Automating a system with no API, without trusting what reads it
A deployment engineer's toolkit, built around a fictional company selling model API access. This page is one real run of it.
Screen automation with safe stops
The demo CRM has no API, so the robot works from screenshots. Before each action it reads the screen with OCR and confirms the app, the screen, and that this is the demo environment. If any of that is off, it stops and sends nothing.
Extraction with hard checks
Statements captured from the screen become structured data. Then plain code recomputes every line and total. It never corrects a figure: a statement passes everything or goes to a person.
Reviewed notes
Transcripts imported from Microsoft Teams and Granola formats become draft notes. Nothing reaches the CRM until a named person approves it, and then the robot types it, reads it back and saves it.
Revenue analytics
Standard metrics for a usage-based API business, from a seeded generator. Revenue here means what was billed: the greater of commit and usage, in the month shown.
What moved revenue each month
Committed revenue and everything above it
Token consumption by model tier
Usage value by model tier
Movement between model tiers
Commit conversations to have
Commits at risk at renewal
What is and isn't AI
| Step | How it works |
|---|---|
| Reading the screen | A local OCR engine (a small neural network, no cloud). Its output is treated as untrusted text. |
| Deciding where the robot is, and whether to act | Rules only. No model is ever asked whether it's safe to click. |
| Structuring a statement | |
| Checking the figures | Arithmetic only. |
| Drafting a note | |
| Checking a note's facts | Rules only: figures, dates, names and citations must match the transcript. |
| Approving a note | A person. |
| Revenue metrics | Arithmetic on seeded data. |