The short version
- Total cost (Year 1): in-house team €350K–€700K. Agency project €25K–€150K + €30K–€90K/year managed.
- Time to first production: in-house 6–12 months. Agency 6–12 weeks.
- In-house wins when: AI is core to your product, you have 12+ months runway, and you can actually hire (and retain) good ML engineers in 2026.
- Agency wins when: speed matters, the AI is one of many initiatives, and you don’t want to run a recruiting process.
- Most successful companies do both: hire one or two senior people for ownership, work with an agency to ship fast.
In 2026, if your business needs custom AI, you have two paths: hire ML/AI engineers in-house, or work with a specialist agency that ships AI systems for businesses. Both work. They have very different cost structures, timelines, risk profiles and outcomes — and the right answer depends on what you’re actually trying to do.
I’ve done this from both sides. I’ve been the in-house lead trying to get ML production-ready while the rest of the engineering org ships the actual product. I now run an agency that has shipped 50+ AI projects for businesses. The trade-offs below are what I’ve actually seen play out.
The real cost — not the “build it for free with our existing team” cost
The most common in-house pitch goes: “Our engineers are smart, they’ll figure it out, the only cost is API tokens.” That math misses three things:
- Salary: a senior AI/ML engineer in 2026 costs €120K–€220K base in Europe / North America, €100K–€180K in Singapore / HK / Australia. With benefits, equipment, manager overhead, true loaded cost is often 1.4× base.
- Hiring time: 4–8 months from posting to productive employee. That’s 4–8 months your problem isn’t getting solved.
- Project ramp: even a great hire takes 3–6 months to ship the first production AI feature. Domain knowledge, eval infrastructure, MLOps tooling, prompt iteration — all built from scratch.
Honest in-house numbers for a one-engineer Year 1:
| Cost | Amount (EUR) |
|---|---|
| Senior AI engineer (loaded) | €180,000 |
| Recruiting cost (agency / referral) | €20,000–€40,000 |
| Manager overhead (15% of an EM) | €25,000 |
| Tooling, infra, model APIs | €20,000–€60,000 |
| Lost months (no production AI for ~6 months) | opportunity cost |
| Year 1, single engineer | €245K–€305K |
And one engineer is a brittle bus-factor. Most serious in-house AI starts with two and adds a third. Year 1 cost for a small team: €350K–€700K, mostly with nothing in production yet.
Agency comparison:
- One project: €25K–€150K (estimate yours).
- In production in 6–12 weeks.
- Optional managed plan after launch: €249–€749/month for hosting, monitoring, and ongoing tweaks.
- Optional retainer for new features: €5K–€20K/month, scaled up or down.
Apples to apples on Year 1 results: an agency typically delivers 2–4 production AI systems in the time and budget that gets one in-house team to its first deployment.
When in-house wins
I’m not anti-in-house. There are clear cases where it’s the right call:
1. AI is core to your product
If you’re an AI-first company — the AI is the product, not a feature — you must own the talent. Iterating on a model that is your product happens daily, not quarterly. Outsourcing that is outsourcing your moat.
2. You have 12+ months and an experienced engineering culture
If you can wait, hire well, and invest in MLOps over 12 months, in-house compounds beautifully. Year 2 is much cheaper than Year 1 because the foundation is in place.
3. Your data is too sensitive for any third party
For some defence, intelligence, or critical-infrastructure clients, no agency engagement is on the table. In-house is the only option, and the cost is the cost.
4. You can actually hire
The ML talent market in 2026 is brutal. Big tech offers €400K+ packages. If you’re a non-tech company in a tier-2 city, you’ll lose every senior candidate to remote-first competitors. Be honest about whether you’ll find and keep good people.
When agency wins
1. Speed matters more than ownership
Most business AI initiatives fail not because the model was wrong but because the project took too long. By the time it ships, the business need has shifted. Agencies ship 4–5× faster because the foundation already exists; you’re paying for delivery, not for building the same MLOps wheels every other team builds.
2. AI is one of many initiatives, not the main bet
If your CTO is also responsible for the website rebuild, the migration to Postgres 17, the SOC 2 audit and three product features, asking them to also build AI competency is overload. Use the agency for the AI work; let the engineering org focus on the product.
3. You want a working result, not a hiring project
Hiring a senior ML engineer takes 4–8 months. Negotiating an agency engagement takes 2–4 weeks. If your CEO needs the customer-support AI deployed before next year’s board meeting, the math is decisive.
4. Your budget is project-shaped, not headcount-shaped
Capex / project budgets are easier to approve than new headcount in most companies. A €60K project with a fixed scope and end date is a different conversation than asking for a €300K/year headcount.
Estimate your custom AI project in 30 seconds
Three questions, an instant cost range and timeline based on real shipped projects. After 30 minutes on a discovery call you have a written fixed-price quote.
The hybrid that actually works
The pattern I see in the best-performing companies:
- Hire one senior AI/ML person — not to build everything, but to be the technical owner. They review the agency’s work, manage handover, run the eval harness, decide when to internalise vs outsource.
- Use an agency for the first 1–3 production systems. They ship fast, they bring patterns, your senior person learns from the implementation.
- Internalise gradually. As the senior owner gets comfortable, hire a second engineer, then a third. Move from agency-led to senior-led to team-led over 12–18 months.
- Keep the agency on retainer for spike work. New AI initiative, integration with a new system, urgent feature — the agency stays on call.
This is “build the team while shipping the product” instead of “build the team for 12 months, then ship the product.” Year 1 has working AI in production, the business gets the wins, and the in-house team grows on real work instead of greenfield setup.
Common objections, addressed
“The agency won’t understand our business.”
True for some. False for any agency that does discovery properly — we sit with your team, watch the work, and write the spec before code. After two weeks we know the workflow as well as anyone except the people doing it daily. Our process page walks through this.
“What about IP? Will we own it?”
You should. Look for an agency where the contract explicitly assigns IP to you. Avoid “you license the system from us” deals — that’s vendor lock-in dressed up as an agency engagement. AI Makers writes contracts where you own everything: code, prompts, evals, integrations.
“What if we want to take it in-house later?”
Plan for it. The right agency builds with handover in mind — clean code, written architecture, recorded walkthroughs, a runbook. We’ve had clients move work in-house after 6–12 months and the handover took a week. The wrong agency builds a system only they can maintain. Avoid them.
“Can we just use freelancers?”
Sometimes — for small, well-scoped tasks. The risk: a single freelancer means single bus factor, no eval infrastructure to inherit, no support if the project hits trouble. For one-off prototypes, freelancers are fine. For production systems your business depends on, the agency model gives you a team and accountability.
“What about offshore development at a tenth of the price?”
The numbers look great until the project ships and you discover the AI doesn’t handle real edge cases, the integrations are unreliable under load, and there’s no-one to talk to in your time zone. Offshore can work for the right kind of well-defined task. Custom AI is rarely that — the work is too specific, the iteration too tight, the stakes too high. Pay for senior people who’ve shipped before.
The 5-question fit test
If you’re trying to decide right now, answer these five:
- Is AI core to your product, or supporting your business? Core → in-house. Supporting → agency.
- Do you need it shipped in 3 months, or can you wait 12? 3 → agency. 12+ → in-house possible.
- Can you actually hire and keep senior ML engineers right now? Yes → in-house possible. Honestly no → agency.
- Is this one project, or a permanent capability? One project → agency. Permanent capability → in-house (or hybrid).
- Is your budget project-shaped or headcount-shaped? Project → agency. Headcount → in-house.
If you got 4–5 “agency” answers, the call is clear. 4–5 “in-house” answers, also clear. Mixed answers → hybrid.
Whichever path you pick
If you’re going in-house: don’t under-budget. €350K Year 1 for one engineer is real. Plan for the ramp.
If you’re going agency: pick one that has shipped your kind of work before, owns the IP transfer cleanly, and is willing to be replaced. Here’s the question list I’d use.
Either way, the worst outcome is the third path: the half-built in-house effort that’s neither finished nor abandoned, eating budget and morale for 18 months. Decide, commit, and ship.
If you want a real number for the agency path
The interactive cost estimator gives you a ballpark in 30 seconds. Or book a discovery call — I’ll tell you straight whether your situation is actually better suited to in-house and walk you through what that would look like.