Hiring an AI agency vs building in-house: the actual math

Hire in-house when your roadmap has years of continuous AI work ahead of it. Hire an agency, or buy a fixed-scope build instead, when you have a finite list of automations and then an operating mode to run them in. Most small and mid-size businesses are the second case, though few vendors will say so out loud, because a finite list caps what they can bill you once it is built.

We sell against in-house building, so read this knowing the bias. But we also ARE an in-house team: everything we sell was built in-house for our own companies first, before we ever sold it to anyone else. That seat gives us a clearer view of the trade than most vendors will give you, because we have paid both prices ourselves and watched what each one actually costs once the invoice arrives.

The in-house math

A capable AI engineer costs $180k to $250k fully loaded in most Western markets in 2026, and the scarce skill is not model calls, it is systems judgment: what to automate, what to keep human, how to make failure loud instead of silent. Ramp time to your business context is typically months before that hire ships anything alone.

If your roadmap contains years of continuous AI work, hire. The math supports it, and nothing beats owned context built up over a long run by one person who was there for every decision, every failed experiment, and every fix that followed it.

But most SMBs do not have years of continuous AI work stacked up. They have five to fifteen automations worth building, then an operating mode: watch the dashboards, fix what breaks, add the next thing when it becomes obvious rather than when a roadmap meeting says it is due.

Hiring a full-time salary against a finite build list is how companies end up with an expensive engineer doing maintenance a part-time arrangement would cover for a fraction of the cost, sitting in meetings that exist to justify the headcount rather than ship the next system. The salary does not go away once the build list is finished. The work does.

MIT NANDA’s GenAI Divide report, published July 2025 from a review of over 300 public AI initiatives and interviews with 52 organizations, found that external partnerships saw twice the success rate of internal builds, the widest single gap the researchers measured. Their own explanation: internal tools tend to stay fragile and dependent on whichever engineer built them, while external builds run through a delivery process designed to survive that engineer leaving the company entirely.

The agency math

AI agencies price a finite build list at $3k to $15k per system, or wrap it in a $4k to $10k monthly retainer. Both structures reward the vendor for slowness: a system that ships faster earns less, and a system kept on the agency’s own platform means leaving costs a full rebuild rather than an export.

Ask one question before signing anything: “if we part ways in month six, what keeps running?” If the honest answer is “not much,” you are renting, not buying, no matter what the contract calls itself.

Menlo Ventures’ 2025 State of Generative AI in the Enterprise report, surveying nearly 500 US enterprise decision-makers, found 76% of enterprise AI use cases were purchased rather than built internally in 2025, up from 53% the year before. The market has largely made the build-vs-buy call already, at scale, across industries that have nothing else in common. What is still unresolved for most buyers is what “bought” should mean: a retainer you keep paying indefinitely for continued access, or an asset you own outright the day the invoice clears.

Not every build belongs off the shelf, either. When the workflow is generic enough that a vendor has already solved it for a thousand companies like yours, buying SaaS beats both hiring and commissioning a custom build, agency or otherwise. The agency question only applies once the job is specific enough that no monthly subscription actually fits it.

The hybrid that actually works

The hybrid is to buy outcomes, own infrastructure: fixed-scope builds delivered onto accounts you control, documented so your existing team can operate them without calling the vendor back for every small change. It gives you agency speed without platform hostage-taking.

Every completed build makes an eventual in-house hire more productive rather than less necessary, because the infrastructure and the institutional memory already exist by the time that hire starts on day one.

That structure is what our AI Workforce Sprint productizes, and what a scoped four-to-six-week build looks like in practice: one system, from $9,500 fixed, on your own GitHub, Supabase, and hosting, with docs and a recorded handover instead of a login you never fully control.

McKinsey’s State of AI in 2026 survey, published August 2026, found 32% of organizations had already decided against buying at least one software product because agentic coding tools let them build the equivalent functionality in-house. That figure cuts the other way for narrow, code-shaped work, and it is a real shift worth naming rather than ignoring because it complicates a tidy argument.

But it does not stretch to a full production system with auth, monitoring, and a handover a new hire could follow without a week of guessing. Writing one internal script is not the same job as shipping a system someone else will maintain after you leave the room. That is a different build entirely, and the gap between the two is exactly where a fixed-scope sprint sits, closer to the agency side of the ledger than the weekend-project side.

Agency vs in-house vs fixed-scope sprint, side by side

Agency, in-house hire and fixed-scope sprint trade cost, speed, and ownership against each other in different amounts, and no single row of the comparison tells the whole story on its own: a cheap system you do not own outright is not actually cheap once you count what rebuilding it elsewhere would cost.

	In-house hire	Agency retainer	Fixed-scope sprint
Typical cost	$180k to $250k a year, loaded	$4k to $10k a month, or $3k to $15k per system	From $9,500, once, per system
Time to first system live	Months of ramp before independent output	Contract-dependent, often a quarter or more	4 to 6 weeks
Who owns the accounts	You, by default	Often the agency’s own platform; confirm before signing	Your GitHub, Supabase, and hosting from day one
What happens if you stop paying	Nothing changes; the hire is yours	Systems may stop running or need rebuilding elsewhere	Everything keeps running; you already own it

A hire is the right answer for continuous work spread across years, where the ramp time pays for itself many times over. A retainer is the right answer only if you value standing availability over a shipped, ownable deliverable. A fixed-scope build is the right answer for a defined list of systems you want to own outright, one at a time, without a headcount decision attached to each one.

Whichever way you go

Whichever way you buy AI work, insist on three things: a written scope with a definition of done, delivery to infrastructure you own, and real numbers instead of demos. Demos are optimized to look impressive in a room. Production systems are judged by what they do to a real number after the room has emptied out.

Our favorite example of why the third thing matters: our ad platform reported a 4.87x return while the reconciled ledger said 0.33 to 0.50. Systems that surface the true number beat systems that look impressive, every time, and that gap is exactly what a written scope and owned infrastructure protect you from missing until it is expensive to fix.

None of these three paths is wrong on its own. They fail when the label on the invoice does not match what gets delivered: a “partnership” that is really a rental, a “custom build” that is really a wrapper around someone else’s API. Name the real category before you sign, then judge it against the numbers above instead of the pitch.

Start async: describe your build list at /start. A human reads every submission and replies in writing within one business day. No call gets scheduled, because none is needed to read a form.