The challenges shaping enterprise AI decisions right now
Decision makers, software engineers, and business leaders are navigating a set of genuinely difficult problems. Each one illustrates why people remain central to responsible AI adoption.
Meeting standards such as BSI C5
As organizations move critical processes to cloud environments, regulators and enterprise procurement teams are raising the bar on security assurance. Andrew Pery, AI Ethics Evangelist at ABBYY, shares in his article how Germany's BSI Cloud Computing Compliance Criteria Catalogue (C5) is emerging as a global benchmark precisely because it goes beyond policy documentation. It requires independently validated operational controls, transparency into data handling, and evidence that security actually functions in practice, not just on paper.
For AI platforms processing sensitive documents at scale, alignment with frameworks like BSI C5 is becoming a competitive differentiator. ABBYY's commitment to trusted cloud operations, including strong governance, access management, monitoring, resilience, and transparency, reflects an understanding that security is not a feature to be added later. It is a foundation.
Detecting document fraud in a world of synthetic identities
Creating a convincing fake identity document once required specialized skills and physical materials. Generative AI has changed that equation entirely. Templates can be cloned, bank statements altered, and metadata tampered with at a speed and scale that traditional fraud controls were never designed to handle.
The problem compounds because fraud review processes have not kept pace with the sophistication of modern attacks. Nick Carr, Director of Pre-Sales, explains how combining Document AI with document forensics addresses this, processing high volumes of documents accurately, surfacing anomalies that isolated review would miss, and escalating only the cases that genuinely require human attention.
Ensuring documents do not corrupt before AI reads them
This is one of the most consequential problems in enterprise AI, and one of the least discussed, according to Maxime Vermeir, VP of AI Strategy. When a typical PDF passes through raw extraction, reading order collapses, tables flatten into meaningless strings, multi-column layouts interleave into noise, and structural hierarchies disappear. The document is effectively corrupted before a model ever processes it.
ABBYY research indicates that when document parsing accuracy drops below 70% to 80%, most downstream natural language tasks become unreliable. Models trained on that kind of data do not learn domain knowledge; they learn the noise. The result is higher hallucination rates, weaker retrieval in RAG workflows, and pipelines that demand constant rework.
Structure-preserving extraction is not an optional enhancement. It is a prerequisite for reliable enterprise AI, and the principle behind ABBYY Document AI. For a deeper analysis, read: Your Documents Are Corrupting Before AI Even Reads Them.
Using AI tokens effectively
Token costs may be falling, but that does not mean organizations should apply LLMs to every workload. The more important question that Slavena Hristova, Director of Product Marketing, asserts in this article is one that enterprise leaders are increasingly asking: is an LLM the right technology for a given task in the first place?
Building a custom AI stack often appears cheaper than purchasing an enterprise platform when teams compare token costs to licensing fees. But the analysis rarely accounts for the full picture, where every custom capability becomes the organization's responsibility. Meanwhile, purpose-built platforms centralize governance, security, monitoring, and operational knowledge so that every new use case benefits from existing infrastructure rather than recreating it from scratch. The strategic question to ask is where custom development creates competitive advantage and where purpose-built platforms deliver better economics.
Navigating common RAG failures
When a RAG system returns a wrong answer, most teams look first at the prompt. In the majority of enterprise deployments, the prompt is not where the problem lives. Retrieval is. Jon Knisley, Director of AI Value Management, describes how three failure modes account for most of what goes wrong: relevance failures, coverage failures, and noise failures.
All three trace back to document preparation quality, not retrieval or prompting. When ingestion does not preserve structure, tables lose their column headers, exception clauses separate from the rules they modify, and the information needed to answer a query never becomes a coherent, retrievable unit. Testing retrieval as its own component and then tracing every failure back to where the corpus was built is the difference between AI that works and AI that you hope works.
The human intelligence series: meet the people behind ABBYY AI
To bring this principle to life in a new way, ABBYY launched a YouTube series called "Human Intelligence Driving ABBYY Innovation." Each episode features the people who shape how ABBYY designs, delivers, and stands behind its solutions.
Neil Murphy, who leads ABBYY's global revenue organization, describes what motivates him after more than a decade in the role: watching the people around him succeed. He builds teams of strong leaders who each bring distinct capabilities, then creates the conditions for them to use those capabilities. When people feel trusted, that trust ripples outward into how solutions get engineered and how customers experience the results.
Ulf Persson, ABBYY's CEO, makes a similar point about leadership. Effective leadership is rooted in collective ambition and the determination to turn individual strengths into shared success. Automation does not build trust; organizations earn trust through transparency, follow-through, and a consistent commitment to customer outcomes. At ABBYY, that means teams worldwide actively incorporate client feedback, remain accountable at every stage, and lead as trusted advisors rather than transaction-oriented vendors.
Hopeful Owitti, VP of Customer Excellence, highlights stories of individuals who combine expertise and creativity to advance AI technology, emphasizing the importance of human oversight in ensuring ethical and effective solutions. Through these narratives, Hopeful demonstrates how purposeful collaboration between humans and AI can address real-world challenges with precision and reliability.
Paula Sanders, SVP of Pre and Post Sales, shares insights into ABBYY's decades of experience in OCR, NLP, and machine learning. They underscore the value of integrating human expertise to ensure AI systems are transparent, compliant, and aligned with business needs. Paula illustrates how human-driven innovation enables ABBYY to consistently deliver exceptional results in an evolving technological landscape.
These leaders see AI hype as one of the most significant risks facing enterprise decision makers. The antidote is not cynicism about AI's potential. It is the rigor to ask hard questions, demand evidence, and commit to solutions that withstand scrutiny. Purpose-built AI, grounded in 35 years of expertise in OCR, NLP, and machine learning, delivers results that are predictable, accurate, and compliant. That track record exists because of the humans who built and maintained it.
What purpose-built AI with human intelligence actually delivers
The real-world results of this approach are measurable. ABBYY customers across financial services, logistics, insurance, and banking have seen:
- An Irish food manufacturer reduced customs clearance times from an hour to five minutes.
- A global brewery increased its touchless order processing rate to 92% and saved more than 140 labor hours per month.
- A healthcare provider saved $6 million in annual late payment costs.
- A financial services firm achieved a 99% compliance first-time pass rate, 40% faster document processing, and a 15% reduction in costs and manual errors.
- A global logistics company saw 40% greater document processing efficiency and 35% straight-through invoice processing.
These outcomes did not happen because organizations deployed a general-purpose model and hoped for the best. They happened because purpose-built AI, designed by people with deep domain expertise, was applied to specific, well-defined workflows with appropriate human oversight at every stage.
The operating principle for the next phase of enterprise AI
The first phase of enterprise AI adoption was largely a race to apply LLMs wherever possible. That phase produced important learning and a necessary shift in organizational mindset. The next phase is about discipline: deciding where AI should be applied, not just where it could be.
That decision requires human intelligence. It requires leaders who understand the difference between a proof of concept and a proof of value, engineers who know how document quality affects model performance, compliance teams who can align AI workflows with standards like BSI C5 and the EU AI Act, and fraud analysts who provide the contextual judgment that automation cannot replicate.
ABBYY has always believed that the most powerful AI is AI guided by people who are invested in the outcome. Every episode of "Human Intelligence Driving ABBYY Innovation" is built on that belief. The series introduces the developers, strategists, and business leaders who ensure ABBYY's AI remains ethical, transparent, and accurate, because technology without human accountability is not a competitive advantage. It is a liability.
Every enterprise faces daily decisions that affect long-term success. Make those decisions based on a full understanding of what AI can and cannot do on its own. The organizations that get this right will not just control costs more effectively. They will build systems that are easier to govern, easier to scale, and ultimately more trustworthy.
Subscribe to the "Human Intelligence Driving ABBYY Innovation" series and watch the latest episodes on YouTube.
