Skip to content

Writing

In-house AI team vs agency: the arithmetic, including when we lose

Real salary numbers, real agency numbers, and the crossover point where hiring wins. Written by an agency, which should tell you how the last section goes.

By Yash Mittal4 min read

Hiring an in-house AI engineer costs roughly $150,000–220,000 a year fully loaded in the US, or ₹25–60 lakh in India, and takes three to six months to fill and ramp. An agency costs $3,000–25,000 per project with no ramp-up and no hiring risk. Agencies win on the first one to three AI features, where the value is patterns you don't have yet. In-house wins once AI is a permanent part of the roadmap — typically past two ongoing workstreams — because context compounds and salary stops being the dominant cost.

This is written by an agency, so read the last two sections first if you're pressed for time. They're the ones that argue against hiring us.

What an in-house AI engineer actually costs

The salary is the number people quote and about two-thirds of the real one.

Line itemUS / Western EuropeIndia
Base salary$140,000 – $190,000₹18 – 40 lakh
Employer taxes, benefits, equipment+25 – 35%+10 – 15%
Recruitment (agency or internal)15 – 25% of first-year salary8 – 15%
Fully loaded, year one$190,000 – $260,000₹25 – 60 lakh
Ongoing years$175,000 – $230,000₹22 – 50 lakh
Fully loaded annual cost of one mid-to-senior engineer with production LLM experience, mid-2026.

Then the parts that don't appear on a budget line. Three to six months to hire, because engineers who have actually shipped LLM features to production are genuinely scarce and know it. One to three months to ramp on your codebase. Roughly four hours a week of somebody senior's time managing them. And the risk that the first hire doesn't work out, which costs the whole search again plus the months.

What an agency actually costs

Project pricing for AI integration work at a small studio in 2026, ours included:

EngagementCostElapsed time
Discovery / feasibility$600 – $1,8003–5 days
One feature into an existing product$3,000 – $8,0003–6 weeks
RAG system over one corpus$4,000 – $10,0004–8 weeks
Full AI-first product or MVP$8,000 – $25,0008–16 weeks
Ongoing retainer$900 – $6,000 / monthContinuous

No recruitment, no ramp, no notice period, and it stops when you stop. The costs that don't appear here are real too: your team's time explaining the domain, the handover, and the fact that when we leave, the deepest understanding of that system leaves with us unless the documentation was taken seriously.

The crossover

Compare over eighteen months rather than a quarter, because that's the window where hiring costs stop being front-loaded.

ScenarioIn-houseAgency
One AI feature, then done~$290,000 and you're still paying~$6,000
Three features over 18 months~$290,000~$18,000 + retainer ≈ $63,000
Continuous AI roadmap, 2+ workstreams~$290,000 for one engineer$150,000+ and you're renting context
Eighteen-month total, US salary basis. The middle row is where most teams actually sit.

The pattern is consistent across every version of this arithmetic we've run: agencies win decisively at low volume and lose at high volume, and the crossover sits somewhere around a permanent two-workstream commitment. Below it you're buying patterns. Above it you're renting an employee at a premium, which is a bad trade for both sides.

The third option most teams should take

Framing it as a binary is the actual mistake. The sequence that works:

  1. 01Validate with an agency. One feature, six weeks, a real number attached. You find out whether AI moves anything in your product before committing a headcount to the belief that it does.
  2. 02Insist the handover is real. Documentation, an eval suite your team can run, and a walkthrough with your engineers. This is the deliverable that decides whether step three is cheap or expensive.
  3. 03Hire once the roadmap is proven. Now you can interview properly, because you have a working system to ask candidates about and someone internal who understands it.
  4. 04Keep a thin retainer. For model migrations, eval upkeep and the occasional second opinion, while your hire owns the day-to-day.

This is a worse outcome for our revenue than a permanent retainer, and it's what we recommend to most teams who ask.

When you should not hire an agency

Plainly, so you can check yourself against it:

  • You already have engineers who've shipped production LLM systems. You're paying a premium for patterns you have. Give them the roadmap space instead — that's cheaper and produces a better system.
  • AI is your product, not a feature of it. If the model pipeline *is* the company, that capability cannot live outside it. Build the team.
  • Your data can't leave and can't be described. Some regulated environments make external help slower than doing it internally, even badly.
  • You need more than ~80 engineering hours a month, indefinitely. That's a full-time job. Retaining it costs more than employing it, and we'll say so on the call.
  • The real problem isn't AI. A distressing share of AI enquiries are data problems, process problems or a search box that was never fixed. An agency that takes the brief without checking is selling you something.

When an agency genuinely is the better call

  • Nobody internal has done this before. The first production LLM system teaches expensive lessons. Buying them once is cheaper than learning them on your own roadmap.
  • You need an answer this quarter. Hiring takes three to six months before anyone writes code. If the window is now, that settles it.
  • You're not sure it's worth doing. Six weeks and $6,000 to find out beats a $260,000 annual commitment to the hypothesis.
  • The work is spiky. One intense build, then maintenance. That shape fits a project plus a retainer and fits a full-time hire badly.

The honest summary

If you'll build one to three AI features over the next eighteen months, an agency is cheaper by an order of magnitude and faster by a quarter. If AI is becoming permanent infrastructure in your product, hire — and use an agency to de-risk the first build and to train the person you hire. Anyone who tells you the answer is always their own model of working hasn't done the arithmetic, or has and would rather you didn't.

If you want the version of this conversation applied to your actual situation, that's what our scoping calls are for — including the ones that end with us telling you to hire someone.

This is what we do. See how we scope AI integration.

See how we scope AI integration

FAQ

Related questions

When does in-house AI development make more sense than an agency?

Once AI is a permanent part of your roadmap rather than a project — practically, once you have two or more ongoing AI workstreams, or need more than about 80 engineering hours a month indefinitely. Below that, an agency is cheaper and faster. Above it, you're paying agency rates for what is effectively an employee, and the person who lives in your codebase daily will build better systems than one who visits.

How much does a fractional AI engineer cost?

Fractional engineers typically charge $100–200 an hour, or $4,000–12,000 a month for one to three days a week. That sits between a studio retainer and a full-time hire. The trade is that you get one person's skill set rather than a team's range, but more continuity than a project engagement — it suits teams who need steady capacity in one area rather than varied capacity across several.

How long does it take to hire an AI engineer?

Three to six months to fill, plus one to three months to ramp on your codebase. Engineers with genuine production LLM experience are scarce and usually employed. If your AI roadmap needs to start this quarter, hiring cannot be the path that starts it — though it can be the path that continues it.

How do we avoid depending on an agency permanently?

Make the handover a contractual deliverable, not a courtesy. You want documentation written for the engineer who inherits it, an eval suite your team can run without us, everything deployed on your own accounts, and a walkthrough session with your engineers. If a studio resists any of those, that's the answer to whether the dependency is accidental.

Keep reading

More on this

Article

What a software maintenance retainer actually costs

Most agencies won't put a number on a retainer. Here are ours, the industry ranges we checked them against, and the arithmetic for deciding when hiring someone is cheaper.

Article

Why your AI proof of concept never shipped

The demo worked. Four months later it's still a demo. Here are the six things that actually block the gap, in the order they usually bite.

Next step

Want this applied to your situation?

Articles generalise. A 45-minute call doesn't — tell us what you're actually dealing with and we'll be specific.

Start a projectRead more articles