loading
Back to the blog

[ Thesis ]

Buy vs. build an in-product AI agent: 100 people, one year — or one script tag

Databricks put ~100 people on an internal agent for a year. A minimum viable team costs $1.3–1.5M before it ships anything. The honest TCO of build vs. buy — and when building actually wins.

rtrvr.ai Team·July 4, 2026·3 min read

Buy vs. build an in-product AI agent: 100 people, one year — or one script tag

Every SaaS roadmap now has the same line item: "an AI agent that can operate the product." Most companies should not build it. Here's the math, with the cases where building does win.

What building actually costs

An agent that reliably operates a real web app is a hard, full-time system — not a wrapper around an LLM API. The evidence is public:

  • Databricks put roughly 100 people on their internal agent for about a year.
  • Instacart spent months and a dedicated team to ship a ChatGPT app.
  • A Forward-Deployed or Applied-AI engineer runs ~$300–400K/yr fully loaded; a minimum viable team of four is ~$1.3–1.5M in year one — before it ships.

Then the part nobody budgets: the models change every quarter, and your agent gets rebuilt against them. Forever. It's not a project; it's a treadmill.

The rule that settles it

Build the agent that is your moat. Rent the one that operates your UI. Driving your own interface is plumbing — undifferentiated, high-maintenance, and exactly the layer that rots as models churn. Your customers don't buy you because your copilot can click your buttons.

Build in-houseRover
Time to live~12 monthsSame day — one script tag
Year-one cost$1.3M–$3M+A fraction; pilots a VP can approve
Model churnYour team, every quarterOurs — it's the whole job
ReliabilityUnknown until builtPublished: #1 on Web Bench
Your engineersConsumed by plumbingBuilding your actual moat

The reliability you'd be renting

This is the part in-house teams spend the year discovering is hard:

Rover on the Halluminate Web Bench: 81.4% task success vs 40–66% for other agents; 3.39% infrastructure errors vs 20–30% for CDP-based tools

81.4% — state of the art, ahead of OpenAI's and Anthropic's agents — at $0.12 a task on Gemini Flash. That's what a specialist gets you on day one, and keeps current as the models change under everyone's feet.

When building is the right call

Honesty cuts both ways. Build in-house when the agent is the product (you're an agent company), when your workflows can't leave your walls even under a private deployment, or when you already run a staffed applied-AI platform team with room on its roadmap. Otherwise the year and the $1.3M+ buy you a worse version of something you could have turned on the same day.

Frequently asked questions

How long does an in-house product agent really take?

Plan on ~12 months to something trustworthy in production — Databricks-scale teams took about that with far more resources than most. The demo takes a weekend; the reliability takes the year.

What does the team cost?

$300–400K per fully-loaded applied-AI engineer; $1.3–1.5M/yr for a minimum four-person team; $2.5–3.2M for eight. Recruiting them is its own project — these are the scarcest engineers in the market.

What would we buy instead?

An agent that already operates real products at published, state-of-the-art reliability — live the same day, judged on one metric you choose. It also covers your marketing site: see Rover vs. Navattic + Drift. And your future agent traffic: see the agentic web, measured.


Spend your engineers where they compound. See Rover on your product →

More from the blog

On this page

  • What building actually costs
  • The rule that settles it
  • The reliability you'd be renting
  • When building is the right call
  • Frequently asked questions
  • How long does an in-house product agent really take?
  • What does the team cost?
  • What would we buy instead?