Contents
AI can read orders received by email, extract the customer, item numbers, quantities and dates, match them to your catalog and prepare the entry in your system, before a person approves it. A first use case starts at $2,300 (€2,000)excl. VAT, in 2 to 8 weeks, plus API usage costs.
At many B2B companies, some orders still arrive by email: a PDF generated by the customer’s system, a table pasted into the message, three lines typed in a hurry. Someone then re-enters them into the ERP. This is a use case where AI genuinely helps, as long as a person stays in the loop. Here is how I design it, what it costs, and when it is not the right answer.
What AI does, and what it does not
Language models can read free text or a document and pull structured information out of it. For an order, that means:
- identifying the customer, from the sender’s address, the signature or the attached document;
- extracting the lines: item number or description, quantity, unit, price if given;
- picking up related details: requested delivery date, shipping address, the customer’s PO number;
- matching each line to your catalog, even when the customer uses their own description.
What AI should not do on its own: approve an unusual order, accept a price that does not match the customer’s terms, or change stock levels without a check. Those decisions stay with people.
The typical flow, step by step
| Step | Handled by | Check |
|---|---|---|
| Arrival in a dedicated inbox | Automatic rule | None |
| Reading the email and attachments | AI | None |
| Extracting customer, lines, dates | AI | Confidence score per field |
| Matching to the catalog and customer terms | Rules + AI | Discrepancies flagged |
| Approving the prepared order | A person | Always approved at first |
| Creating the order in the ERP | Automatic connection | Creation log |
| Acknowledgment to the customer | Message template | Optional review |
The key piece is the review screen: the prepared order is shown next to the original document, uncertain fields are highlighted, and the person approves it in seconds instead of retyping everything. Over the following weeks, you may decide to let simple orders from regular customers go through without review, if the measurements support it.
What it costs
These amounts exclude VAT and come from my pricing.
| Situation | What is delivered | Indicative budget | Timeline |
|---|---|---|---|
| First use case on your current tools | Extraction, review screen, export or simple connection | From $2,300 (€2,000)excl. VAT | 2 to 8 weeks |
| Built into a business application | Full order module with AI, ERP connection, dashboard | $6,800 to $22,400 (€6,000–€20,000)excl. VAT | 1 to 3 months |
In the estimator, built-in AI counts for $2,800 (€2,500)excl. VAT, a connection to another tool for $2,300 (€2,000)excl. VAT and a dashboard for $1,700 (€1,500)excl. VAT, on top of the application’s core. AI API usage is added on top, billed directly by the provider. I do not promise a specific gain upfront: time saved is measured during the prototype, on your real orders.
Precautions to take
- Limit the data sent. Only the content needed for extraction goes to the API, with a provider whose business terms exclude using your data to train its models. The topic is covered in what data to share with AI.
- Keep a trail for every order: original document, extraction result, corrections, approval. If there is a dispute, you know what happened.
- Measure quality: share of orders approved without correction, most frequently corrected fields, average processing time.
- Plan a fallback: if the API is down, emails stay in the queue and are handled as before.
When it is not the right solution
- Your key accounts use EDI: electronic data interchange is more reliable than reading emails. The page on EDI for SMBs explains when it is the way to go.
- Your customers could order online: a B2B ordering portal removes re-entry at the source, with structured data from the start.
- Volume is low: a few orders a day are quick to retype, and the investment does not pay off.
- Orders are all different and ambiguous: if someone often has to call the customer to understand the order, AI will not solve that.
Often the right answer combines all three: EDI for the largest customers, a portal for regulars, and AI for those who will keep sending emails.
Where to start
- Gather about fifty orders received by email, representative of your customers.
- Describe how entry works today: who does it, in which tool, with what checks.
- Start with an audit and a prototype: in 2 to 3 weeks, your team tests the extraction on its own orders, and you decide based on real measurements.
The full approach, from the one-week audit to deployment, is described on the AI integration page.
Frequently asked questions
Can AI read scanned or photographed purchase orders?
Yes, current models read PDFs, scans and photos of documents, with reliability that depends on image quality. A clean, structured purchase order reads well; a blurry photo of a handwritten order needs human correction more often.
What happens when the AI does not recognize an item number?
The order is not approved automatically: it lands in a review queue with the uncertain field flagged. The person picks the right item, and that match can be saved for that customer’s future orders.
Do we need to change our ERP to set this up?
No. The automation connects to your ERP or management software if it has an API or a file import. If not, the prepared order can be exported in a format your team imports in one click.
How much does day-to-day AI usage cost?
The AI provider bills by usage, based on the volume of text processed. For a small or mid-sized company, that is often a few euros to a few tens of euros a month. Actual volume is measured during the prototype, before any rollout.