AI agency or in-house AI engineer? A decision guide

The short answer

Hire in-house when AI is the product and you will build it for years; hire an agency when you need a first system in production within months and want a written plan before committing a salary. Most teams end up with both: the agency ships and documents the first system, an in-house engineer takes it over.

We are an agency, so read this with that in mind. We have also handed over every system we built to the client's own people, with the code in their repositories and the plan in their wiki, and we would rather tell you to hire than take a project that should have been a hire. The guide below is the conversation we have on most first calls.

What you are deciding

The decision has four parts: how soon you need a working system, how long the AI work will go on, who owns the knowledge afterwards, and how much you can afford to learn by trial. An in-house engineer is a long-term asset with a long lead time. An agency is a short lead time with a cost per project.

Most people frame it as a cost question first. Cost matters, and a senior AI engineer's salary in your market is a public number you can look up; so is the starting price of a scoped build. The comparison is less useful than it looks, because the two buy different things. A salary buys time and attention, including the time to learn. A project price buys a defined system, and the learning has already happened on someone else's projects.

Time to the first system in production

An agency that has shipped the kind of system you need will have it in production sooner than a new hire, because the hire has to be found, start, learn your codebase and then learn the problem. Recruiting a senior AI engineer alone takes months in most markets, and the first months in the seat go to your systems.

On our side the clock starts at the written plan. The enterprise knowledge platform we took from proof of concept to production ran six months and 1,145 hours, and the proof of concept stayed live throughout. The AI writing assistant, a single in-place rewrite feature inside an existing editor, took 162 hours. Those numbers are from our invoices, and they are what we point to when someone asks how long a first system takes.

Ownership and the knowledge that stays behind

The honest risk of an agency is that the knowledge walks out at the end. The fix is contractual and practical: code in your repositories from the first commit, a plan you can forward, and a handover that is a deliverable. Our standard agreement assigns the IP to you on payment, and nothing is subcontracted without telling you.

An in-house engineer carries the knowledge in their head, which is wonderful until they leave. The fix is the same: documentation and a codebase other people can read. We have inherited systems in both states, and the ones written by a careful agency with a handover plan have been easier to pick up than the ones built by a brilliant solo engineer who left no notes. The variable is the discipline, not the employment status.

Breadth and depth

An AI product needs more than a model call. The knowledge platform needed multi-tenant row-level security, background document conversion, prompt caching, a context budget, streaming chat and a human approval flow: database, infrastructure, API and front-end work around the AI. A single hire is rarely strong at all of it; an agency team is several specialists on one engagement.

The in-house advantage is depth on your problem. An engineer who has lived with your data for a year knows things no agency will learn in a discovery sprint. For a product whose edge is a model trained on or tuned to your own data, that depth is the business, and it belongs in-house.

The hybrid most teams end up with

The arrangement we see most often is an agency for the first system and the hard parts, an in-house engineer for ownership and the daily work, and a monthly retainer that shrinks as the internal team grows. The system reaches production before the hire is found, and the hire starts from a working codebase and a written plan.

The order matters. Agency first, then hire, works because the hire inherits something real and the agency's handover is written for them. Hire first, then agency, usually means the agency is called in when the first attempt has stalled, and the first job is an audit.

The decision table

The table is the one we sketch on most first calls. Read your own situation down each row before you read our column.

Questions we get about this

Is a freelancer a third option? Yes, and a good one for a defined feature with one skill set. The AI writing assistant was that kind of job. For a system that needs database, infrastructure and AI work together, a single freelancer is an in-house engineer without the long-term commitment, with the same breadth problem.

What does an agency cost, concretely? Ours publishes its starting prices: a discovery sprint from $2,500 for the written plan, fixed-scope builds from $6,000 priced against that plan, and an ongoing team from $45 per hour. The plan is yours whether or not you continue.

How do we avoid being locked in to the agency? Code in your repositories from day one, a plan you can forward, an IP clause that assigns on payment, and a handover deliverable. Ask for all four before signing anything, with us or anyone.

When would you tell us to hire instead? When the AI work is the product and will not end, when the edge is in your own data, or when the problem needs someone living with it daily for a year. We say so on the call.

What we would do on your project

We start with a free architecture call about what you are building and who you have. If it should be a hire, we will say so and tell you what to look for. If it should be a build, the discovery sprint, from $2,500, produces the written plan an in-house engineer could execute, with or without us. Book a free architecture call to have the conversation. The two builds quoted above are the enterprise knowledge platform and the AI writing assistant, and the AI agent development service is where most of these projects start.