Build your AI agent in-house, or hire a partner? The real math
MIT's Project NANDA studied 300 public AI deployments, surveyed 350 employees, and interviewed 150 leaders — and found that only about 5% of enterprise AI pilots deliver measurable business results. The same research found externally sourced AI succeeds roughly twice as often as internal builds. That does not mean you should never build in-house. It means your first agent — the one that has to prove AI works for your company at all — is usually the wrong project to learn on.
What the research says about success rates
The MIT findings, reported by Fortune in 2025, draw a sharp line: AI solutions bought from specialized vendors reached deployment successfully about 67% of the time, while purely internal builds succeeded only about a third as often. The failure driver was not model quality but what the researchers call a learning gap — teams underestimating evaluation, integration, and workflow design, which is precisely the unglamorous work that separates a demo from a production system. A partner who has crossed that gap repeatedly carries the lessons into your project on day one.
The real cost of an in-house AI team
In Germany, a machine-learning engineer earns a median of €57,800 gross per year, and experienced profiles average around €74,800 (StepStone, 2026) — before employer contributions, recruiting fees, equipment, and cloud spend. A minimal credible agent team — two senior engineers plus part of a product owner — puts you at several hundred thousand euros a year of committed cost, and that meter starts months before the first line of code, because senior AI engineers are among the hardest hires in Europe. None of this is wasted if AI becomes your core competence. All of it is at risk if the first project stalls like the 95%.
The hidden cost: time to first value
The salary line understates the real gap. An in-house team must first be hired, then form, then learn agent-specific engineering — evaluation suites, tool-use guardrails, human-in-the-loop design, cost control at inference time — and every one of those lessons is paid for on your payroll, on your first project. A specialized partner amortizes those lessons across many builds. The practical difference is not code quality; it is whether your first agent reaches production in one quarter or is still a promising prototype a year later.
When building in-house is the right call
In-house wins when AI is strategically core rather than instrumental: the agent is your product, you have a roadmap of many agents, you already employ senior engineers who want to own the capability, or your governance genuinely requires everything inside your walls (though on-premises delivery by a partner usually solves that too). In those cases, the higher cost of the learning curve is an investment in a lasting competence — and worth it.
The third option: build with a partner, own everything
Build-versus-buy is a false binary. The model that captures the partner success rate without permanent dependency is a transfer build: a partner ships the agent to production, and you own every artifact — the repository, the infrastructure and accounts, the evaluation suite, the documentation, the runbooks. Your engineers shadow the build, support continues after launch, and if you later insource, you inherit a documented, tested system instead of starting from zero. You get the 67% path without signing up for a vendor relationship you cannot leave.
A decision rule that holds up
If this is your first production agent and it needs to prove value this year: partner. If it is your fifth agent and you have a hired, formed team: in-house. In between: partner build with your engineers embedded, then take over operation after handover. And if an off-the-shelf tool genuinely covers the use case, buy the tool — the MIT numbers favor it, and a partner who tells you that in the first call is one worth keeping.
Can our developers maintain a custom agent after handover?
Yes — if the delivery includes the evaluation suite, documentation, and runbooks, which is exactly what to demand from any vendor. A capable in-house web team can operate a well-documented agent; what they should not have to do is reverse-engineer an undocumented one. That is the test of a good build.
Isn't an agency more expensive than doing it ourselves?
Compare full costs, not the invoice: a year of team salaries plus months of hiring plus the documented risk that the first internal build stalls — against a fixed project price with a contractual delivery date. For a first agent, the fixed price is usually the cheaper number and always the more predictable one.
What about just buying an off-the-shelf AI tool instead?
If a tool fits your use case, buy it — the research says exactly that. Custom agents earn their cost when the work runs through your systems, your data, and your rules, where generic tools stop. We tell prospects in the first call which side of that line they are on.