The short version
One mid-senior engineer in the US costs ~$120,000 a year fully loaded. Salary plus benefits, taxes, equipment, the all-in number your finance team will recognize.
At today's Anthropic pricing — roughly $3 per million input tokens for Sonnet, $15 per million output, with prompt caching dropping input to ~$0.30 — that same $120K buys somewhere in the band of 20-30 billion tokens a year, depending on cache hit rates and model mix.
At our actual burn rate of 3-10 billion tokens a month — which sounds insane until you do the math — we're spending the cost of one to four engineers on AI that does the work of a small team running 24/7. Not a flex. Actual ratio.
What we're buying, in plain English
This isn't ChatGPT-for-everyone. It's not a "let's give Sales a copilot" line item. It's an operating layer that runs between the tools the company already pays for.
Concrete: every weekday morning at 7:30 AM, an agent reads our calendar, the overnight Slack signal, HubSpot deal motion, and posts a 250-word canvas to my DM. The team gets the same brief in their channels. Total cost: ~$40/month. Replaces 20 minutes of triage per operator per morning.
Multiply across the team. The morning briefing alone, deployed to 8 operators, saves ~50 hours of human triage per week. At a $150/hour operator rate, that's $30K/month in opportunity cost reclaimed, for a $40 line item. The ROI isn't 10x. It isn't 100x. It's the kind of ratio that breaks spreadsheets.
The receipts
| Workflow | Monthly cost | Human equivalent | Ratio |
|---|---|---|---|
| Morning briefing × 8 ops | ~$40 | ~50 hr/week of triage | ~750x |
| Deal post-mortems (per quarter) | ~$15 | ~12 hr of sales-leadership review | ~120x |
| Codex 24/7 bug watcher | ~$200 | ~0.25 of an on-call eng | ~50x |
| Friday weekly synthesis | ~$25 | ~3 hr of operator time / week | ~150x |
None of these numbers come from a vendor deck. They come from running this stack across a portfolio for 18 months. The bills are real. The hours saved are observable in calendar archaeology.
The two questions a CFO actually asks
"Why not just give the team ChatGPT Plus?"
ChatGPT Plus is $20 per seat per month. It's a chat window. It doesn't run on a schedule. It doesn't talk to our CRM. It doesn't remember what we decided last Tuesday. Fifty seats of ChatGPT Plus is $1,000/month for fifty operators each individually re-typing the morning briefing context every day. The AI operating layer is one budget line that delivers a finished canvas to all of them, every morning, before they open Slack. Same cost, 50x the output, zero training.
"Why both Codex AND Claude Code?"
They're not competing — they're shifts. Codex runs against Sentry and GitHub 24/7, catching regressions and drafting fixes overnight. Claude Code is the day driver where humans ship features. Together they cost less than half an engineer's salary and they cover both shifts. Cutting one to save $200/month forfeits coverage that costs $60K to replace with a human.
The verdict
Every dollar in this category should clear two bars before it ships: (1) measurable hours reclaimed, and (2) a kill criterion at 60 days. We don't do open-ended AI spend. Every line item has a "what stops this" rule. That's why our footprint is operator-grade, not hype-grade.
If the answer to "is this worth it" isn't obvious from the receipts above, we shouldn't be spending. But it is. Approve the line.
Excerpted and condensed from Vlad's Ultimate AI Dive Deep — Vlad Podoliako's field manual for operators. Specifically: Ch 2 — The Five-Tool Stack, Ch 29 — Why Is My Bill So High?, and Ch 35 — Codex or Claude Code — or Both?. If your finance person wants to argue the numbers, point them at those.