How do you build an AI workflow for document processing?

Your team retypes the same fields from invoices, claims, contracts and statements every week, and the work never gets faster. Building the job once as a written workflow moves your people from typing to checking. You build an AI workflow for document processing by naming the fields you need, giving the assistant the document and your rules as source material, and checking every extracted field against the original before anyone uses it. Processing documents by hand costs money and time, and generative AI can classify, extract and analyse the data in them instead. 1 2 You check the output against the information your business supplied, and we help you build and test a workflow you can use again next week.

Key takeaways

• AWS reports that manual document processing is expensive and time consuming, and that generative AI can classify, extract and analyse document data in place of that repetitive work. 1 2
• Start with one repeated document type such as invoices, claims or financial records, where the fields you need are always the same. 3
• Write the workflow as five inputs: Context, Task, Source, Rules and Output.
• Test on a safe example and reconcile names, dates and figures back to the source document before the output is used. 4
• The four-hour Sunshine Coast workshop is A$495 including GST with the date to be announced, and private onsite training is quoted after a scoping conversation.

What is an AI workflow for document processing?

An AI workflow for document processing is a written set of instructions that takes a document, pulls out the fields you name, and hands you a draft to check. The volume of unstructured data such as documents, audio, video and images is growing quickly. 5 AWS Intelligent Document Processing uses optical character recognition, computer vision, natural language processing and machine learning to automate the handling of that data. 6

With large language models and generative AI added, this kind of processing extracts and classifies information from unstructured data and also produces short summaries and points you can act on. 7 Ready-to-use APIs process unstructured data at scale, classify documents, extract critical information, validate insights and generate summaries and reports. 8

The reason to build one is the cost of the manual version. Processing documents by hand is expensive and slow. 1 Organisations have to put people on high volumes of documents, which reduces their agility. 9 Generative AI classifies, extracts and analyses that document data and takes over the repeated manual work. 2

Which document tasks should you automate first?

Automate the document type you handle most often, where the fields you need never change. You can process invoices, taxes, benefit claims, licences and financial records to extract the data points a decision needs. 3 Those are the jobs where one set of instructions pays for itself the same week.

Several document types are already known to be slow by hand:

• Insurance work, where quotes, forms, claims and receipts arrive in varying layouts and formats, which makes extraction difficult. 10
• Contracts, which often arrive in non-standardised formats. 11
• Legal filings, where loading, reading and extracting the case number, parties involved or legal entities takes hours of manual effort. 12
• Lending, where incomplete loan packages, tax forms, paystubs and other missing data found during underwriting create more work and raise the risk of bad loans. 13
• Sensitive material, since generative AI understands context and can identify and secure sensitive data across large volumes of unstructured text. 14

Choose one specific piece of work to start: one supplier invoice, one claim form, one signed contract. Bring the document and the reference material that job needs. Your business task determines the training example. If the job runs across several stages, take the extraction step first. You can write the instructions for that step without connecting every system in your business.

How do you write AI instructions that extract the right fields?

Write the instructions as five inputs: Context, Task, Source, Rules and Output. The five inputs describe the job and give you something to check the draft against. Write them before you attach the document.

• Context describes the business, customer and situation.
• Task names the job the AI assistant needs to do.
• Source supplies the notes, documents or approved facts for that job.
• Rules set the limits and identify missing information to flag.
• Output describes the format and information the finished draft needs to contain.

For an invoice, that reads as one short brief. We are a service business checking supplier invoices before payment. Read the attached invoice and pull out the fields we approve against. Use only the attached invoice and our purchase order. Copy every figure and date exactly as printed, change nothing, and flag any field that is missing or unreadable. Return a table of supplier, invoice number, invoice date, purchase order number, line items, GST and total, plus a list of questions for me to check.

The rules do the heavy lifting. Keep supplied figures unchanged, and have the assistant flag anything it cannot find rather than filling the gap. We build training around your business tasks, source material and the output you need, so the five inputs describe your documents, not a generic example.

How do you check extracted data against the source document?

Check every extracted field against the original document before anyone acts on it. Compare names, dates, figures and reference numbers line by line with the source. Recalculate totals. Confirm the draft answers the request and stays inside the scope you set. The person responsible for the work approves the final version.

Test on a safe example first. In our sessions, participants run the workflow on a safe example and check the result against their own source material, which shows where the instructions are loose before real documents go through. Using OCR and NLP to extract text and specific terms automates the process with higher accuracy. 4

When a field comes back wrong or unsupported, point to it. Supply the missing fact, or tell the assistant to leave it as a question for you. Then check the next draft for the same problem.

Do you need a custom model or will a general AI tool do?

Start with a general tool and a written workflow, and move to a fine-tuned model when your document type keeps beating it. Document AI uses a model that provides both zero-shot extraction and fine-tuning. 15 Zero-shot means the foundation model is trained on a large volume of various documents, so it broadly understands the type of document being processed. 16 It can locate and extract information specific to that document type even if it has never seen the document before. 17

With Document AI you can prepare pipelines for continuous processing of new documents of a specific type, such as invoices or finance statements. 19 Snowflake's own guide trains on documents from the Contract Understanding Atticus Dataset (CUAD) v1. 20 Most small teams get their answer from the written workflow long before they need a pipeline, because the five inputs expose whether the problem sits in the model or in vague instructions.

How do you turn one document workflow into a process your team repeats?

Turn it into a repeated process by keeping the instructions and the tested example, then agreeing on who reviews the output. We help your team agree on the source material, instructions and review process for a repeated task, so the workflow belongs to the business rather than one person's chat history.

Store the five inputs where the team works. Store the approved reference material next to them: the price list, the purchase order rules, the fields your finance system needs. Name the person who checks the draft before it is used. Run the workflow on your next batch of invoices, claims or contracts, check the draft, and update the rules when a field slips through. Build workflows your whole team can run the same way.

Where can you get hands-on help building your document workflow?

You get hands-on help at our public Sunshine Coast workshop or through private onsite training at your workplace.

We bring private AI training to your workplace on the Sunshine Coast and in Brisbane, Gold Coast, Ipswich, Noosa, Moreton Bay, Logan, Redlands, Toowoomba and Gympie. Sunshine Coast onsite training covers Maroochydore, Caloundra, Mooloolaba, Buderim, Birtinya, Nambour and Coolum. People from other areas travel to the Sunshine Coast public workshop.

After your enquiry, we talk through your team's tasks, location and experience. We agree the training scope and prepare a fixed quote. Team size, session length and travel determine the private-training price.

We have done this work on our own documents and our clients'. We helped automate Radiant Roof Repairs' internal processes using AI. In our own business, Searchmaxxed, we built AI workflows for creating blog posts, doing SEO and preparing custom proposals. We used AI to automate SEO, enquiries and admin for Glen House Homoeopathy. You can see how we scope sessions by industry, including roofing, plumbing and electrical, real estate and accounting and bookkeeping.

Frequently asked questions