Examples
Each of these workflows began as a task that a person typed into a chat. The person did the task once with Claude and the Toolscaled connector, then asked for it to run by itself. Each page shows what the workflow does, what you fill in, what it needs, where the result goes, the messages that the person typed, and its numbers.
Answer support mail with your refund policyLabels each new support mail, finds the order, and saves a draft reply. 8 messages, 27 minutes to build.
Check supplier invoices against purchase ordersEach Monday, a sheet with one row per invoice and a summary mail. 6 messages, 30 minutes to build.
Watch a stock list for price rules and new filingsEvery 15 minutes, a mail, a sheet row and a calendar event for each alert. 6 messages, 14 minutes to build.
Rate customer accounts each weekEach Friday, a rating in HubSpot and a draft to the owner of each account at risk. 5 messages, 18.5 minutes to build.
Research a company after its quarterly reportFor a ticker, a company report, a peer sheet and an industry report. 17 messages, 99 minutes to build.
Make a dossier for each contact-form mailA Google Doc on each sender, made or updated on each new form mail. About 45 minutes to build.
To build your own, tell your chat what you do now by hand, as these people did: where the data is, what you check, and where the result goes. Then ask it to do the task once and to make it run by itself. Your first workflow walks you through it.
How we measured
Section titled “How we measured”- Each task has two data sets. The build chat saw only the setup set. The workflow ran on a test set, with other names, amounts and dates, and with cases that the setup set did not have. No step changed between the two.
- The workflow ran by itself, with no chat and no person. Its numbers are the median of 5 runs.
- As built, these workflows use several AI models. For the numbers on these pages, each AI step used GPT-6.1 Terra, the model of the agent.
- The agent is Hermes Agent on a computer, with GPT-6.1 Terra, the Toolscaled tools and file tools. It got the task as one prompt, with no person present. Each task has one agent run.
- The time of a run goes from its start to its last action.
- Workflow cost is the credits of the run at the Starter price of $19 for 10,000 credits. What a workflow costs explains the credits.
- Agent cost is its tokens at the list price of GPT-6.1 Terra. A ChatGPT plan paid the real bill.
- Build cost is an estimate of the build chat at list price. When you build in your own Claude or ChatGPT, that plan pays for the chat.
- Payback is the build cost divided by what one run saves against the agent.
- “Same answer in 5 runs” compares the 5 runs of a workflow with each other: the share of the graded fields that are the same in every run.
- The
[bench]tags in the messages mark test data.