Ampwise AI
    Back to blog

    Processing Free-Text Orders: From Inbox to ERP Without Retyping

    Processing free-text orders means turning messy customer messages, such as “Send 40 of the 2-inch brackets like last time, deliver Friday”, into clean, validated order lines in your ERP. The reliable way to do it is a four-step flow: read the message and its attachments, extract the fields, validate them against your master data, and send only the doubtful cases to a person. Done well, your team approves instead of retyping.

    We built Ampwise around exactly this problem, so we will be specific about how it works, where people stay in control, and how to start.

    Ampwise
    ampwise.ai
    Move Free-Text Orders Into Your ERP
    Ampwise processes orders from Outlook or Gmail, including free-text emails and varied documents, to reduce manual data entry and errors.
    Explore Ampwise

    TL;DR

    • Free-text orders arrive without item numbers or a fixed layout, so templates and simple OCR break down. You need extraction that understands context.
    • The workflow that works: capture, extract, validate against ERP master data, route exceptions to a person, then post on approval.
    • Validation matters more than extraction. A field that was read correctly can still be the wrong price, unit or delivery date.
    • Start with a golden set of real historical emails, measure per-field accuracy, and widen automation as the numbers hold.
    • Ampwise works inside Outlook and Gmail and posts verified data to your ERP with one-click approval.

    Table of Contents

    What counts as a free-text order

    A free-text order is a purchase request that comes without structured catalogs or item numbers, often as unstructured text in an email, a form or an ERP field. Mercanis describes it this way in its 2025 write-up on automating them.

    In wholesale and manufacturing, that looks like:

    • A buyer replying to your quote with “OK, go ahead with option B, 200 pcs.”
    • An email body listing products by nickname, old part number or description.
    • A PDF purchase order in the customer’s own layout, with your SKU nowhere on it.
    • A photo or scan of a faxed order with handwritten changes.
    • An Excel sheet whose columns change from customer to customer.

    The same customer may order five different ways in one month. That variety is why manual handling is slow and why the first generation of automation often disappointed.

    Why manual handling hurts

    Someone has to read the message, work out what the customer means, look up the right item, check the price and stock, key it into the ERP and reply. Each step is small. Together they are the bottleneck between a customer’s email and a confirmed order.

    The cost shows up in three places:

    1. Time. Orders wait in a shared inbox until someone has a free hour. Technical sales teams spend their day on data entry instead of selling.
    2. Errors. A transposed digit or a wrong unit of measure becomes a wrong shipment, a credit note and an annoyed customer.
    3. Hidden costs. Mercanis reports manual procurement carrying an error rate of roughly 5%, and notes knock-on costs such as late-payment and express-delivery fees. Its own example has the cost per order falling from €115 to €67 with automation. These are one vendor’s figures for its own context, so treat them as a way to frame your own calculation, not a promise.

    The better benchmark is your own: count orders per week, minutes per order and corrections per month. Those three numbers make the business case.

    Why templates and plain OCR fall short

    It helps to separate three technologies that often get lumped together.

    Approach What it does Where it works Where it breaks
    OCR Turns pixels into characters Clean scans and printed forms It reads text but does not know what it means
    Templates and rules Pulls fields from fixed positions or patterns Stable forms from a few customers A layout change or a new customer breaks the rule
    AI document models Interpret layout and language in context Variable PDFs, scans, free-text emails Ambiguity still needs escalation

    The sequence in intelligent document processing is OCR, layout interpretation, field extraction, normalization and validation. As Neuronify’s guide on rules, OCR or AI for document processing puts it, reading text is not the same as knowing whether a number is a price, a quantity or a delivery date. That gap is the heart of free-text orders.

    For a wholesaler with hundreds of customers, maintaining a template per customer is a job in itself. Ampwise is designed to work without templates, so a new customer’s first email is handled like the hundredth.

    The workflow, step by step

    Here is the flow we recommend, whichever tool you use.

    1. Capture from where the order lands

    Orders live in Outlook and Gmail, so capture should start there rather than asking customers to use a portal. Ampwise works directly in the mailbox, with smart email labeling and status tracking so everyone can see which messages are new, in review or posted. It also extracts metadata such as sender, thread and attachments.

    2. Extract the fields

    The system reads the body and every attachment, whether PDF, Excel or scan, and pulls out customer, reference number, line items, quantities, units, prices, delivery dates and ship-to details. For PDFs specifically, our guide on email PDF parsing to ERP walks through how this works. Spreadsheets have their own quirks, covered in Excel order import automation.

    3. Match and validate against your data

    Extraction gives you what the customer wrote. Validation checks it against what you know:

    • Does the customer exist, and is the ship-to address one on file?
    • Does “the 2-inch brackets” resolve to exactly one active item, or three?
    • Is the unit of measure plausible for this item?
    • Does the price match the customer’s agreed price list?
    • Is the quantity in a normal range for this customer?
    • Has this order already been entered?

    That last check deserves attention. Our article on duplicate order detection in ERP covers why the same order arriving by email and again as a follow-up PDF creates duplicates. For purchase orders on the buying side, see the three checks before ERP release.

    4. Route what is unclear

    Missing, ambiguous or conflicting details are normal in free text. The right design does not guess. It flags the line, says why, and puts it in front of a person.

    Typical triggers for review:

    • Missing: no delivery date, no quantity unit, no ship-to.
    • Ambiguous: a description that matches several items, or “same as last time” with two possible previous orders.
    • Conflicting: the email body says 100, the attached PDF says 120.
    • Out of range: a price or quantity far from the customer’s norm.

    A good review screen shows the original email next to the extracted fields, highlights only the doubtful ones and lets the reviewer fix them in seconds.

    5. Approve and post

    Once fields are verified, the reviewer approves in one click and the data goes straight into the ERP. Nothing is posted without passing validation, and clean orders can move faster than doubtful ones. Our piece on automating email orders with human review before posting shows what this looks like for manufacturers.

    A worked example

    Imagine this email to a fastener wholesaler:

    Hi, please ship 500 of the M8 hex bolts we always get plus 20 boxes of the washers. Need by the 14th. Same address.

    What a capable system does:

    1. Identifies the sender and matches the customer account.
    2. Looks at that customer’s order history to resolve “the M8 hex bolts we always get” to one item. If history shows two variants, it flags the line.
    3. Notes that “boxes” for washers is a unit conversion question and checks the item’s pack size.
    4. Reads “the 14th” as a date in the right month and checks it against lead time.
    5. Uses the default ship-to on file, which satisfies “same address”.
    6. Presents a draft order with one yellow flag on the washers, and the rest green.

    The reviewer clears the flag, approves and the order is in the ERP. The human handled one decision instead of eight keystroke tasks.

    Measuring accuracy before you trust it

    Overall accuracy is a weak metric. A system can read 98% of characters and still get the quantity wrong on the lines that matter. Measure per field and weigh the critical ones, such as item, quantity, unit, price and delivery date.

    A practical method:

    1. Build a golden set. Pull 100 to 300 real emails from your mailbox history, covering your messiest customers, and record the correct order for each. This approach is recommended by Balázs Csorba, who also advises idempotency keys so the same order cannot be created twice in the ERP.
    2. Run the system against it. Compare each extracted field with the correct one.
    3. Set thresholds per field. A wrong note field is cheap. A wrong quantity is not, so critical fields get stricter confidence thresholds and more review.
    4. Track the right operational numbers. Straight-through rate, review rate, correction rate and time from email to posted order.
    5. Test with realistic traffic. If you want to rehearse the pipeline end to end, an email API for test inboxes lets you send and receive real messages programmatically against a sandbox before anything touches production.

    For how to think about the error side, see our guide on reducing data entry errors with AI as escalation, and for invoice-side measurement, field-level extraction for AP.

    What to look for in ERP integration

    Automation that stops at a spreadsheet only moves the retyping. Check these points before you commit:

    • Direct posting. Does verified data create the sales order, PO confirmation or goods receipt in your ERP, or does it export a file someone must import?
    • Master data access. Validation needs your customers, items, units and price lists. Without them, the system can only read, not check.
    • Duplicate protection. Retries and repeated emails must not create twice the order.
    • Audit trail. Every posted field should trace back to the source email and the person who approved it.
    • No retraining. Staff should keep working in Outlook or Gmail and the ERP they know.

    Our operations playbook for AI order processing covers realistic touchless targets, and order processing automation for ops teams goes through the rollout in more detail.

    A rollout plan that works

    1. Weeks 1 to 2: Pick one flow. Start with the order type that costs you the most hours, for example emailed customer orders or PO confirmations.
    2. Weeks 2 to 3: Build the golden set and baseline your current minutes per order and error rate.
    3. Weeks 3 to 6: Run in review mode. Everything goes through human approval. Track corrections to see where extraction needs tuning.
    4. Then: Widen. As critical fields prove reliable for a customer or order type, let more orders through with lighter review. Add the next document type.

    The point is steady trust, not a big switch. Time spent in review tells you honestly how far you can go.

    Our perspective

    We have seen the same pattern in wholesale and manufacturing teams: the order is rarely hard, the inbox is. Customers will not change how they write to you, and you should not ask them to.

    That is why we built Ampwise to meet the email where it arrives. It works in Outlook and Gmail, reads free-text emails, PDFs, Excel files and scans without templates, validates against your ERP data and puts a one-click approval between the AI and your system of record.

    We believe AI should do the reading and the checking, and people should keep the judgment calls. That balance is what lets teams cut manual entry by up to 90% while keeping control, and it is why many see a return within about three months.

    Evert

    See it on your own orders

    If your team spends its mornings retyping emails into the ERP, Ampwise is built for that exact job: it extracts, validates and enters orders, inquiries, invoices, PO confirmations and incoming goods straight from Outlook or Gmail, with your people approving the result. You can read more on the Ampwise AI page or see how it merges your email and ERP.

    Ampwise

    FAQ

    Can software really handle orders that have no item numbers?

    Yes, when it combines language understanding with your master data and order history. It resolves descriptions to items, and when more than one item fits, it flags the line for a person instead of guessing.

    How is this different from OCR?

    OCR only converts an image to text. Processing a free-text order also means understanding which words are quantities, prices or dates, matching them to your items and validating the result before it reaches the ERP.

    Do I need to change how customers send orders?

    No. The aim is to accept emails, PDFs, Excel files and scans as they come. Ampwise works in Outlook and Gmail, so customers keep their habits and your team keeps its tools.

    What happens when an order is unclear or incomplete?

    It goes to a review queue with the doubtful fields highlighted next to the original message. A reviewer corrects or confirms, approves, and only then does the data post to the ERP.

    How do I know it is accurate enough?

    Test on a golden set of your own historical emails and measure per field, with strict thresholds for item, quantity, unit, price and date. Start with everything reviewed and relax review as the numbers prove out.

    Sources

    📅 Book an Ampwise AI demoBack to blog