In-house AI team vs. AI development agency: how to decide

In-house AI team or agency isn't a philosophy question — it's two numbers. How many weeks until you need a working result, and does this capability need to live inside your company permanently once the project ships, or is it a single, defined build? If the honest answer to the first is "this quarter" and the second is "just this one," hiring is the wrong tool before you've even priced it: specialized engineering roles are taking an average of 62 days to fill in the US in 2026 — the slowest of any function SHRM tracks — and that's before the new hire has written a line of code against your actual systems (SHRM, 2026 Recruiting Benchmarking: Attracting Critical Talent, 2026).

AI roles specifically are worse than that engineering-wide average, not better. ManpowerGroup's 2026 Global Talent Shortage Survey — 39,000 employers across 41 countries — found AI model and application development has become the single hardest capability to find anywhere in the world, ahead of every traditional engineering and IT category, with 72% of employers overall reporting they can't fill the roles they currently have open (ManpowerGroup, Global Talent Shortage Reaches Turning Point as AI Skills Claim Top Spot, 2026). François Lançon, ManpowerGroup's Regional President for Asia Pacific and the Middle East, framed the practical choice in exactly the terms this article is about: organizations "need to build, buy or borrow the skills to recruit 'AI enhanced candidates.'" Build is hiring in-house. Buy is an agency engagement. Borrow is a fractional or contract hire. The survey's own finding is that build is currently the slowest and least certain of the three — not the safe default it used to be when engineering talent was easier to find.

Hiring the person is also the easier half of "in-house." Getting them productive is a separate, ongoing cost most build decisions never price. Deloitte's State of AI in the Enterprise 2026 report — 3,235 business and IT leaders across 24 countries, fielded August–September 2025 — found workforce AI-skill readiness at just 20% "highly prepared," the lowest score of any readiness dimension the survey tracks and down 2 percentage points from the year before (Deloitte, The State of AI in the Enterprise — 2026 AI report, 2026). Even a company that fills the seat is, by its own data, more likely than not to have filled it with someone the organization doesn't yet consider fully prepared for the work — which means the 62-day hiring clock isn't when the capability actually starts, just when the ramp-up clock does.

The market has re-priced this trade-off once already, and it's worth being precise about who that shift actually applies to. McKinsey's November 2025 State of AI survey found the internal-build share of enterprise AI initiatives has risen to 47%, against 53% still bought from a vendor — up sharply from a year earlier, when the large majority of organizations relied on third-party software alone (McKinsey & Company, The state of AI in 2025: Agents, innovation, and transformation, November 2025). That move toward building internally is concentrated in organizations running enough AI initiatives, on a long enough horizon, that a permanent team's hiring cost amortizes across many projects instead of one. A company shipping its first production AI system is not that company yet. For a single, time-boxed build, the 47% figure is describing a different buyer's economics than yours.

Three questions surface which side of that line you're actually on, before you spend a hiring cycle finding out the hard way. First: will this capability be needed on more than one project over the next 12 months, or is it a single, defined build with a clear finish line? A recurring need justifies amortizing a hire's ramp-up cost; a one-off doesn't. Second: can the business case survive a 2–4 month hiring runway before month one of actual development even starts, or does the opportunity — or the budget cycle — expire before someone new could plausibly be productive? Third: has your existing engineering team shipped a production AI system before — with real evaluation, permissioning, and monitoring — or would this be genuinely new ground for them too? A team that's never built the guardrails around an agent that writes to a live system will spend the first project learning that discipline from scratch, in-house or not.

Answer "recurring, time-flexible, and we've done this before" to those three and building in-house is usually the right call — you're the organization McKinsey's 47% describes. Answer "one-off, time-pressured, and this is new ground" and an agency or a dedicated AI development team gets you a working result inside the hiring runway you'd otherwise spend just finding a candidate, with the evaluation and permissioning discipline already built rather than being invented under deadline pressure on your first attempt.

The middle path is real, and it's the one buyers skip most often: a dedicated external team that runs at agency speed but reports and ships the way an internal team would, without the 62-day search or the 20%-readiness gamble. We staffed exactly that shape of team on a multi-assistant AI platform — five specialists, six months, shipping domain-tuned assistants with full prompt observability from day one, not month four of onboarding. That timeline is what "borrow" looks like done properly: capability available immediately, evaluation and monitoring built in from the start because the team has done it before, and no permanent headcount commitment if the project turns out to be a one-off after all. A fractional CTO AI engagement covers the smaller version of the same borrow decision — vendor and architecture oversight without a full-time executive hire — when what's missing is judgment, not build capacity.

If you do decide to buy or borrow, the diligence doesn't disappear — it moves to a different question. Ask a prospective agency the same thing you'd ask about an internal hire's onboarding plan: how quickly can your team actually be productive against our systems, and what's already built versus what gets invented for the first time on our project? A vendor who has shipped the evaluation harness, the permission model, and the audit trail before — not just the model integration — is offering you the readiness Deloitte's data says most in-house hires don't yet have. That's a fair trade for the agency premium, and a specific, checkable claim you can ask them to back up with a reference client, not take on faith.

None of this argues that hiring in-house is the wrong answer generally — it argues for pricing the hiring clock honestly before assuming it's free. If the true cost of "build it ourselves" is a 62-day search followed by a ramp-up period against systems that person has never seen, and the true cost of "bring in a team" is a vendor premium against a team that's done this exact evaluation-and-permissioning work before, that's a real trade-off worth running the numbers on — not a values question about whether you trust outside help. An AI readiness assessment is built to answer the harder version of this question before either budget is committed: not just build vs. buy, but which specific parts of your first AI project are ready for either path today, and which need more groundwork regardless of who staffs it.