AI agency vs in-house team: what each costs, and the embedded option
Ask whether to use an AI agency or build an in-house team and you will collect confident answers in both directions, mostly from people selling one of the two. Here is ours, with our bias stated up front: we run the third model described below. If AI matters to your business, you will want in-house capability eventually. The real question is what you do during the 12 to 24 months it takes to hire that team and make it effective, because your market will not pause while you recruit.
That question has math in it. So let us do the math first, look at what each path trades away, and then at the middle option we think most mid-size companies should take.
What hiring in-house costs
Robert Half’s 2027 technology salary guide puts the US midpoint salary for an AI engineer at 175,000 dollars, with the high end at 196,750. That is base salary. Add payroll charges, equity, tooling and a recruiter’s fee, and the fully loaded figure sits well above it. European salaries are lower on paper, but for the profile you want, the competition is the same, because that profile is rare everywhere.
The bigger problem is the size of the pool. "AI experience" on a CV usually means coursework, a fine-tuning tutorial, or a chatbot demo. People who have shipped LLM systems to production, with evals, monitoring, cost control and an incident history behind them, are a small subset, and they field offers from labs and well-funded startups that a mid-size company will struggle to match.
Then there is time. A search takes months, ramp-up takes months more, and your first AI hire is the hardest to evaluate because nobody in the building has done the job before. We have seen a company spend the better part of a year hiring a head of AI whose first deliverable was a strategy document, while a competitor shipped.
What the agency route trades away
The classic agency model: you write a spec, they build off-site, they deliver, they leave. For AI systems this fails in a specific way. The spec cannot be written up front, because what the system must do gets discovered by putting a prototype in front of real workflows and real data. Each iteration routed through an account manager adds a week of latency to a loop that should turn in days, and that latency is one of the quiet reasons most AI projects die between the demo and the workflow.
The second failure is that the knowledge leaves with the builders. You receive working code, nobody on your side can explain why the retrieval behaves the way it does, and the maintenance retainer quietly becomes a dependency. You end up renting a capability you thought you had bought. If you are evaluating that classic model anyway, run the twelve questions to ask an AI vendor before signing.
The middle path: an external team inside your walls
The third model is the one we run: senior engineers embed in your team, in your repos and on your infrastructure, and build with your people rather than for them. The prototype runs on your real data within 2 to 4 weeks, production hardening happens in your CI and your monitoring, and your engineers pair on the system so the operating knowledge stays when the engagement ends. The mechanics are laid out in forward-deployed engineering.
This changes the hiring problem instead of replacing it. Once a system is in production, you know what the job requires, you can interview candidates against a running system instead of a hope, and the role you post is "own this" rather than "figure out where to start". Hiring after is easier, cheaper to de-risk, and more attractive to good candidates than hiring before.
How to decide between agency, in-house and embedded
Three questions settle most cases. Is the AI differentiating or generic? For generic needs (meeting notes, standard support macros) buy off the shelf, following the logic of build vs buy. How urgent is the value? If a competitor is cutting quote turnaround from days to hours, a 12-month hiring plan is a decision to concede that ground. Do you already have strong engineers near the problem? If yes, they can carry more of the build with a smaller external team. If not, an embedded team doubles as a training vehicle.
Be clear-eyed about the invoice, too. An embedded senior team is not the cheap option on a monthly statement. The fair comparison is against the fully loaded cost of hires you have not yet made, plus the cost of the quarters in which nothing ships. For what the build itself costs, ranges and drivers are in what custom AI development costs.
If you are weighing these paths against a specific project, that is a conversation we have most weeks. Our custom AI engagements run on the embedded model, and we will tell you plainly when hiring first is the better answer for your case.