Density Labs / AI pilot: in house vs outside team

Run the first AI build in house, or bring in a team?

Your engineers can do this. That is not the question. The question is whether the first one should be theirs, given that first builds carry the learning cost and your team is already committed to a roadmap.

In houseOutside team
Direct costLooks like zero. It is the roadmap work that does not ship.Visible and fixed
Time to first production useThree to six months typical, longer if it competes with feature workSix to twelve weeks
Learning curvePaid once, kept foreverPaid by the vendor, transferred only if you require it
Biggest riskIt becomes the thing everyone gets to next quarterYou end up dependent on someone who leaves
Who runs it after launchAlready clearThe thing to negotiate up front, not later
Best whenYou have slack capacity, a clear owner, and time to be wrong onceThere is a deadline, or the first build must be right

The cost that gets missed

In house reads as free because no invoice arrives. The real price is the roadmap item that slips, and the slip is usually larger than expected, because a first AI build has an unfamiliar shape: evaluation, prompt and retrieval iteration, and a review loop that is not like normal QA. Teams routinely estimate this like a normal feature and land at two to three times the estimate.

That is not an argument against in house. It is an argument for estimating it honestly, then deciding.

The risk that gets missed on the other side

Outside teams can leave you with a system nobody internally understands. This is a real and common outcome, and it is avoidable, but only if you make it a term of the engagement rather than a hope: your engineer in the code from week one, documentation as a deliverable, and a named internal owner before launch, not after.

If a vendor resists that, the resistance is the information.

A reasonable default

For a first build with a deadline, the shape that works most often is neither pure option: an outside senior engineer builds alongside one of yours, and the handover is a contract term with a date. You buy the speed and the pattern, your team keeps the system. The second build is then genuinely in house, and it is much cheaper than the first.

If there is no deadline and you have real capacity, build it yourself. The learning is worth more than the months.

Related reading: why 95% of AI pilots never reach production · what an AI diagnostic includes · a forward deployed engineer vs hiring