cleandataengineering.aiAgent Harness, drawn simply
You are here: Cost › At scale

Agent Harness Tutorial · Cost

What it would cost at scale

The demo costs cents. The Series also works out a real deployment: 2,000 conversations a day, 6 turns each, on Azure with GPT-5 models. These are its numbers, per day.

ALL-AGENTHYBRID (fixed flows + RAG + agent)Models$28.75$10.67AI Search S1 × 2$16.13$16.13Semantic ranker$11.97$5.40Layout extraction$5.00$5.00Hosting$6.30$6.30Operations$4.42$4.42People (0.2 FTE)$54.79$54.79
Azure per day: $72.57 all-agent, $47.92 hybrid. People are the same in both and are the biggest line.

Hybrid sends 45% of turns to fixed flows, 35% to retrieval and only 20% to the agent loop. It cuts the Azure bill by a third and model spend by nearly two thirds. Prompt caching at 60% is assumed; losing it is one of the biggest jumps.

Cost per outcome. If 1,000 of those conversations would have been phone calls and 60% are resolved, that is 600 calls avoided at $5: about $3,000 a day of value against about $103 of total cost. The number to watch is total cost per resolved conversation.

The Series has a calculator for this; change the volume, the split, the models and the cache rate to see what moves the total.

Want every detail? The long version is in the Series: the cost section and TCO calculator.