Anyone who sells orders knows the step after the signature: the order confirmation arrives as a PDF, and someone copies it into the project tool. At t3n, every booked service needs its own task in Asana, with subtasks and due dates. For large bookings that took up to 40 minutesM. Since mid-August 2026, a Claude agent has been creating these entries. The team only reviews them.
The starting point
The Asana board already had fixed fields and templates. That is the key prerequisite: the agent does not invent a structure, it fills an existing one. According to t3n, Claude is "connected to the company's internal tools". The article does not say which Claude product or which integration.M
How the agent was trained
The project lead first asked Claude what it needed to automate order entry. An internal skill made Claude ask detailed follow-up questions. Claude then received the process documentation and ten real order confirmations.
The decisive step: Claude rebuilt orders the team had already entered by hand and compared its results with the real entries. The differences became its first set of instructions. In the first test, Claude found two errors in an entry the team had created itself.M
After that, training happened in a Slack thread. The team described errors in plain language, and Claude added each correction to its instructions. After 24 hours they were at version 12.M The team never created a test copy of the board.
The workflow
- 01StartOrder confirmation arrives as a PDF
- 02PersonA team member tags Claude in the Slack thread for the confirmation
- 03AIClaude reads the PDF and maps the booked services to the template
- 04SystemClaude creates one task per service with subtasks and due dates on the Asana board
- 05CheckTeam checks: are services, dates and fields correct?no: describe the error in the thread; Claude fixes the entry and its instructions
- 06ResultOrder is on the board; review takes under 5 to at most 10 minutes
What it delivers
- Share: The agent creates over 80 % of advertising ordersM.
- Time: Instead of up to 40 minutes of entry, review takes under 5 minutes for simple orders and at most 10 minutes for large ones.M
- Quality: According to t3n, no error reached daily operations in the first two-plus months.M
- Cover: "The team notices the difference most clearly when someone is on holiday."
All figures are the newsroom's own statements and have not been independently verified. t3n gives no costs, and does not say which orders fall outside the 80 %.
What you can adopt
- Structure first, agent second: Without fixed fields and templates, the agent has no target. Build the template before you train.
- Test with past cases: Completed orders entered by hand are your test bench. The comparison shows differences without affecting a real customer.
- Corrections go into the instructions: Every correction belongs in the instructions, not just in the single entry. Otherwise the agent repeats the error.
- Review stays: A person reviews every entry. That keeps errors out of daily operations and makes clear who is responsible.
The pattern is not tied to Claude or Asana. What matters is that the assistant may write to the project tool and can be reached in the team chat.
You create orders in our project tool. Here are the process documentation, ten order confirmations and the entries we created for them by hand. Rebuild each order without looking at the finished entries. Then compare your result field by field with our entry. List every difference and turn it into a rule for your instructions. Ask if a rule is ambiguous.
Limits and risks
- Data: Order confirmations contain contact names and terms. Before you start, check the data processing agreement under Art. 28 GDPRR and whether the vendor uses inputs for training.
- Write access: The agent writes straight to the live board. Give it access to that one project only, not to the whole workspace.
- No test copy: t3n did without a test copy. If your board triggers invoices or production, plan one.
- Works council: If the review logs the performance or errors of individual people, it is a technical facility under Section 87(1) no. 6 BetrVG.R
- Competence: Since 2 February 2025, Art. 4 of the AI Act has required measures on employees' AI literacy. The omnibus Regulation (EU) 2026/1744 reworded the duty: what is required are measures to promote it.R Whoever trains and reviews the agent should be trained, and the training should be documented.
How to start
- Pick a document that arrives regularly and is always copied into the project tool the same way.
- Define the template and required fields if they do not exist yet.
- Collect ten completed cases with the document and the finished entry.
- Have the assistant rebuild the cases and write down the differences as rules.
- Go live with a review of every entry and measure review time for four weeks.