Risk and compliance

What happens to an order the software can't read? How exceptions, failures and approvals work

By · Published

When automation can't read an order, the order is not guessed at and not lost. It is held as an exception with the original attached, a person handles it the way they do today, and the system records what happened. Nothing financial or irreversible happens without a person approving it. Failures are retried safely, duplicates are blocked, and a named person is alerted when something needs attention.

For 14 years Sandra has typed every emailed order into Sage 50 herself. She runs the order desk at a food wholesaler, and she knows which café means sunflower when it writes “the usual oil”. Any mistake was hers to catch, because every order passed through her hands. Three weeks ago the firm switched on software that reads the orders first. This morning, at 7.40, before she was in, a PDF order arrived that it could not read. Sandra has carried one worry since the project was announced, and she has not said it out loud. When the software gets one wrong and a van goes out short, who notices? And who gets the blame?

Sandra is made up. Her worry is not.

Key facts

  • Every automation Better Systems Co. builds has a held list, which engineers call an exceptions queue. Anything the software cannot read or match with confidence waits there for a person, with the original document attached.
  • Nothing financial or irreversible, such as raising an invoice, taking a payment or releasing stock for despatch, happens without a person approving it.
  • Duplicate protection, safe retries, a record of who did what and alerts to a named person are part of the build, not optional extras.
  • Accuracy is measured on your own documents during a side-by-side run, and the pass mark is agreed in writing before the software goes live.
  • In June 2026 the Federation of Small Businesses found that 54% of small firms worried about AI cite inaccurate responses (FSB). The design below answers that worry.

What happens to an order the software can’t read?

It is held, and a person deals with it. An order like that is called an exception.

Take the PDF that reached Sandra’s firm at 7.40am. The software works out the customer account from the email address and finds four of the five product lines. It cannot match the fifth, because the café has written “the usual oil”. It does not pick the nearest oil. It holds the whole order, with a note against line five saying no product matched. When Sandra sits down at 8.30, the order is at the top of her held list, with the original PDF beside the four matched lines. She knows this café. She picks the sunflower oil, or phones to check, approves the order, and it is created in Sage 50. Two minutes went on the line that needed her, none on the four that did not.

That answers her first question. The software noticed, then Sandra did, and no order existed until she approved it. A wrong order costs a credit note, a second van run and a cross customer. This one cost two minutes.

This is the worry behind a reply in the Sage Community Hub, where one user who had tried PDF readers found them “a lot less reliable than just getting the orders in excel format” (Sage Community Hub). The experience is real. No reader gets every order right. A safe one knows when it has failed and says so. Where one customer’s layout keeps landing on the held list, we fix the reader for that layout under a managed support plan.

What is a confidence threshold?

It is the point at which the software stops trusting itself and asks a person. Every detail it reads comes with a score for how sure it is, and below an agreed line Sandra decides.

The line is set detail by detail, because some mistakes cost more than others. The customer account must match exactly or the order is held. A product line must match a code in your product file, directly or through a name that customer has used before; a near miss is held. A delivery date that is slightly unclear can be flagged for a glance without holding the whole order. The lines start strict. They are loosened only where the side-by-side run shows the software was right every time it was unsure, and only with your agreement.

How are duplicates prevented?

By assuming they will happen and checking three ways. Customers send the same purchase order twice, send an amended version with the same order number, or a colleague forwards an email that was already in the orders mailbox.

  1. The software takes a fingerprint of each email and attachment, so an identical document is recognised whatever the subject line says.
  2. The customer account and purchase order number are checked against orders already created in your system and orders already waiting on the held list.
  3. Every order put into your system carries its own unique reference, so trying again after a fault cannot produce a second order.

A suspected duplicate is never created. It is held with the reason “possible duplicate of order 10482” and the two documents side by side, because only a person can tell a resend from an amendment.

What happens when something fails halfway through?

The software tries again safely, and if that does not work, somebody is told. Systems fail for dull reasons. The accounts package is locked during its backup, the email server does not answer, the internet drops. The software treats each as expected.

  • The order stays on the list until your system confirms it has it. The inbox is the record; nothing is deleted or moved until then.
  • Retries are spaced further and further apart, so a two-minute outage costs two minutes and a 3am fault is retried until the office opens.
  • A retry can never create a second order. If the first attempt did succeed and the confirmation was lost, the second attempt finds the order already there and stops.
  • After an agreed number of failed attempts, the order is marked as failed, left on the list for a person, and an alert is raised.

The worst case looks exactly like today. Sandra keys the order from the email. The best case is that the fault passes and nobody notices.

What is in the audit log?

The audit log is the record of everything that happened to every order, and who did it. It shows when the email arrived, what the software read and how sure it was, every detail a person changed and who changed it, who approved the order and when, when it went into your system, and the order number your system gave back. The original email and attachment are kept with the record.

This is where Sandra’s second question starts to get an answer. When a customer says they never ordered twelve, the record shows the PDF that said twelve. It shows your accountant how a price reached an invoice, and it shows why one customer’s orders keep being held. Nobody has to argue from memory. The log is yours and stays readable without us.

How is it monitored, and who gets called?

Monitoring watches three things: that the software is running, how long the oldest unprocessed order has been waiting, and how many failures there have been in the last hour. Each has a limit agreed with you, such as thirty minutes’ wait in office hours; an alert goes out when it is passed.

Who gets called depends on the kind of problem.

Kind of problem Example Who is alerted
An order that needs a decision Unknown customer, product not found, price mismatch Your named person, through the held list
A technical fault Accounts system unreachable, mailbox login failed Better Systems Co., and your named person
A change of layout A customer’s PDF layout changed and nothing matches Your named person sees the held list fill; Better Systems Co. fixes the reader under support

Under a managed support plan the response times are written down and Better Systems Co. fixes technical faults within them. Without a plan the automation still runs, alerts still reach your named person, and fixes are quoted on a time basis. Either way, a named human being is told, not a shared inbox nobody reads.

What is a side-by-side run?

A period, agreed in the statement of work, which is the written agreement for the project. During it your team keys orders exactly as today, and the software reads the same orders alongside them without putting anything into your live system. Each day the two sets are compared.

Every order falls into one of three outcomes: read correctly and would have gone through; held for a person, correctly; or wrong and not held. The first two are fine. The third is the only number that matters, and the side-by-side run exists to count it on your real orders before anything is at stake. For those weeks Sandra is the judge. Her typed order is the right answer and the software is marked against it. The reader is adjusted, the lines are moved, and the names your customers use for your products are learned. Only when the count meets the agreed pass mark does the software start creating orders, and even then every order waits for approval. The how it works page shows where this sits in a project.

What does “accuracy” mean, and how is it agreed?

It means the number of orders that were wrong and not held, counted on your orders, during the side-by-side run. We do not quote an accuracy percentage in advance, because a figure quoted before seeing your customers’ documents would be a guess, and the FSB survey above shows what small firms think of confident guesses.

The pass mark is written into the statement of work in plain words. Over the side-by-side period, out of all orders received, no more than an agreed number may be wrong and not held. And none of the details that carry money, which are the customer account, product code, quantity and price, may be wrong on an order that went through. The agreed number is yours to set; a business shipping perishable goods to supermarkets will set it lower than a merchant whose customers check every delivery note. If the pass mark is not met, the software does not go live. The reader is adjusted until it is, or we tell you plainly that your order formats are not yet suitable.

What never happens without a person?

Anything that costs money or cannot be undone. In every build this list is explicit:

  • Raising or sending an invoice.
  • Taking or refunding a payment.
  • Releasing stock for picking or despatch.
  • Changing a price, a credit limit or a customer’s terms.
  • Sending anything to a customer or supplier.
  • Deleting or overwriting a record in your system.

On its own the software reads, matches, drafts, checks and records. It creates an order only once a person has approved it, or once you have chosen, after the side-by-side run, to let routine orders from known customers through on a single click. See Will automation replace our office staff? for the human side and automating email orders into Sage 50 for one common case. The terms used here are defined in the glossary.

So who notices, and who gets blamed?

The software notices first, a person second, and the record settles the rest.

Picture Sandra’s morning once this has bedded in, still as an illustration. She gets in at 8.30 and the overnight orders from her regulars have been read and are waiting for her approval. A few sit on the held list, each with its reason and its original beside it. She does those first, because they need someone who knows the customers. When the phones get busy she is answering them, not typing.

As for blame, an order the software held is the design working. A fault in the software is ours to fix. Sandra is responsible for what she approves, as she always was for what she typed, but now she checks with the original beside her and a record behind every decision. A firm with those safeguards agreed in writing can let software do the reading, and confirm its orders sooner than a competitor still typing them.

What we would do

We would design the held list, the lines, the duplicate checks and the alerts before the reader, because they are what make the reader safe to use. Then we would run the software alongside your team on your real orders for the agreed period, show you the three counts each week, and go live only when the written pass mark is met. The sales order processing automation service page describes the build; the how it works page describes the stages; and managed automation covers who answers when something fails once it is live.

Sources

Questions people ask

Can we see every order before it goes into our system, not just the ones the software is unsure about?

Yes, and that is how every project starts. During the side-by-side period and for as long as you want afterwards, every order waits for a person to approve it on screen. Only once the agreed pass mark has been met, and only if you choose, do routine orders from known customers start going through on a single click or in bulk. The default is a person on every order.

What if the automation stops working at 8am on a Monday?

The inbox is still the inbox. Orders are never deleted or moved out of your mailbox, so your team can key them by hand exactly as before while the fault is fixed. Monitoring raises an alert when the system stops processing, and under a managed support plan Better Systems Co. responds within the agreed time. Without a plan, the alert still reaches your named person and the fix is quoted on a time basis.

How do you stop the same order being created twice?

Three checks. The software takes a fingerprint of the attachment and email, so a forwarded copy is recognised. The customer account and their purchase order number are checked against orders already created. And every order put into your system carries a unique reference, so a second attempt after a fault cannot produce a second order. A suspected duplicate is held for a person, never created.

Who decides what counts as accurate enough?

You do, with us, in writing, before the software goes live. During the side-by-side run we count orders that were read correctly, orders that were correctly held for a person, and orders that were wrong and not held. The last category is the one that matters, and the acceptable number is agreed in the project's written agreement on your real orders, not promised in advance.

Related guides

Find out what the typing is costing you

A 20-minute call. You tell us about your Sandra, your Dave or your Phil, and how the job gets done today. We tell you honestly whether automating it would pay, and what it would cost.