Internal build Back office 2026

Three automations built to be safe to leave running

Invoice extraction, lead triage and client reporting, each with a deterministic gate between the model and anything irreversible.

Running, internal

A diagram showing a model output passing through a deterministic gate before any irreversible action, with a budget check before each call and an append only audit trail.
3
Automations on one shared kernel
0
API keys needed to run the demos
Every demo replays a recorded run.
Every call
Budget reserved before it is made
Append only
Audit trail

The pitch is not that AI can read an invoice

Anything can read an invoice. The question is what happens on the day it reads one wrong, at two in the morning, with nobody watching.

Invoice extraction is gated on arithmetic the model cannot override: if the line items do not sum to the total, it does not quietly pick one. Lead triage scores leads with a model but routes them deterministically, and holds any drafted reply for a human. Report generation refuses to render if the narrative contains a number that is not in the computed facts.

Four controls that make unattended running reasonable

Budget is reserved before each provider call rather than counted afterwards, so a loop cannot run up a bill before anyone notices. Everything that happens is written to an append only audit trail. Anything sensitive is redacted on the way to disk, not on the way to a screen. And between the model and any irreversible action there is a deterministic gate that the model has no access to.

None of that is exotic. It is the difference between a demo and something you can leave switched on.

What this involved

  • Automation
  • API integration
  • Budget controls

Built with

  • Node.js
  • TypeScript
  • Anthropic API