AI Agency vs In-House Developers: The Year-One Cost of Each, Worked Out

Most “agency vs in-house” articles stop at a pros-and-cons list. This one puts numbers on the in-house side using public US salary data, explains how agencies actually structure their pricing so you can compare like with like, and ends with the contract terms that decide whether outsourcing leaves you with a working capability or just a bill.

It covers one decision: whether to build your own machine learning team or hire an outside firm. If you have already decided to commission a custom build and want to know what the phases and deliverables look like, read our guide to custom AI development. If what you need is off-the-shelf AI software rather than a build, see how to choose an AI software vendor.

What a minimal in-house ML capability looks like

One ML engineer is rarely enough. A model that runs in production needs someone who can frame the problem and build the model, someone who can get data into shape and wire the model into your systems, and someone accountable for priorities and quality. In a small company those jobs collapse into two or three people:

• Data scientist / ML engineer. Explores data, selects and trains models, evaluates results.
• Data or software engineer with MLOps duties. Builds data pipelines, deploys the model, sets up monitoring and retraining.
• Technical lead or manager (often optional at first if an existing CTO can absorb it). Sets priorities, reviews work, owns the relationship with the business.

Product ownership, QA and security review are also needed, but most companies borrow them from existing staff for a first project.

Year-one cost of an in-house team, using BLS data

The figures below use median annual wages for May 2025 from the US Bureau of Labor Statistics Occupational Outlook Handbook. BLS does not publish a separate “machine learning engineer” category, so we map each role to its nearest BLS occupation.

To get from salary to what the employer actually pays, we use the BLS Employer Costs for Employee Compensation release. For private-industry workers in management, professional and related occupations in June 2026, wages and salaries made up 68.5% of total compensation and benefits the other 31.5%. Dividing a salary by 0.685 gives a rough fully loaded cost.

Read these as a floor, not a quote, for four reasons:

1. Medians cover every experience level and every region. Experienced ML specialists in competitive markets earn well above them. Levels.fyi, which collects self-reported pay including bonus and equity, showed a US median total compensation of about $250,000 for ML/AI software engineers when we checked in October 2026.
2. The 31.5% benefits share is an average across a broad occupational group. Your own benefits package may cost more or less.
3. Salary is not the whole bill. Recruiting fees, laptops, software licenses, cloud compute for training and serving, and data-labeling costs all sit on top. Put your own numbers against each of these. They vary too much between companies to estimate here.
4. Year one is not a full productive year. Every month spent recruiting and onboarding is a month with no output. Salaries also start before the first model ships.

Demand is not easing either. BLS projects data scientist employment to grow 35% from 2025 to 2035, against 3% for all occupations, so expect to compete for these hires.

How agencies price ML work

Agency rates vary too widely by region, seniority and specialty for a single number to be useful, so we won’t quote one. What you can compare is the pricing structure, because each one moves risk to a different side of the table.

Many engagements combine these structures: a short paid discovery phase at a fixed fee, then a fixed or capped build, then a retainer for support. If you use time and materials, ask for a not-to-exceed ceiling and a burn report at every review.

To compare an agency with in-house, ask for a proposal priced by phase, then line it up against the in-house table above for the same scope and the same time frame. Also price what happens after launch. An agency quote that ends at deployment is not comparable with a salaried team that keeps maintaining the model.

The hybrid model: the agency builds, your owner learns

For many mid-sized companies the realistic answer is neither option on its own. An agency builds the first version while one internal person, often an existing engineer or analyst, works alongside it and takes over afterwards. That keeps the first-year cost closer to a project fee than to a full team, and still leaves you with someone who understands the system.

For this to work:

• Name the internal owner before kickoff, not at the end. Give them real time on the project, not a few hours on Fridays.
• Have the owner pair with the agency on real work: running training jobs, reviewing evaluation results, deploying to staging.
• Define “handed over” as something the owner can do unaided, for example retrain on new data, redeploy and roll back. Don’t define it as the date the agency stops invoicing.
• Keep a small support retainer for a period after handover, so the owner has a safety net while gaining confidence.

If the owner is not technical enough to retrain a model, the realistic outcome is an ongoing agency retainer. That is a legitimate choice, but budget for it rather than discover it later.

Decision matrix

Whichever way you go, the most common failure has nothing to do with who writes the code. In interviews with 65 experienced data scientists and engineers, RAND researchers found that misunderstanding or miscommunicating the problem to be solved was the leading root cause of AI project failure. Write down the business problem and the success metric before you compare costs. If you are unsure whether the problem calls for machine learning at all, this explainer on AI, ML and deep learning is a useful starting point.

Handover clauses to insist on

If you outsource, the contract determines whether you can ever bring the work in-house. Ask your lawyer to check for each of these:

• IP assignment that names source code, trained model weights, training and evaluation scripts, prompts and configuration, not just “deliverables.”
• Your accounts from day one. Repositories, cloud projects, model registries and third-party API keys sit in accounts your company owns, with the agency added as users.
• Documentation as a paid, accepted deliverable: architecture overview, data sources and lineage, how to retrain and redeploy, known limitations, and monitoring runbooks.
• Reproducibility. Your team can rebuild the model from the delivered code and data, and the contract says so.
• Knowledge transfer schedule with named sessions, pairing time and a handover acceptance test that your internal owner performs.
• Data handling and return. What client data the agency may keep, and confirmation that copies are deleted at the end.
• Exit assistance. A defined period and rate for help after termination, so switching providers or moving in-house is not held hostage.
• Hiring terms. Check any non-solicitation clause. If you might want to hire someone who worked on your system, negotiate that now.

If a model will make or influence decisions about customers or staff, also agree who reviews its outputs after handover. Our post on human-in-the-loop AI covers how to design that review.

Comparing providers

If the numbers point toward an agency or a hybrid setup, you can compare firms by specialty in our AI/ML development provider directory. How we review providers explains how listings are assessed. Empirical Edge, which runs this site, is one of the listed firms.