. Then check you need a person. If you do, pay for a small slice with an eval set, then the production build, then the handover.

Pages that rank for this query are recruiting guides: job ads, salary bands, Boolean searches. If you already want me to build it, the offer is the AI engineering page. This one is how to choose.

Do you need an AI engineer at all?

Often, no. Rebuilding a product you could subscribe to is the slow way to get a Zap.

You haveDo this
A fixed workflow you already run: a form, a sheet, a notificationStay in no-code.
WhatsApp replies, qualification or booking, and a product's script and calendar fitBuy the product.
A prediction you must train, or a research questionHire an ML engineer or a data scientist.
Your own rules, writes into systems you run, a failure you cannot shrug offHire an AI engineer for the project.

The second row is a product I sell, so read it that way. Claient is the WhatsApp assistant I built and run: it replies on the business's number, qualifies the lead, and books in the chat. Sogas Mascaró, a family winery in the Penedès, uses it for visits to the winery and for a meeting with the CEO. If that script and calendar fit, subscribe. Prices, Meta's included, and the build-or-buy cut are in WhatsApp appointment booking with AI.

An ML engineer trains and serves a model. An AI engineer ships the product around one, usually a model they did not train. One project is the usual buy. No rate card here: what moves the quote, and the run cost, is in AI agent development cost.

AI engineer, ML engineer, agency, or a full-time hire

The same title covers all four, which is why a recruiting search and a project search collide.

Project AI engineerML engineerAgencyFull-time hire
You buyOne system, then the handoverA model and a measured resultCapacity and a scheduleA person who stays
Fits whenOne feature, agent or app, with an endA hosted model cannot make the predictionSeveral workstreams at onceYou will still be building in a year
You should hearAn incident, an eval set, cost per task, who owns the accountsData, drift, a serving budgetThe name of the person writing the toolsThe same questions, plus who reviews them later
Walk away whenThe work is a chat on a laptopThe deliverable is a slideYou never meet who does the workThe role is six jobs

A full-time hire is a year on payroll. One agent, and nobody to manage a new employee, is a project. If you use an agency, ask who writes the tools. "The team" is not a name.

What to ask before you hire

  • What is in production, and who uses it? On Claient, Sogas Mascaró's customers book a visit in the chat. You want a user, a channel, and something that broke.
  • Where is the eval set? Which cases must pass before a prompt or a model change ships, and what did the last change break? If the failures were never written down, you are the test set.
  • What does one conversation or task cost? Model name, what is cached, who pays the provider. The figures I could check are in AI agent development cost.
  • What happens when it is wrong? Timeout, duplicate delivery, a person asking for a human, a tool that wrote the wrong row. Retries and a handoff are the work.
  • Where do keys and data live? Keys do not belong in a transcript. A stored name, phone number or chat needs a place, an access list and a way to delete it.
  • What do you hold on the last day? Repo, cloud project, model account, prompts, eval set, and a note on how to run it.

If they open with six agents, ask which one earned its own context. Most of this work is one loop with tools: single agent vs multi-agent. If a coding agent will write much of the repo, ask how that code is reviewed: how to review AI-generated code.

Red flags

  • Demo-only work. A happy path and a screenshot. No user, no log, no incident.
  • No evals. "It feels sharp", or three hand-picked examples after a prompt edit.
  • No cost sketch. Not even on a named model. The invoice arrives on your card.
  • Vague ownership. Repo, cloud account, model key or WhatsApp number "on our side for now".
  • A diagram instead of a write path. Several agents, and silence on a tool that fails halfway.
  • Training first. Fine-tuning before a hosted model has met an eval set.

How a good first engagement is scoped

Start with a slice you can reject.

PhaseYou pay forYou hold
ProblemTask, systems it may write to, metric, cost of a wrong answerOne page you can both point at
Paid prototypeA thin slice on realistic inputs, cases it must pass and cases it must refuseA demo and the failure list
ProductionWrite path, logs, human takeover, cost per conversation or taskSomething a real user can hit
HandoverRepo, accounts, keys, prompts, eval set, a short runbookYou can run it without them

Skip the prototype only when the path and the data are already clear. Without a handover, the build stays in someone else's accounts. The prototype is paid.

What to prepare before the first call

  • Data access. A sample export, a sandbox, or a plain "production is off limits for now".
  • A success metric. "Answer after hours" is a wish. "Visits that do not double-book, and a human gets a stalled chat" is a metric.
  • A budget range. A ceiling tells you if this is a prototype or a product. Read the cost post before you borrow a number from a salary survey.
  • An owner. One person who can approve the scope, the data and the handover.

Hiring an AI engineer in Barcelona

I work from Barcelona, in English, Spanish and Catalan. On 18 November 2026 the city is on Central European Time (CET, UTC+1), six hours ahead of New York. Daylight saving ended on 25 October 2026; while it runs, the clock is Central European Summer Time (CEST, UTC+2). Source: timeanddate, checked 18 November 2026.

Two EU rules matter before a customer's chat is stored. This is not legal advice, and it does not pick a city.

A name, a phone number or the conversation is personal data under Regulation (EU) 2016/679. Ask where it sits, who can read it, and how a person gets it deleted.

Article 4 of Regulation (EU) 2024/1689 tells providers and deployers to support AI literacy for staff and others who operate the system for them. The Commission's AI literacy Q&A, checked 18 November 2026, says that duty has applied since 2 February 2025, that national market surveillance authorities enforce it from 2 August 2026, and that a Digital Omnibus amendment in force since mid-July 2026 requires no specific literacy level of any one person. EUR-Lex still shows the older "sufficient level" sentence, so the Q&A is the later account. It names no job title. Someone at the company still has to understand the system.

For the engagement itself, the offer is AI engineering. This page stays the guide.

FAQ

How do you hire an AI engineer?

Write down the task, the system it has to touch, and the metric that would show it worked. If a no-code flow or a product already does that, buy it. If it does not, hire for a small paid prototype with an eval set, a cost per conversation or task, and a handover of the repo, the accounts and the keys.

Do I need an AI engineer, or will a product do?

A product or a no-code flow is enough when the rules fit tools you can already buy. Hire an AI engineer when the system has to write into software you run, follow rules a template does not have, and fail in a way you can see. Training your own model is a different hire: an ML engineer or a data scientist.

What should I ask before I hire an AI engineer?

Ask what they have run in production, where the eval set is, what one conversation or task costs on a named model, what happens on a bad answer or a timeout, where the keys and the personal data live, and what you own at handover. A portfolio of demos with none of those answers is a demo.

How is a first AI project usually scoped?

Four beats: a written problem, a paid prototype with an eval set, production (the write path, logs, a human takeover, a cost line), then a handover of the repo, accounts, keys, prompts and eval set. There is no universal number of weeks. Scope follows how many systems the agent can write to.

About the author

I'm Joan Comadran, an industrial engineer (UPC, Tecnologías Industriales) turned AI engineer and full-stack developer in Barcelona. I am a problem-solving engineer who ships webs 10x faster with AI, plus apps, agents and AI tools. I built and run Claient; Sogas Mascaró uses it to book winery visits and a meeting with the CEO. That is the work.

Work with me

Want this in your product?

I build AI agents, apps and websites for companies.

Let's talk →

Keep reading

03