AI processes orders automatically

For those in a hurry

• Orders, invoices, bills of materials, customs information, bills of lading, transport orders, delivery instructions, etc., from PDFs, scans, or emails can be captured fully automatically in three ways: Discrete OCR (rule-based), Predictive AI (ML), and Generative AI (GPT).
• As a result, companies save themselves the manual entry of this data into ERP, TMS, OMS, TOS, CRM, etc. The productivity gain is enormous, and the cost savings are significant.
• Discrete OCR is best suited for structured, standardized documents, while generative AI is best for unstructured, highly variable documents.
• Generative AI (GPT) achieves recognition rates of over 98%, nearly as high as discrete OCR with processing rules, though it carries a slightly higher residual risk of errors.
• The trend is clearly moving toward generative automation methods combined with error handling, as this allows almost all orders to be captured automatically—regardless of how they arrive or how they are structured.

A tip for trying it out

Anyone wanting to process sensitive data should selectively anonymize it first. Personal data, confidential pricing information, or customer details fall into this category. The cloud service pdfFiller redacts fields in the best CIA fashion. This allows training data to be prepared to provide additional protection when processed in AI models.

The Gold Standard

Until now, powerful OCR platforms like docparser combined with an automation platform like Workato were the gold standard for the automated processing of PDF documents or scans. docparser works on the principle of marking every relevant area in the document and, if necessary, applying further evaluation rules.

This method is wonderfully suited for similar documents. If a company receives, for example, numerous orders that are similar in layout and thus have a predictable structure, we can achieve an accuracy of nearly 100%.

The disadvantage of these discrete OCR solutions is their dependence on a predictable document structure. If a company has orders from many different customers, each using completely different PDF layouts, a separate evaluation logic would have to be configured for each PDF layout.

Discrete OCR shines through its traceability and transparency.

The New Gold Standard

Since mid-2023, with the availability of GPT 3.5 Turbo and GPT 4.0, a new gold standard has begun to establish itself: Generative AI. Unlike discrete OCR, we "trust" that the GenAI model has already "seen" countless documents with similar content during its training and has been calibrated through reinforcement training to distinguish an item number from a supplier number and to understand complex price table structures.

Anyone who can use generative AI to solve problems in day-to-day business has a clear advantage. The industry is striving to harness the power of AI.

GenAI is conquering document processing.

Relieving Customer Service, Helpdesk, and Back Office

Where is the best place to start? Where work is not part of the core business and does not provide a competitive advantage.

PDFs, Word documents, emails, and scanned orders are still manually entered in many places. This ties up resources and costs money. Thanks to generative AI, these tasks can be more than 98% automated. The workflow:

1. PDFs, etc., are made machine-readable via an OCR service.
2. Prompts with relevant instructions are run against this read-in data with exact specifications.
3. The generative AI model performs its work and generates a result (completion).
4. Completions are received and converted into standardized order structures.
5. These orders are imported into the ERP or TMS via an interface. Job done!

All of this happens fully automatically. Orders can be received via any channel—email, file server, or interface (API).

Manual work in document entry is eliminated.

A further expansion stage: The AI evaluates responses from the ERP system and takes further actions—acting as an "automated dialogue" with itself. For example, after transmitting an incoming order, an automated search for an available transport slot can be triggered via another API, a suitable proposal can be automatically selected, and the order can be confirmed.

This is where artificial intelligence shows its true strength, as it mimics human behavior. It can combine decision-making factors perfectly because it can also learn them.

Successful Projects

There is an effective "recipe" as a best practice:

1. Quantify the volume of orders to be automated. How large is the economic potential?
2. Describe the functional target process as simply and completely as possible. What exactly should happen operationally and how?
3. Acquire or bring in technical expertise. Which technologies and providers are suitable?
4. Test success-critical points. How do we achieve the desired result quality?
5. Budget, staff, and execute the AI project.

One downside must not be concealed: You will certainly encounter problems. That is completely normal. Keep in mind that you are setting a large machine in motion. OpenAI, for example, required a double-digit billion-dollar investment to achieve today's performance.

Problems happen, and problems must be solved. This can range from unfavorable recognition rates ("accuracy") to questions of result interpretation and standardization. There are solutions and methods for this.

The universe is infinite—and so are the possibilities of AI.