Agency vs in-house AI team
Two ways to get AI engineering done. One question decides which, and it is not cost.
The short answer
Build an in-house AI team when the work is permanent and central to what you sell; use an agency when you need the capability sooner than a hire can start, or when the work has an end date.
Most teams argue about cost. Cost is the wrong axis, because the two options fail in different ways. An in-house team fails slowly, through a hiring process that takes longer than the opportunity. An agency fails at the end, when the engagement closes and nobody left understands the system.
Both failures are avoidable, and knowing which one you are exposed to is the whole decision.
Side by side
Compared on the dimensions that change the decision, not on a feature count.
| Dimension | In-house AI team | Agency or embedded engineers |
|---|---|---|
| Time to a working engineer | A search, a notice period, then ramp-up | Weeks, and you interview before committing |
| Where the knowledge ends up | Inside your company, permanently | Inside your repo and docs, if you insist on it |
| Cost shape | Fixed. Salaries continue between projects | Variable. Stops when the work stops |
| Range of skills | What your hires happen to know | Whoever has already shipped that specific thing |
| Who reviews the code | Your team | Your team, if the engineers embed properly |
| Main failure mode | Hiring takes longer than the opportunity lasts | The engagement ends and nobody owns the system |
| Best when | AI is permanent and core to the product | The work is real but bounded, or urgent |
In-house AI team
Strengths
- The knowledge compounds inside your company and stays there.
- Employees care about the roadmap in a way contractors do not.
- Context accumulates. Year two is far more productive than year one.
- No renegotiation every time the scope moves.
Trade-offs
- Senior AI engineers are hard to find and harder to assess if nobody on your side has done the work.
- You pay for the capability between projects, not only during them.
- One or two hires cannot cover retrieval, agents, infrastructure and evaluation.
- A single person leaving can take the only understanding of a live system with them.
Agency or embedded engineers
Strengths
- People who have already shipped the specific thing you are building.
- Capacity that stops when the work stops.
- You can meet them and decline before anyone starts.
- Breadth on demand: a retrieval specialist for six weeks does not need to be a permanent role.
Trade-offs
- The knowledge leaves unless you require documentation and reviews in your own process.
- Priorities are yours to set, and a vendor who is not embedded will drift from them.
- Quality varies enormously, and the sales conversation is a poor way to tell.
- If AI is your core product, outsourcing it permanently is a strategic mistake.
- Permanent?Core to the product, indefinitely.
- Urgent?Needed before a hire could start.
- Bounded?Real work, but it ends.
- Assessable?Can anyone here judge the candidate.
- DecideHire, borrow, or do both.
The fourth question is the one people skip. If nobody in your company can tell a strong AI engineer from a confident one, hiring first is expensive guesswork.
Which one fits you
Four questions about your situation, not your preference.
Will this AI work still be running in three years?
Could someone on your team interview an AI engineer today and tell if they are good?
When do you need working software?
What happens to the system after launch?
Every outcome
- Hire in-house
- The work is permanent and central to what you sell, and you can already judge the candidates. Build the team and keep the knowledge.
- Embed outside engineers
- You need the capability sooner than a hire could start, or the work is real but has an end. Interview them, put them in your repo, keep your review process.
- Do both, in that order
- Bring in senior engineers now and hire alongside them. Your first hire learns on a live system with someone experienced next to them, and the handover is written into the engagement from the start.
How to choose
We sell the second option, so read the first rule with that in mind. It is still the right rule, and we have told people to go and hire instead.
- 01Hire in-house if AI is what your product is, and someone on your side can assess the candidates.
- 02Bring in outside engineers if you need the capability before a hire could realistically start.
- 03Do both if you intend to own it eventually: outside engineers now, your hire learning beside them, handover written into the contract.
- 04Do neither if you have not yet found a problem worth solving with AI. A diagnostic costs less than a team.
Systems built this way
Questions, answered
01Can you use both an agency and an in-house team?
Yes, and it is the most common good answer. Outside engineers cover the work now while you hire, and your first employee learns on a live system instead of a greenfield one. Write the handover into the engagement at the start, not when it is ending.
02Is an in-house AI team cheaper?
Cheaper is the wrong comparison, because the two costs behave differently. Salaries continue whether or not there is AI work that quarter, while an engagement stops when the work stops. Decide on permanence first and the cost shape follows.
03What is the biggest risk with an agency?
That the knowledge walks out with them. Three requirements remove most of it: the code lives in your repository from the first commit, your engineers review the pull requests, and the evaluation suite and runbook are deliverables rather than favours.
04How do we assess an AI engineer if nobody here has done the job?
Ask what happens when the model their system runs on is deprecated. A strong answer names an evaluation suite with real cases and a pass mark agreed in advance. A weak answer talks about testing in general terms, and you do not need to be an AI engineer to hear the difference.
05How long until an in-house hire is productive?
Longer than most plans assume, because the search, the notice period and the ramp-up all stack. That gap is the real argument for bringing people in first, and it disappears entirely if you have no fixed date.
