# An AI agent pre-screens order emails and turns a clear customer PO into a draft sales order in Zoho Inventory (Northline Components)

> In-house demo by Frédéric D'Amours with Grok Bot: an AI agent sorts order emails as clear or uncertain, turns the clear PDF PO into a draft sales order in Zoho Inventory, and the rep confirms and sends.

- Canonical URL: https://aiagentsformanufacturing.com/use-cases/ai-agent-order-emails-draft-sales-order-zoho-inventory/
- Department: Office & management (https://aiagentsformanufacturing.com/departments/office/)
- Evidence level: In-house demo. Developed and tested by Frédéric D'Amours himself, in a test environment with fictional data, using Grok Bot. Not a customer case.
- Added: 2026-11-02
- Last updated: 2026-11-02
- Publisher: AI Agents for Manufacturing (https://aiagentsformanufacturing.com)

> In-house demo on test data, developed and tested by Frédéric D'Amours himself in a test environment, using Grok Bot. Not a customer case; no time savings were measured.

## Role

Sales and customer service (order desk: entering customer orders into the ERP). Typical users: order desk or customer service representative, inside sales, sales coordinator, small-business owner who enters orders

## The problem

The order inbox mixes complete purchase orders, with the customer's own part numbers, descriptions and units, and vague requests with no PO. The rep reads everything, retypes the lines into the ERP and guesses on the unclear ones. Blind automation would get the vague requests wrong. This page documents a demo Frédéric D'Amours built himself to test a simple version of that flow on a small ERP, run directly in the chat with the agent.

## How AI helps

An AI agent built on Grok Bot, the agent platform, and set up as an order desk assistant, reads the order emails that arrived and pre-screens each one as clear ("I can create the draft") or uncertain ("I'm not sure"), with its reasons. It then asks the rep which orders to process through a clickable option list in the chat. For the orders ticked, it matches each customer line to a Zoho Inventory SKU despite different wording ("SSR 25 amp" = solid state relay 25 A, "24 VDC 2.5 A" = 60 W power supply, "RTD Pt-100" = PT100 probe), checks prices and stock, and creates a draft sales order in Zoho with the customer's PO number and the discrepancies written in the notes. The agent's role stops there: the rep reviews the draft, decides on the warnings, confirms it and emails the confirmation to the customer from Zoho. The agent sends nothing.

## The demo in video

https://aiagentsformanufacturing.com/cases/northline-order-booking/demo.mp4 The demo in 43 seconds: two order emails, the agent's pre-screening, the rep's choice, the draft in Zoho Inventory, and the confirmation sent from Zoho with the PDF. Faithful reconstruction of the real session, same data; email addresses replaced.

## Step by step

1. Give the agent read access to the order mailbox and to your item list (SKU, description, unit, price, stock), and write access only to create draft sales orders.
2. On request, have the agent summarise each new order email with a letter (A, B, C…) and mark it clear or uncertain, with its reasons. Nothing is written to the ERP at this step.
3. Let the rep choose which orders to process from a clickable list: each letter, useful combinations and "None, I'll handle it myself". Uncertain requests stay with the rep.
4. For each order ticked, have the agent match every customer line to your SKU and explain why, and flag price, stock, unit and unknown-item discrepancies. It never substitutes an item without saying so.
5. Have the agent create a draft sales order with the customer's PO number and the discrepancies in the notes, then stop.
6. The rep reviews the draft in the ERP, decides on each warning, confirms the order and sends the confirmation to the customer from the ERP.
7. Start with your most common customers and PO formats, then add more once the matching is reliable.

## Tools

- Zoho Inventory (free plan) through its REST API v1, with OAuth: used in the demo to read items, prices and stock and to create the draft sales order
- Grok Bot, the agent platform: the agent reads the order mailbox, runs the pre-screening and the option list in the chat, and never sends email
- Zoho Inventory's own email button: used by the rep to send the sales order confirmation with the PDF
- Other tools to evaluate: the same flow works with any ERP that has an API to read items and create draft sales orders, and a mailbox the agent can read
- Any LLM (ChatGPT, Claude, Copilot, Gemini) with the prompt below, on the order emails and an export of your item list

## Result

In-house demo developed and tested by Frédéric D'Amours with Grok Bot, not a customer case. Northline Components Inc. is a fictional company on a free Zoho Inventory organisation (Canadian data centre); the customers Arpent Automation inc. and Atelier Lac-Fictif inc., their emails and their PO are fictional, written by Frédéric. Tested on 2 order emails (Nov 2, 2026): A, PO ARP-24117 from Arpent Automation as a PDF with 6 lines, was screened as clear; B, an urgent request from Atelier Lac-Fictif with no PO, no prices, no part numbers and "the big one or the small one", was screened as uncertain and left to the rep. The rep ticked A only. The agent matched the 6 lines to Northline SKUs and created draft sales order SO-00003 for CAD 3,130.90 (the PO subtotal; 3,599.75 with GST and QST), with two warnings for the rep: the 60 W power supply is out of stock (-50, backordered), and the customer price for the 25 A relay is 29.00 versus a list price of 32.14 (-9.8%). The rep then sent the confirmation from Zoho, with the sales order PDF attached. Processing time and manual time saved were not measured.

## Next steps

- Advanced version for higher volumes: a small local service does the mechanical work (mailbox, PDF extraction, SKU lookups) and the agent only judges, with fewer tokens and a board to follow each order.
- More PO formats and customers, including unit conversions and items not in the catalogue.

## Human review: cautions

- This is an in-house demo on test data, tested on two order emails. It shows that the flow works, not how much time it saves; that was not measured.
- This simple version runs in the chat, which suits a low volume of orders. For higher volumes, the mechanical steps (reading the mailbox, extracting PDF lines, SKU lookups) should move to a small dedicated service so the agent only handles judgment, and costs fewer tokens.
- Matching customer wording to your SKUs is the risky step: keep the reason for each match visible and let the rep check it before confirming.
- Keep the agent on drafts only: confirmation and the email to the customer stay with a person, in the ERP.
- Give the agent only the API scopes it needs, and tag every record it creates so it can be found and removed.

## Try it yourself (prompt)

```text
You are an order desk assistant at a manufacturing company. A person approves every order and sends every confirmation.

Inputs: today's order emails (text and attached POs) and our item list (SKU, description, unit, price, stock).

Step 1: summarise each email with a letter (A, B, C…): customer, reference, and CLEAR ("I can create the draft") or UNCERTAIN ("I'm not sure"), with your reasons. End with a multiple-choice question I can tick (or a numbered list if your tool cannot): each letter, useful combinations, and "None, I'll handle it myself". Do nothing else.

Step 2: process only what I ticked. Extract the header and every line, match each line to our SKU and explain why (synonyms, abbreviations, V × A = W, unit conversions), and flag price, stock and unit discrepancies and items not found. Never substitute an item without saying so. Present a draft sales order as a table ready to enter in our ERP; I will confirm and send the confirmation myself.

If data is missing, write "not found" and ask. Do not invent prices or dates. Do not send, order or change anything in any system yourself: produce drafts for my review and approval.
```

## Related use cases

- [Customer PO and quote emails → sales orders in the ERP (drafts or automated entry) (RUKO, Gems Setra, Stelrad, Siemens Healthineers)](https://aiagentsformanufacturing.com/use-cases/customer-po-emails-to-erp-sales-orders/)
- [An AI agent chases a late PO in Zoho Inventory and updates it after approval (Northline Components)](https://aiagentsformanufacturing.com/use-cases/ai-agent-late-po-follow-up-zoho-inventory/)
- [AI sorts the shared customer inbox and routes each email (order, quote, after-sales, invoice) to the right person (Safran Landing Systems, Dossche Mills)](https://aiagentsformanufacturing.com/use-cases/ai-shared-inbox-email-triage-routing-safran-landing-systems-dossche-mills/)
- [AI drafts answers to customer questionnaires and RFPs from a library of approved answers and proof documents (Haworth, Ingredion, Cuisine Solutions)](https://aiagentsformanufacturing.com/use-cases/ai-customer-questionnaires-rfp-sustainability-answers-haworth-ingredion/)

## Sources

- In-house demo developed and tested by Frédéric D'Amours with Grok Bot, on a test Zoho Inventory organisation (Nov 2, 2026). Video rebuilt from the screen recording of the session; email addresses replaced.
- https://www.zoho.com/inventory/api/v1/

---
Please cite the canonical URL: https://aiagentsformanufacturing.com/use-cases/ai-agent-order-emails-draft-sales-order-zoho-inventory/