Remember that lesson every engineer learns the hard way? Garbage in, garbage out. It never went away. It just scaled up.
This month, we're looking at what happens when that problem meets agentic AI, token economics, and document workflows that look fine on the surface.
Jon is dialing into Westworld for enterprise workflows. Microsoft Research tested 19 frontier models on long document workflows, and even the best corrupted 25% of content. AI playing telephone, while no one catches the drift.
Then Slavena calls time on the "throw an LLM at it" reflex. Enterprises are livid about paying for tokens that create zero value. Her case for hybrid AI is simple: match the tool to the task, not the hype.
Finally, Marlene shines a light on the hidden layer powering RAG and agentic systems. It is not model size. It is input quality. Nail the data foundation, and you get real automation. Miss it, and you are just amplifying noise at scale.
Microsoft Research tested 19 large language models on long, delegated document workflows across 52 professional domains. Even the best-performing frontier models corrupted an average of 25% of document content. Across all models tested, average degradation reached 50%. Giving models agentic tool access made no measurable difference.
The most dangerous AI failure in the enterprise right now does not crash a system, trigger an alert or produce obviously garbled text. Instead, it produces a document that looks polished, reads coherently and travels through approvals while the meaning has shifted underneath.
The Reckoning for "Throw an LLM at It" Has Arrived
Between Palantir CEO Alex Karp telling CNBC that enterprises are "livid" about paying for tokens that "create no value," and analysts like Apollo's Torsten Slok warning that cheaper tokens are driving spending up, not down, the last few months have made one thing clear: throwing an LLM at any task will not be acceptable for much longer.
The LinkedIn eulogies for every category AI supposedly killed can retire along with it. What's replacing that reflex is a harder question every enterprise now has to answer: does this task actually need a large language model, or does it need something purpose-built, auditable, and cheaper to run at scale? That's the case for hybrid AI—using the right technology for the task, not the flashiest one.
Tokenmaxxing may be over. The harder discipline—matching the tool to the job before the invoice forces the conversation—is just getting started.
We keep hearing the same lesson because it keeps proving true: garbage in, garbage out. From databases to early AI to today's RAG and agentic systems, the principle has never changed, only the consequences have.
What is different now is scale. Agentic AI does not stop at a single response; it makes decisions, passes information along, and builds on previous outputs. So when extraction is wrong, the error gets amplified downstream.
To avoid that, businesses do not just need smarter prompts or larger models, they need a reliable data layer that can transform messy documents into structured, trustworthy inputs. In an AI pipeline, that foundation is what separates real automation from accelerated noise.
Fun Fact
Bots Have Officially Taken Over
Website traffic from AI agents and bots has eclipsed its human-generated counterpart for the first time.
The organizations in IBM's new study that have the greatest control across their AI stack protect 55% more operating profit from AI-driven disruption than organizations with less control.
How intelligent automation is transforming insurance claims processing | Journal Du Net (French) > The article in Journal Du Net explains how intelligent automation is transforming insurance claims processing by combining technologies such as Intelligent Document Processing (IDP), OCR, AI, NLP, computer vision, machine learning, RPA, and process mining. These technologies automate document handling, data extraction, fraud detection, and end-to-end workflows, enabling insurers to process claims faster, more accurately, and at lower cost.
Jon Knisley on why many AI failures stem from using LLMs where simpler tools would work better | Computer Weekly (German) > Jon Knisley highlights in Computer Weekly that many AI failures stem from using large language models (LLMs) where simpler, more appropriate technologies would deliver better results. He shares examples of organizations that initially relied on LLMs for claims processing and document extraction but found them costly, inconsistent, and inaccurate. By replacing or complementing LLMs with business rules, regex, purpose-built document processing, and human review, they achieved faster, more accurate, and more cost-effective outcomes.
ABBYY in Action: Innovation Spotlight
Agentic AI in Banking: Opportunity, Risk, and the Growing Importance of Model Risk Management
Model Risk Management in banking is a reinvention where the document-centered nature of banking is no longer an impediment but a foundation. When augmented with intelligent Document AI, banks can innovate confidently while maintaining the trust, transparency, and discipline that regulators and customers expect.
AI in Practice: Real-World Applications & Case Studies
How to Achieve Accurate, Reliable KYC at Scale with Document AI
KYC programs need to be fast and accurate. Miss a red flag and you are exposed. Move too slow and customers leave. The latest ABBYY AI at Work video shows you how ABBYY Document AI handles the heavy lifting.
Ascend in Action: Ashling Partners Scale Global Invoice Automation
From ABBYY Ascend Nashville, Morgan Conque of Ashling Partners shares how her team is pushing the boundaries of intelligent automation with ABBYY that scale across global finance and accounting processes, including complex, multilingual invoice workflows.
Event: Technology Executive Summit | AI & Intelligent Automation | Palo Alto, CA – Wednesday, August 5th | 6:00 PM PT
This exclusive event brings together senior technology leaders for expert insights on AI, intelligent automation, and emerging technologies shaping the modern enterprise. Attendees can expect high-level discussion, strategic networking, and actionable takeaways in an intimate dinner-roundtable setting. If you are in the San Francisco Bay Area, we invite you to come by.
Webinar: Stop Document Fraud at Ingestion, Before It Moves Downstream – Wednesday, July 29 | 1:00 PM ET / 10:00 AM PT
Learn how ABBYY and Resistant AI help financial services teams stop document fraud at ingestion—before forged data contaminates onboarding, lending, claims, and downstream decision-making. This webinar covers today's document fraud landscape and includes a live demo of an integrated approach to authenticity verification, forensic detection, and more confident automation.
Discover how purpose-built Document AI turns business-critical documents into trusted, actionable data—powering intelligent automation, informed decision-making, and next-generation AI use cases.
You have received this e-mail because you are an ABBYY customer or partner, have requested further information on our products, solutions or services, registered for an ABBYY trial version, and/or agreed to receive news from ABBYY.
This e-mail is sent to you by:
ABBYY USA Software House Inc., 600 Congress Avenue, Suite 15015 Austin, Texas 78701
