A note on the numbers upfront, because I'd rather you catch this than a client does: $239,000 is illustrative. It's a composite built from published benchmarks on manual data-processing costs, applied to a mid-market company on a legacy ERP (SAP, NetSuite, or Dynamics). Not a universal constant. Recalculate it against your own client's headcount and error rates before you quote it out loud. The method's laid out below so you can defend it if someone pushes back.
Everyone's tuning prompts. Almost no one's fixing the handoff 🤔
Scroll through any AI strategy conversation right now and it's all prompt engineering, model selection, fine-tuning, which LLM reasons best, which agent framework to standardize on. Easy stuff to obsess over honestly, it's visible, it's measurable, it's fun to tinker with.
Here's the problem though. None of it matters if the AI's output still has to get manually copied, reformatted, and re-keyed into your ERP by a person at 6pm reconciling the day's exceptions.
That's the handoff. The unglamorous seam between "the AI figured it out" and "the system of record actually has it." And it's exactly where the money leaks out.
