Build vs. Buy vs. Managed

Three doors, one bill

Once you've built a recruitment AI agent, the choice in front of you isn't really "AI or no AI." The build is the easy part. The honest question is narrower: who runs this thing once it exists?

There are three doors.

• Buy an off-the-shelf product. There's a whole reckoning for why this one disappoints: generic isn't yours, and the tool that scores everyone's CVs the same way scores yours badly. We're not relitigating it here.
• Build and run it yourself. Stand up the code, own the deployment, carry the pager.
• Managed. Someone else designs, deploys and runs it; you keep the leash on the decisions and hand off the operations.

This chapter prices the last two honestly. No thumb on the scale. We model a managed option, add up what doing it yourself actually costs once you count the bits people quietly drop off the spreadsheet, and stand both next to the failure data. Then you decide.

The build is a weekend. The bill is for the years after.

The number people quote, and the number that bites

Let's start with the figure everyone fixates on, because it's the smallest one in the room: infrastructure.

A single .NET agent service on Cloud Run (AWS App Runner / Azure Container Apps) can scale to zero when idle, and that's a genuine architecture win, not a cost claim you can lean on. A real production deployment isn't the near-zero hobby case. The moment you keep a warm instance (min-instances ≥ 1, so a CV doesn't wait on a cold start), add a managed audit store (Cloud SQL), a queue (Pub/Sub) and log ingestion, the bill lands at roughly $120 a month at the low end, $300–$350 for a mid-size agency, and $500–$750+ under heavier load or HA. Scale-to-zero only reaches near-$0 with negligible traffic; under real volume it doesn't. Observability on Grafana Cloud sits between $0 and $50 a month at this size; you only reach Datadog's $80–$120-per-host territory if you choose to layer on its full product stack. The LLM calls are thicker than people expect once it's a real agentic loop: on a GPT-5-class model the agent makes roughly 2–4 calls per CV (~3,000 input and ~700 output tokens each), which works out to about $20–$75 per 1,000 CVs, a few cents per CV. A budget model (GPT-4o-mini) is 10–15× cheaper, around $2–$4 per 1,000. At agency volume, say 10,000 CVs a month, that's a few hundred dollars a month on a frontier model (~$250–$700) versus about $25 a month on a mini one. (OpenAI list prices: GPT-5 ≈ $1.25 in / $10 out per 1M tokens; GPT-4o-mini $0.15 in / $0.60 out per 1M. Batch mode roughly halves it.)

Add it all up and the technology to run three agents for a single agency lands in the low hundreds a month. Real money, but still a rounding error next to the row that comes next. That's the part vendors wave at to make DIY look free.

It isn't the part that bites.

Infra, tooling and tokens are the cheap parts. People are the expensive part. They always were.