Hire an AI Engineer or an Agency? The Year-One Math Compared

Hire an AI engineer if three things are true. First, AI is your product rather than your tool, and will be for years. Second, you have a tech lead who can manage and challenge them. Third, there will be steady work in year two, after the first build is done.

If any one is missing, have a team build the first live system on a fixed price. Build it on your own accounts, and train your people to run it. The year-one math most often agrees. Most people leave three things out of the sum. One is the full, loaded cost of the salary. One is the months the seat sits empty. The last is who pays when the date slips.

How much does an AI engineer cost?

Start with pay. The Bureau of Labor Statistics (BLS) doesn't track "AI engineer" as its own job. The nearest job type is data scientist. Median pay was $120,230 in May 2025. Jobs are projected to grow 35% from 2025 to 2035. Treat that pay as the floor. Our working figure is $150,000 a year. That is for a full-time AI developer who has shipped live systems. And every firm that wants that person reads the same growth chart.

Then add the loading. That is what a job costs on top of the wage. BLS runs a survey of what employers spend on pay and benefits. In June 2026, wages were 70.0% of that spend for private employers. Benefits were 30.0%. That is about 43 cents of benefits on top of every wage dollar.

At that mix, a $150,000 salary costs about $214,300 a year. And that is before a laptop, a recruiter's fee, the tools or a single model bill. The mix is an average across all private jobs. Use your HR number if you have one.

How long does it take to hire an AI engineer?

Six months is our working figure. That runs from posting the role to a start date. The hire is someone who has shipped AI into production. The search is not the costly part. The empty seat is.

Put a price on the wait. Say the constraint you are hiring for costs $40,000 a month. That is a made-up figure, so use your own. Six months of searching is $240,000 of constraint. You carried it while no one built a thing. Add the first year on payroll, and that is $454,300 over eighteen months. For one person.

That $40,000 is the line people skip, because it never shows up as an invoice. It belongs in every option's column. Count it once for each month until that option has a live system running.

The four ways to get AI built, compared

Read the third row twice. In three columns the date is your problem. And you pay for the slip twice. First you pay in the vendor's hours or the employee's salary. Then you pay in the constraint, which keeps costing what it cost on the day you started.

What goes wrong with a freelancer, and with an hourly agency

Two composites, both common. Each one is drawn from many cases, not one client.

The freelancer. You have messaged them three times. The last reply said "almost there, just finishing the integration." You open the project folder to see how almost. There is a half-built integration and no notes. The only person who knows how it works isn't answering. They weren't dishonest. They got a bigger job. That is what freelancers do. It is why something that has to run for years shouldn't live on one person's calendar.

A good freelancer is right for a small, closed piece of work with a clear end. The failure is asking one person to be a whole team.

The hourly agency. The monthly invoice lists every hour, line by line. Integration work, 31 hours. Integration work (continued), 44 hours. Project management, 12 hours. Status meeting about the project, 6 hours. Nothing on it is dishonest.

That is the trouble. An hourly agency is paid more when the project takes longer than it quoted. Its margin grows with your hours. Your timeline wants fewer. And nothing in the contract settles it. The full math is in fixed price vs hourly AI development.

When you should hire the AI engineer

You can hire a full-time AI engineer, which is great. It is the right call when the three conditions hold. AI is your product, not your tool, and will be for years. You have a tech lead who can manage and challenge them. And there is steady work in year two, when the first build is done. An engineer with nothing to build in year two is a costly engineer.

There is also a hybrid. For most mid-sized firms it is the best of the options. Have the first system built on your own accounts. List the docs, SOPs and training in the scope. SOPs are written how-to guides for each task. Then hire or train someone to run it. You get the skill in the house without paying for an empty seat first. And you hire for a job you can now describe.

What to ask any team before you sign

Before you give anyone your data, ask for four things in one meeting. If they can't show all four, you are hiring one person with a team's logo.

1. Named engineers. Not "our team". Ask who writes the code and what they have shipped. Ask if they will sit in a shared channel with your people from the first day.
2. A delivery record. Builds of the same weight as yours. Ask how many, and for a reference you can call.
3. Security, built into the design. Not badges. Row-level security at the database engine on every query. That means the database itself checks who may see each record. Every value the system pulls out is stored with its source document, page and a snippet of proof. Database changes are numbered, and each one can be rolled back. Every model call is traced. People approve where it matters.
4. A handover pack. Code, IP, accounts, docs, SOPs and training. They should be listed in the scope as things you get, not promised on a call. IP is the legal right to what was built.

There is a sales-call version of this list in AI agency red flags. It gives the question that exposes each warning sign.

Is it better to build AI in-house?

The instinct is right about ownership and wrong about what is hard.

RAND asked 65 seasoned data scientists and engineers why AI projects fail. Here is the most common root cause. Leaders misread the problem the AI should solve, or explain it badly. A new hire doesn't fix that. They get the same brief. A written constraint with a price on it does.

The second hard part is range. The skills change from build to build. Look at our own recent work. One was a research copilot for a 50-person team of editors. It works across 10,000+ documents. One was a voice receptionist that answers every inbound call. And one was an agent for a property manager. It makes 14 live function calls into the system of record. That is the main system where the business keeps its data. Each one called for its own specialists. One hire is one slice of that list.

So the question that matters is who owns what gets built. Build on your accounts from day 0. Get the code, the IP and the docs handed over. Then moving it in-house later is a choice, not a rescue.

When not to hire anyone for AI

Here are three cases where none of the four options is the answer.

• The job is a landing page or a basic CRM setup. No AI creates the return. A good freelancer will do either well.
• You already have a senior in-house team and a clear plan for the build. Build it yourselves. The questions above still apply to your own team.
• You want a promise about revenue. No one honest can give you one. A team can promise a date and a definition of done. It doesn't control your market.