Why Document Automation Is a Practical Starting Point for Enterprise AI

“For most of ERP history, the human adapted to the system,” says Bastian Schiele, Senior Manager – Technology at cbs Corporate Business Solutions. “You learned the transaction, you followed the process, you entered the data in the format the system expected.”

“AI inverts that dynamic,” he says. “The system starts coming to meet you, working with natural language, understanding context, processing content in the form it already exists rather than demanding it be translated first.”

Over nearly a decade at cbs, moving from Solution Architect through to his current role, Schiele says this is the shift that matters most, and it “has nothing to do with integration patterns or platform choices.” It reframed the question he brings to clients.

The work is “no longer just ‘how do we connect system A to system B,’” he says. “It is ‘how do we redesign the interaction layer so that the system absorbs more of the friction that humans are currently carrying.’”

Document automation, he argues, is where that idea becomes most practical.

A Problem That Kept Showing Up

Document automation was not something Schiele set out to own.

“It was never a strategic priority,” he says, but it kept appearing because, as he puts it, “almost every client has the same problem: high-value business processes being held together by manual document handling.”

It was a problem that recommended itself, he says, because everyone in the room could see it: “It is concrete, it is measurable, and it resonates immediately.”

When SAP’s Hack2Build initiative came in 2024, and cbs looked for a problem worth solving, document automation was the obvious candidate. What started as a Hack2Build submission grew into cbs Advanced Integration of Documents (cbs AID), and the effort went on to earn the SAP Business AI Hack2Build Award and a “Built with SAP Business AI” certification, among the first solutions globally to receive it.

Where Old Tools Hit a Ceiling

Plenty of technologies already promise to read documents. Schiele’s contention is that reading was never the hard part.

“The document problem is not really a technology gap,” he says. “Clients have tried OCR, they have tried RPA, they have tried template-based tools. The gap is that none of those approaches understand SAP. They can read a document, but they cannot connect what they read to how an SAP process works.”

Each established approach, he says, has a ceiling. Template-based tools “break the moment a supplier changes their layout,” and maintaining hundreds of templates “is a burden that compounds quickly.” OCR-only tools convert images into text “but have no understanding of what that text means in an SAP context.”

Even pre-trained layout models fall short when extracted information enters a configured SAP process. “Knowing a field is a ‘total amount’ is not the same as knowing how it maps to a company code or document type in your S/4HANA configuration,” Schiele says.

What he describes as the real differentiator is the reasoning that follows extraction. cbs AID is agentic. “AID orchestrates a chain of reasoning steps: understanding the document in context, validating against SAP master data, applying business rules, and determining what needs human review versus what can flow straight through.”

“It is not a pipeline that reads and dumps,” he says. “It is an approach that reads, reasons, and acts. That is what makes the output usable inside an SAP process.”

The Documents Nobody Designed for a Machine

The true stress tests, Schiele says, “are the ones where the document itself was never designed with automation in mind.”

He offers certificates of origin as a case in point. “We are talking about documents that can run up to 200 pages, with nested relationships between data points, large amounts of irrelevant content mixed in with what actually matters, and no consistent structure to lean on.” The system has to distinguish relevant information from noise, follow relationships across different sections, and produce results that remain reliable at scale.

If certificates of origin are a test of sheer complexity and scale, the next challenge Schiele describes is about something else entirely: speed, language, and a moving target.

That case involved a US-based company under quarter-end pressure. Its customer contracts come in across multiple languages and layouts, all due in the same window. Many include upgrade agreements whose prices have to be assembled from different parts of the document, and a marketing team that keeps coining new phrasing means the system can never settle into a fixed pattern. Every contract is, in effect, a new challenge.

That customer, Schiele notes, “had gone through two previous automation solutions that both fell short on accuracy and scalability. With cbs AID they finally have something that holds up across languages, across layouts, and under the time pressure of quarter end.”

A publicly discussed deployment at IDEXX Laboratories shows similar pressures.

The pet-healthcare company uses cbs AID to feed customer-contract data into SAP Revenue Accounting and Reporting, and reached roughly 98% accuracy in the feasibility phase after two earlier solutions fell short.

What changes once cbs AID is in place, Schiele says, is where people spend their time. “Instead of keying data and chasing discrepancies, the team shifts to reviewing exceptions that genuinely need human judgment from data entry to exception management.”

Where to Start and What Comes After

The case Schiele makes to a prospective customer is modest: you can begin now, with the landscape you already have. “Almost any customer, and that is intentional,” he says.

AI projects can stall when customers believe they must first complete a larger cloud or ERP transformation, but cbs AID is designed to avoid that dependency.

For the large base of organisations still running SAP ECC or fully on-premise: “The manual effort tied to those processes does not disappear because you have not yet migrated to SAP S/4HANA. So why should the ability to address it?”

What makes that first step safe to take, especially in regulated industries, is that nothing disappears into a black box. Every extraction carries a confidence indicator and a clear status, so reviewers can see where the system is certain and where it is flagging something for a human. Rather than sending everything to a manual queue, the system flags only what genuinely needs judgment, letting customers automate at scale without losing oversight.

That combination of a low entry barrier and a process people can trust is what turns one project into something larger. Once a document-heavy process runs reliably, Schiele says, the discussion shifts “from ‘Can this work?’ to ‘Where else can we apply it?'”

Customers typically expand in two directions. They apply the same approach outward, to other document-heavy workflows such as order confirmations, delivery documents, contracts. And they push inward, bringing AI deeper into the SAP process itself, into the decision points where data is checked and only true exceptions reach a person. That deeper integration, Schiele says, is where the value really compounds.

cbs then helps customers assess which additional use cases are practical, where they can create value, and how they fit into the organisation’s transformation plans. “The goal is not to chase isolated AI experiments,” he says, “but to identify areas where AI can improve end-to-end processes and then scale those improvements in a controlled, value-driven way.”

Document processing may begin with a familiar pain point. But its larger value lies in showing customers how AI can participate reliably in SAP processes.

What This Means for SAPinsiders

AI adoption no longer depends on ERP modernisation. Deploying across SAP ECC, on-premise, and SAP S/4HANA separates immediate automation value from long transformation timelines. Document AI can therefore build the business case and organisational confidence for deeper modernisation.

Documents become the gateway to transactional AI. They connect unstructured business inputs with SAP’s governed processes, giving companies a practical route from generative AI experimentation to systems that can validate, decide, and execute.

Process design overtakes extraction accuracy. Once document recognition becomes reliable, the limiting factors shift to master data quality, business-rule consistency, and exception ownership. Automation becomes a diagnostic test of SAP operating discipline.