You cannot run a production system on a salary figure off a job board. The salary is the part you see; the loaded cost is the part you pay.
The US Bureau of Labor Statistics' Employer Costs for Employee Compensation data puts benefits at 29.9% of total compensation, a wage multiplier of roughly 1.43× before you've bought a single laptop or paid for a single training course. Apply that to a mid-level software engineer and the fully-loaded cost lands at about $10,000–$15,000 a month in the US.
Hold that number, because it's the spine of the whole decision.
Here's where the DIY sum quietly falls apart.
A production agent that touches your ATS and sends CVs to clients can't only work during office hours. APIs drift at 2am, models get deprecated on the vendor's schedule, a queue backs up on a Saturday. Someone has to be on the hook when it breaks.
The tooling for this is trivial: PagerDuty is $21–$41 per user per month. The human is not. On-call pay runs $500–$1,200 per engineer per month on top of salary, and here is the load-bearing point: sustainable 24/7 coverage needs roughly four to six engineers so that no one person is permanently tethered to a phone.
A solo DIY build has no sustainable on-call. One engineer covering a production system around the clock isn't a rota; it's a resignation letter waiting to be written. So your real choices are: accept that the system is unmonitored outside working hours (and discover the failures from your client), or staff a rota you were never going to staff for three small agents.
"No sustainable solo on-call" isn't a caveat. It's the whole DIY problem in five words.
And on-call is just the alarm. Behind it sits the work of actually running the system: monitoring dashboards someone has to read, the guardrail gateway someone has to keep current, the evals someone has to run, the API change someone has to absorb when ~15% of library updates break backwards compatibility. None of that is a one-off. It's a standing cost that never ends.
Put the honest version of each option next to the other. We'll model the managed option at ~$2,500 a month, and to be precise about what that number is: it's the midpoint of the prevailing $500–$5,000/month range for SMB managed-automation services (Digital Agency Network 2026; Latenode 2025; Arsum 2025; SalemWise 2025). It is a market assumption you can adopt for the maths, not a price from us.
| DIY, run in-house | Managed | |
|---|---|---|
| Infra (Cloud Run, production) | ~$120 low / ~$300–$350 typical / $500–$750+ heavy | included |
| Observability | ~$0–$50/mo | included |
| LLM API | ~$20–$75 / 1,000 CVs (GPT-5-class); ~$2–$4 (mini) | included |
| On-call tooling | $21–$41/user/mo | included |
| Engineer (fully loaded, BLS 1.43×) | ~$10,000–$15,000/mo | included |
| Sustainable on-call | needs ~4–6 engineers; solo = unsustainable | covered by the team |
| Who reads the dashboards? | you, or no one | them |
The cheap rows are a rounding error. The expensive row is one fully-loaded engineer, and that single line item is already four to six times a $2,500 managed retainer, before you've solved on-call at all.
Look at where the weight sits. Running it yourself doesn't cost what the infra costs. It costs what a person costs, every month, forever. And one person can't actually cover it.
Download the full PDF for free?