Creating a process from a document

Pick the Hanbit Office Bulk Order Quote Approval Policy in the Knowledge Base and ask the Deep Agent to build a process that follows it. Approve the draft and pick two agents for the stock check and quote calculation, and the process, agents and input forms are created in one go. Then check the flow against the policy and ask for a fix where it differs.

How this differs from Chapter 3#

In Chapter 3 you described the steps one by one in words. But companies already have documents such as policies and work manuals. Pick a document and say "follow this policy", and the Deep Agent reads it and pulls out the steps, owners and gateways.

The key points of the policy (Hanbit-Office-Quote-Approval-Policy.pdf, uploaded in Chapter 8):

Try it 1 — Pick the policy and make the request#

1. In the chat box on Home (the pencil icon), make sure the agent is Deep Agent.
2. ＋ › Knowledge Base › Expand all › check Hanbit-Office-Quote-Approval-Policy.pdf › Use 1 file.
3. Send the text below.

Search the selected knowledge base for the Hanbit Office Bulk Order Quote Approval Policy and create a bulk order quote process that follows it. Keep its steps and owners, the approval by amount, and sending rejected quotes back to quote calculation. Let agents handle the stock check and quote calculation steps, and make an input form for each step.

Try it 2 — Approve the draft#

In our run, the draft window appeared after 26 seconds. Its title is the Deep Agent's opening sentence: "Here is the proposed process draft for the "Hanbit Office Bulk Order Quote Approval Process," derived directly from the official policy and following process best practices. Please review — if you want to add or change any steps, let me know." The draft follows the policy in five steps. The sales representative receives the quote request and fills in the quote request form. The stock checker (an agent) confirms the stock, the expected arrival date and whether the requested delivery date can be met. The quote calculator (an agent) calculates the unit price, volume discount, installation fee and total from the price list. Under KRW 30 million, the sales team lead approves or rejects; at KRW 30 million or more, the sales team lead approves first, then the finance officer. A rejected quote goes back to quote calculation with the approver's reason. Last, the sales representative sends the quote and records the date sent and the 30-day validity. The body ends with Shall I proceed as drafted? Tell me if you want to add or change any steps. In the current version, the opening sentence is repeated twice in the body and the headings show raw ** marks. You can ignore both.

If it matches the policy, choose Approve and click Submit response. If something needs changing, choose Reject and write what to change in the box below (Add your reasoning or requested changes (optional)).

Try it 3 — Pick agents for the steps#

Next, the Pick the items to add to this process (skip if none) window appears. These are candidates the Deep Agent suggests after reading the policy. In our run, it appeared 70 seconds after the request.

In our run, the window also suggested a [DMN] item (a rule as a decision table, Chapter 15). Other runs can suggest [Skill] items (documents describing how to do the work, Chapter 11, coming soon). This chapter picks only the agents.

Check [Agent] Stock check automation agent and [Agent] Quote calculation automation agent (2 selected), leave the [DMN] item unchecked, and click Submit response.

Try it 4 — Review and save#

In our run, about 128 seconds after the request, the Deep Agent said "I've created the Hanbit Office Bulk Order Quote Approval process, including agent-powered stock check and quote calculation steps, amount-based approval routing, and rejection loops, with dedicated input forms for each step." and asked us to review it and press "Save". The Work folder on the right already held 11 items. We replied "Yes, please create the process with this flow." anyway, and at about 157 seconds it answered "I've created the process with the specified flow—every step, owner, approval logic, agent-based automation, and input form is now implemented in files and ready for review." The Work folder still showed the same 11 items. This time no "holding the result back" notice came. If you get one, reply "Yes, please create the process with this flow."

Click Save at the top right of the Work folder. In our run, saving took a little over 2 minutes. "✅ Saved. Check the generated processes, skills, and agents." appears, with the saved items as cards: Generated BPMN process (1) (Hanbit Office Bulk Order Quote Approval Process, ID bulk_order_quote_approval) and Generated agents (2) (Stock check automation agent and Quote calculation automation agent). The Save button turns into Saved. This time both agent steps got new agents; the Hanbit Office Quote Advisor you made in Chapter 9 wasn't used. Once you go to another screen or refresh, the new agents appear under Agent Colleagues in the sidebar (for example Quote calculation automation agent, with the role "Quote Calculator (Agent)").

Try it 5 — Check the flow against the policy#

Hanbit Office Bulk Order Quote Approval Process now appears under Uncategorized in the Process Definition Map on Home. Click it to open it.

The last row differs. The amount split and the order of approvals are right, but on the KRW 30,000,000-or-more path a quote the sales team lead rejects has nowhere to go. Just as you checked the draft quote against the price list in Chapter 10, check flows the AI makes against the source too.

Try it 6 — Ask for a fix#

1. Stay in this chapter's chat room (Hanbit Office Bulk Order Quote Process), or reopen it from Conversations in the sidebar.
2. Type what differs into the chat box at the bottom (not the form preview on the right) and send it. We sent the Article 5 rule below, which restates the approval order:

This differs from Article 5. If the quote total is under KRW 30,000,000, the sales team lead approves and it goes straight to the reply to the customer; if it is KRW 30,000,000 or more, the finance officer must approve after the sales team lead before the reply. Fix the flow in that order.

Our message didn't name the missing rejection, but the Deep Agent fixed the flow against Article 5 and added it. If your flow has a different gap, describe that gap in your own words. In our run, the notice "Something went wrong while putting the process files together, so I'm holding the result back." came back after 26 seconds, with the reason "There's no passing validate_process_definition record yet. Try again, or tell me what to change." We replied "Yes, please fix it that way.", and at about 42 seconds the Deep Agent said the process had been fixed according to Article 5, with both approvers able to send a rejected quote back to quote calculation. We sent the same reply once more, and at about 63 seconds it confirmed: "The process flow has been fixed exactly as you specified". It also said "This revision is saved and ready for review", but nothing is kept until you click Save in the Work folder. Saving took about 2 minutes again.

In our run, the fix was saved over the same definition, so the map still shows one Hanbit Office Bulk Order Quote Approval Process.

If your run saved a second definition with the same name, delete the older one:

1. Open each one. The older one is the one without your fix.
2. With the older one open, click the red trash can at the top right (Delete Process) and confirm the deletion.

In the fixed flow, a new Sales Team Lead Approval Outcome (over 30M) decision comes after the sales team lead's approval on the KRW 30,000,000-or-more path. The rest stays as before.

In the current version, the Rejected arrows are still drawn into the top-left corner, and the label of Finance Officer Approval Outcome (over 30M) hangs below the pool. Only the drawing is off.

Next chapter#

You built a process with agents in it from a policy document, checked it against the source and fixed it. The policy's numeric rules (volume discounts, approval limits) are easier to change when kept separately in a decision table. Next up: decision tables (DMN) (Chapter 15).