What you need to know
- The ten stages are an illustrative workflow; tools may combine or omit them.
- A missing quantity cannot be extracted. Keep corrections, calculations and estimates separate from source values.
- Confidence is a model estimate. Test your own tickets, field by field, and give exceptions a named reviewer.
Picture a construction delivery note photographed on a phone in the rain.
It's the pink carbon copy, and the supplier's logo is smudged. Someone has written a purchase order (PO) number in biro across the top at an angle. The box for the weight is empty, because the supplier's printer didn't fill it in.
Asking an AI tool to scan it can give you the supplier, the date and perhaps the PO number. The weight is different, because nothing is printed there to read. The best result is a blank, and the worst is a confident number that was never on the paper.
How does AI read a delivery note? Ten stages
Document tools can extract text and fields from a delivery-note photo. Some also return confidence scores or route exceptions for review. Those features depend on the tool and how the workflow is configured. A good scan still cannot verify that the load matches the ticket. Sources: Google Enterprise OCR documentation; Amazon Textract best practices.
The ten stages below are our explanation of a possible workflow, not a vendor standard or a claim that every product implements all ten. Some stages are combined, and checking, review and export need to be designed into the process.
- Capture. Someone photographs the ticket.
- Prepare the image. Correct orientation or lighting where the tool supports it.
- Find the text. Locate readable text regions.
- Read the text. Convert the visible marks into characters.
- Understand the layout. It works out what's a heading, a table or a column.
- Pick out the fields. It decides which number is the ticket number and which is the weight.
- Score certainty. Return confidence information where available.
- Check. Apply required-field, unit and duplicate checks; compare order/receipt data where connected.
- Person reviews. Route missing or doubtful values to a named reviewer.
- Export. Send approved data to the agreed destination, retaining a link to the original ticket.
Follow the ticket from the top of this page through the chain (we use "ticket" and "note" to mean the same thing):
On the ticket | What can happen | Where |
|---|---|---|
Smudged logo on a carbon copy | Characters missed or misread | Stages 3 and 4, or stage 1 if the photo is poor |
PO number in biro, at an angle | Handwriting is harder, and skew can hurt | Stages 2 and 4 |
Empty weight box | Nothing to read: a blank at best, an invented number at worst | Stage 6; a required-field check at stage 8 should flag the missing quantity |
Stages 1 and 2: getting a usable image
Some pipelines correct rotation, perspective or contrast before reading the page. These are possible processing steps, not a universal description of every OCR engine. The simplest improvement is often a better photo.
Fill the frame without cutting off the ticket, keep the print in focus and avoid glare. Google’s Enterprise OCR can report image-quality defects when that feature is enabled. A quality score is another check, not assurance that every field is correct. Source: Google Enterprise OCR documentation.
If a thumb or glare hides a number, retake the photo while you still have the ticket. Do not treat an AI reconstruction as an extracted value.
Stages 3 and 4: finding text and reading it
A common OCR pipeline locates text regions and recognises the characters within them. The implementation varies: newer models may combine these steps. The commercial question is whether the important fields were read correctly, not which architecture the vendor uses.
Is OCR the same as AI?
OCR describes the task of turning an image of text into machine-readable text. It can use AI. Field extraction is another task: identifying which text represents the supplier, quantity or order reference. Some products combine both.
Why carbon copies are hard
A faint carbon or carbonless copy can be harder to read because the marks are less distinct. That is a practical risk to test, not a measured accuracy claim about all copies. Use the clearest available copy or request the supplier’s digital record.
Include faint copies, dot-matrix print and handwritten additions in your test batch. Results from clean printed invoices do not establish performance on those tickets.
Can AI read handwritten delivery notes?
Some document tools support handwriting, but support is not a guarantee that a handwritten PO number or quantity will be correct. Test the handwriting on your own tickets and require a person to confirm important values. Source: Google Enterprise OCR documentation.
Stages 5 and 6: working out what the text means
OCR gives you text and where it sits on the page. It doesn't say which number is the weight. There are three ways to work that out.
- Rules and templates. Match labels, positions or patterns on expected layouts. Changes to the layout may require adjustment.
- Trained models. Learn fields from labelled examples. Microsoft supports custom models trained on examples; the minimum training set is a starting point, not evidence of accuracy on new suppliers. Source: Microsoft custom-model documentation.
- Generative extractors. Use a field schema and instructions, sometimes without initial labelled examples. Whether this works well on your tickets still needs testing. Source: Google generative-extractor documentation.
The output may contain fields and line items, such as a net weight and its unit. Check that figures belong to the right row and that the unit has not been lost. Complex or merged tables can be harder to extract reliably. Source: Amazon Textract best practices.
Keep training examples separate from the documents used to judge performance. Include unfamiliar layouts, poor photos and blank fields in the test set. Recheck results after changes to the model or supplier documents.
Ask how the tool handles a layout it has not seen and how corrections are incorporated. A minimum number of training documents is not a meaningful buying criterion on its own.
A general model can return a plausible answer that was not read from the ticket. Ask the supplier to preserve the source image and show where each extracted value came from. Any calculation or inference should be labelled separately from extraction.
Stage 7: what is a confidence score?
A confidence score is the tool’s estimate of certainty in a prediction. Microsoft reports field confidence on a 0-to-1 scale; Amazon uses 0 to 100. Scores are model outputs, not measured success rates on your site’s tickets, and not every field has one. Sources: Microsoft confidence guidance; Amazon Textract best practices.
Set review thresholds using a labelled test set and the consequence of a wrong value. A score of 0.95 does not establish that precisely one in twenty of your tickets is wrong. Measure errors and review workload before letting results post automatically.
A high score can still accompany the wrong value or wrong field. Check both the text and its assigned meaning, and test what happens when a required field is absent. Source: Microsoft confidence guidance.
The blank field problem
If a quantity is absent, extraction should return a missing value. A quantity derived from another record or a calculation is different evidence and should carry its own source and method. Never present an inferred weight as one printed on the ticket.
A blank may mean the value was never printed, this copy is incomplete, or the wrong field is being requested. For example, a concrete load may need volume rather than weight. Resolve the unit and source before accepting a number.
Require the tool to distinguish missing, unreadable and inferred values. A plausible number is not a substitute for the original record.
For each important field, count correct values, incorrect values, missing values and incorrect values approved without review. Break results down by supplier and image quality. That gives you a useful acceptance test on your own records.
Stages 8 to 10: checks, people and export
Checks may include required fields, units, totals, duplicate ticket references and comparison with the order. Three-way invoice matching compares the order, receipt and invoice; it is a finance control, not a standard feature of every ticket scanner. Source: HMRC procure-to-pay guidance.
A ticket supports the receipt record. If the site has not checked the goods, matching the extracted ticket to the invoice can simply confirm that two supplier documents contain the same mistake.
Decide who owns exceptions and which failures block export. Keep the original ticket, corrections and reviewer’s decision accessible. Automation only helps if the unresolved rows do not disappear into a queue.
What "99% accurate" means
“99% accurate” is incomplete without the unit, test set and scoring method. Ask whether the figure concerns characters, fields, whole documents or results after human correction. Do not compare unlike percentages from vendor marketing.
The unit matters. If a tool reads each digit right 99 times in 100, and errors are independent, a ten-digit reference comes out right about 90 times in 100. Real errors bunch together, so treat this as an illustration only.
A benchmark on forms, receipts or invoices is not a benchmark on your construction tickets. Test representative deliveries and report the rate of wrong values accepted without review, alongside accuracy and the proportion needing a person.
Confirm it while the lorry is still there
The best time to query a missing quantity is when the load arrives. Check the ticket, units and load reference against the receiving procedure. A visual check or pallet count cannot independently verify every bulk weight or concrete volume.
If the quantity cannot be verified at the gate, record that limitation and request the appropriate batch, weighbridge or supplier record. Keep verified, supplier-declared and estimated quantities separate. The materials delivery log template can support that routine.
The receiving team can add context the ticket lacks: the intended activity, unloading location, condition and any instruction reference. A photo helps preserve the document; recording actual use needs a later site or stock record.
When the ticket is already digital
A supplier’s digital ticket can reduce scanning work. A PDF still needs field extraction unless it arrives with usable structured data. Ask whether you can obtain CSV/API data, who can access it and whether corrections are included.
Heidelberg Materials UK describes ticket-data exports and its Datalink integration. That is a supplier service, not evidence of a Gather integration or a comparison showing other suppliers lack one. Ask each supplier what is available for your account. Source: Heidelberg Materials digital solutions.
Five questions to ask any vendor
Before you buy, run a batch of your own tickets, including the worst carbon copies, and score each field separately, counting wrong values approved without a flag. Then ask:
- What does your accuracy figure count? Characters, fields or whole documents, and on which ticket types?
- What do you return when a field isn't printed? A blank, a flag or a guess?
- What does your confidence score mean? How were the thresholds set on tickets like mine?
- How does it learn a new supplier's layout? A template, a set of examples (how many?) or a general model?
- Where does a person come in, and where does the data go? Into finance matching, the site diary or your carbon return, and does a failed check stop the row?
How Gather records deliveries
Gather records delivery notes within the shift record, alongside the day’s work and resources. Its Record product page also describes photos with timestamps and GPS metadata. These records support a delivery history; they do not, by themselves, confirm the quantity received or prove that materials were used. See the Gather Record product page or book a demo.
FAQ
How does AI read a delivery note?
A configured workflow can capture an image, extract text and fields, check the results and send exceptions to a person before export. Tools vary. The ten stages in this article are our explanation, not a mandatory architecture.
How accurate is AI at reading delivery notes?
There is no single useful accuracy percentage for every ticket. Test your supplier layouts, image quality and required fields, and count wrong values that pass without a flag. Separate automatic results from results corrected by people.
Can AI read handwritten delivery notes?
Some tools support handwriting, but the result depends on the writing, image and field. Test handwritten references and quantities separately and review critical values.
What is a confidence score?
It is the tool’s estimate of certainty, not proof that the value is correct or a measured accuracy rate on your documents. Source: Microsoft confidence guidance.
Is OCR the same as AI?
OCR is a task, not an alternative to AI. Modern OCR may use AI, and document tools may combine reading with layout analysis and field extraction.
Sources
- Technical sources: Google Enterprise OCR documentation; Microsoft custom-model documentation; Google generative-extractor documentation.
- Confidence and extraction limits: Microsoft confidence guidance; Amazon Textract best practices.
- Workflow source: HMRC procure-to-pay guidance. The ten-stage grouping and purchasing questions are Gather’s explanatory framework and recommendations.
- Supplier example: Heidelberg Materials digital solutions. No comparison of supplier coverage or vendor headline accuracy figures is claimed.
- Product source: Gather Record product page.
.webp)
