Sales order entry automation: email to ERP without re-keying

A customer sends an order: three items typed into the body of an email, a spreadsheet attached, sometimes a scanned order form. A customer service representative reads it, translates the descriptions into your product codes, and types the lines into the ERP. At sixty orders a day that is two people’s full-time work, and none of it appears anywhere as a cost.

This article covers how that work gets automated, which decision must stay with a person, how to measure accuracy in a way that means something, and when the right answer is something other than automation.

What manual order entry costs

Typos become orders. 1200 entered instead of 120. Nobody notices until the shipment leaves, and the cost of putting it right dwarfs the cost of the entry.

Wrong variants get matched. The customer writes in their own vocabulary; the representative picks the closest-looking code. The wrong item ships.

Delay compounds. The order arrives in the morning and gets entered in the afternoon. Stock is allocated to something else in between.

Nothing is traceable. When an order goes missing, the search happens in mailboxes. Which email became which order was never recorded.

All four get filed under “carelessness”. None of them is. They are what happens when a process depends on typing speed and short-term memory.

What the automation does, step by step

Five steps, each independently verifiable. If they are not separable, you cannot tell where an error came from.

1. Capture. A dedicated mailbox is monitored. Body text, PDF attachments and spreadsheet attachments are taken separately.

2. Classification. Is this an order, an order amendment, a quote request, or something else? Misclassification is the most expensive failure mode, so the confidence threshold is set high here and anything uncertain goes to a person.

3. Extraction. Item description, quantity, unit, requested delivery date, delivery address, customer reference. In spreadsheets, columns are read by meaning rather than by position, so a renamed header does not break anything.

4. Matching. The customer’s wording is bound to your product code. Three sources are used together and each match carries a confidence score.

5. Writing. The order is written into the ERP as a draft. It is not released.

Template-based versus semantic extraction

Situation	Template / rule-based	Semantic (AI-assisted)
Fixed-layout PDF	High accuracy, low cost	High accuracy, unnecessary cost
A different layout per customer	Configuration per layout	Reads without configuration
Layout changes	Breaks silently	Generally survives
Free text in the email body	Does not work	Works
Approximate item description	No match	Proposes candidates with confidence
"Same as last time"	Does not work	Derives candidates from order history, asks
Cost profile	High to build, low to run	Low to build, per-document to run
Auditability	Rules are readable	Needs confidence scores and source citations

The right answer is usually a mix: structured channels — portal, EDI, standard spreadsheet templates — handled deterministically, everything irregular routed to the AI layer. That split materially lowers running cost, and it should be visible in any proposal you receive.

A worked scenario: a distributor taking 60 orders a day

The input. 60 orders daily. 25% arrive structured through a dealer portal, 45% as spreadsheet attachments, 20% as PDFs, 10% as free text in the email body.

Today. Two representatives read and key the orders. On busy days entry slips to the following morning.

With the system.

1. Portal orders go straight to the ERP and never touch the automation.
2. Spreadsheet attachments are read by column meaning.
3. PDFs and free-text orders go through extraction.
4. Each line is matched to a product code and assigned a confidence score. High-confidence lines arrive ready; low-confidence lines are highlighted.
5. The order is created as a draft in the ERP and lands in the representative’s queue.
6. An acknowledgement goes to the customer: order received, confirmation to follow.

Where the human approves. The representative sees the original document on the left and the extracted lines on the right. They correct highlighted lines, check stock and delivery date, and release the order. The confirmation email goes out only after that.

The output. The representative’s job shifts from entering data to checking it. Time per order falls, the backlog that used to spill into the evening disappears, and which email became which order is now on the record.

Why doesn’t the system release the order? Because a released order allocates stock, enters the production plan and creates a commitment to the customer. That is not reversible. Even at 98% accuracy, the remaining 2% is a broken promise rather than a typo.

How product matching gets reliable

Extraction is usually easier than expected; matching is usually harder. The reason is simple: customers describe products in their own language, not in your codes. “10mm black cable trunking”, “trunking 10x20 blk” and “KK-1020-S” can all be the same item.

Four sources make matching dependable, and all four should be in use:

1. Product catalogue and synonyms. Code, description, dimensions, colour, unit. If a single product exists three times in the catalogue, matching will be wrong regardless of the model — catalogue cleanup is a prerequisite, not an optional extra.

2. The customer’s own order history. The strongest signal available. If this customer used this phrasing before and that order was accepted, the match is close to certain.

3. Customer-specific code tables. Larger customers order using their own material numbers. Once mapped, that customer’s orders pass automatically almost every time.

4. Correction history. Every correction a representative makes is recorded and used in subsequent matching. Without this loop, accuracy is frozen at day one.

The practical consequence of the fourth point: accuracy in the first month will be below average, and that is normal. Judge the system on the trend in the correction rate, not on the first week.

Measuring accuracy honestly

“99% accurate” is the most quoted and least informative number in automation proposals. Three questions make it meaningful.

At what level? Field, line or document? 99% at field level in a 20-field document means roughly one error per document.

On which data? The vendor’s clean sample set, or your real orders?

What is the straight-through rate? Industry guidance targets over 80% of standard orders posting to the ERP without human touch. The number that matters more is how quickly the remainder gets resolved.

A practical acceptance test: take 100 random real orders from the last six months, write down the correct output by hand, and run the pilot against that set. Do not remove the ugly examples — they are precisely what needs testing.

Four weeks to go-live

Week 1. 100 real orders are selected and their correct outputs written down. The channel mix is measured. The ERP write path is verified.

Week 2. Extraction is built and the first accuracy measurement taken against the test set. Matching sources are connected.

Week 3. Review screen, exception queue and the draft-order write path are built. The acknowledgement flow is defined.

Week 4. Parallel run. The system processes real orders while representatives continue entering them manually; the two outputs are compared. When the gap closes, manual entry stops.

Skipping week four buys a week and costs the first incident. Parallel running is a trust-building step, not a technical one.

Cost

A single-process build lands in our $3,000-6,000 band over two to four weeks. Three factors decide where within it: the channel mix, ERP access method, and the difficulty of product matching. A dealer portal, customer-specific pricing rules or delivery scheduling are quoted as additional modules; a multi-module order management system sits in the $8,000-15,000 band. Annual maintenance is optional at 12-25%, licences are perpetual, source code is handed over. Full bands on the pricing page.

When not to automate this

Under about 15 orders a day. A well-designed entry screen with keyboard shortcuts delivers the same gain.

When most orders already come through a portal. If the remaining email channel is small, the build will not pay back. Increasing portal adoption is cheaper — the case for that is in B2B ordering portals for export distributors.

When the product catalogue is a mess. Three codes for one product guarantees wrong matches. Cleaning it is cheaper than automation and necessary regardless.

When pricing or allocation would be decided automatically. Order entry and pricing are different jobs. The system reads quantity and item; price and stock allocation belong in a rules engine or with a person.

When the confirmation email would be automated. An acknowledgement can be automatic; a confirmation is a commitment.

When the ERP is being replaced. Build order entry automation on the new system, not the old one, and pay once.

Measure your own situation

Five numbers, one week.

1. Daily order count and channel mix. Portal, spreadsheet, PDF, free text — as percentages.
2. Average entry time per order. Measure for a week. Count times minutes times 250 working days gives the annual figure.
3. Incorrectly entered orders in the last six months, and what they cost. Returns, cancellations, credits and lost accounts.
4. Average time from order arrival to ERP entry. Most companies have never measured this, and it is the number that moves fastest.
5. How many distinct customers and how many distinct layouts? Ten customers with ten layouts is a very different project from 200 customers with ten.

Convert the annual minutes to cost, add the correction cost from item three, and divide the build band by that annual figure. Under two years to payback is worth pursuing; above it, your channel mix probably is not suited to this.

Next step

Our approach to operational automation is on the AI process automation page, the organisations we build for on who we build for, and the bands on the pricing page. The invoice side is covered in invoice matching automation, and where approval points belong in human-in-the-loop AI approval design. Send the five numbers above through the quote form and we will tell you which channel to start with.

Frequently asked questions

Our customers do not use a standard form. Will this work?

It will, but do not go live without measuring it. Template-based extraction needs configuration per customer layout; semantic extraction absorbs the variation. Either way, plan for a correction period in the first weeks and make sure those corrections feed back into matching, or accuracy will stay at day-one levels.

What if it enters an order incorrectly?

It creates a draft, not a released order. The representative sees the original document, the extracted lines and a confidence indicator side by side, then confirms or corrects. Because a person releases the order, a misread becomes a correction rather than a shipment error.

How should accuracy be measured?

At line level, not document level. In a 20-line order, 19 correct lines still means a human reviews the whole thing, so the document counts as a failure. Use at least 100 real historical orders as your test set, and measure the straight-through rate separately from raw extraction accuracy.

How does it write into our ERP?

In order of preference: a documented API, the ERP's own import format, and only as a last resort direct database writes. SAP, Microsoft Dynamics, Netsuite and most regional ERPs offer at least one of these. Which path is used determines both build time and fragility, and it should be stated explicitly in any proposal.

Will the customer know their order was received?

An acknowledgement can be sent automatically when the draft is created, and we recommend it, because that is the question customers ask most. A confirmation, however, should only go out after a representative releases the order. Conflating the two means confirming orders you may not be able to fulfil.

Wouldn't a customer portal be better?

Usually yes, and the two are not competing. A portal makes orders structured at source, which is the cleanest solution available. The problem is that some customers will never use it and will keep emailing. Build the portal and automate the email channel; together they cover the whole population.

Related guides

• Gulf e-invoicing: what ZATCA and the UAE mandate actually require from your systems
• Automating repetitive back-office tasks: a practical method
• A B2B ordering portal for your distributors: what actually needs to be in it

Service page: Custom software service