---
title: The tools were fine, the rollout was not
---

| View in browser Hello from ABBYY's AI Team AI had a rough month with its performance reviews. Anthropic started watermarking every Claude output, claims adjusters are trashing AI tools with the enthusiasm of a Rotten Tomatoes pile-on, and McKinsey found that 90% of organizations use AI while only six percent see actual profit from it. Notice the pattern? Adoption keeps outrunning accountability, like an outie racing through Lumon's hallways with no idea what the innie signed up for. This month, the team dug into all three stories: what provenance really requires, why "AI backlash" is usually an implementation problem wearing a disguise, and where the productivity-to-profit gap breaks down.   Read on for the full picture of what’s been moving in the AI world this month.     Max Vermeir, VP of AI Strategy AI News You Need to Know: Expert Take Jonathan Knisley, Global Process AI Lead, ABBYY 90% Adopt AI. Only 6% Really Profit From It. There's a widening gap between adoption and enterprise value in McKinsey & Company's latest State of AI survey. While 90% of organizations report using AI, only 37% see EBIT-level impact and just 6% qualify as high performers with more than 5% EBIT improvement attributed to AI.   AI use continues to boost individual performance at work. 80% report that AI has improved their individual productivity, and 50% report that AI helps them make better decisions.   As AI becomes embedded in more business processes and employees become increasingly fluent in using it, the organizations that translate individual productivity gains into lasting enterprise-level financial performance are likely to be those that transform their businesses, not just adopt AI tools. Learn more  > Slavena Hristova, Director of Product Marketing — Document AI, ABBYY When Bad AI Decisions Kill Good Automation "You Know Who Really Hates AI? Insurance Claims Adjusters" caught my eye. How did we get here?   A recent analysis found that 98% of claims adjusters mentioning AI expressed negative sentiment, citing misclassified claims, faulty summaries and ultimately more work instead of less. A terrible outcome for a technology with so much potential.   But it quickly becomes clear that this is less about poor technology than poor implementation.   Especially when errors can materially affect someone's livelihood, people deserve better than organizations falling for hype and pushing the wrong technology into production.   We talk a lot about upskilling employees to work with AI, and rightly so. But organizations also need to become more discerning buyers and users of AI, with a clear understanding of what different technologies are designed to do, where they fail, and what controls they require.   Otherwise, poor implementations risk creating AI fatigue and distrust. We waste more than time and money. We lose trust, momentum and opportunities to automate work that genuinely should be automated. Learn more  > Dr. Marlene Wolfgruber, Product Marketing Lead AI, ABBYY When AI Leaves a Trace: What Anthropic's Watermarks Mean for Document Governance Anthropic's decision to watermark every Claude output to me reads like the confirmation of a broader pattern: AI systems that touch documents are being expected to leave provenance behind. In document-heavy workflows, that expectation has been building for some time. Teams already pair OCR and layout-aware parsing with structured exports so that what enters an AI pipeline and what leaves it can be traced, validated, and, when needed, challenged. Anthropic's watermarks and C2PA-style metadata simply extend that logic to the generation step: the model's contribution becomes another link in the chain.   That is highly relevant for organizations running high-volume document processing: Because if ingestion preserves the structure and provenance, but the generation doesn't, the chain breaks exactly where risk concentrates: in the final text, summary, or decision that humans act on. Watermarking is the first step to fix that.   My takeaway: governance is moving from abstract commitments about "responsible AI" to concrete, verifiable signals embedded in the outputs themselves. Watermarks, signed metadata, and structured document formats all point in the same direction: every artifact should carry enough context to show how it was created and transformed. This direction is the right step towards a more transparent and reliable usage of AI, especially in the context of automated decision making. Learn more  > Fun Facts   Explain It to Me Like I'm 5 Years Old   Today's AI can outperform humans at extraordinarily complex tasks, yet still struggle with things a five-year-old finds easy. In one recent benchmark of basic visual and reasoning skills, five-year-olds scored 96%, while the best tested AI model managed just 52%.   Source: AAAI > AI by the Numbers By 2028, the average Fortune 500 company will have more than 150,000 agents, according to Gartner. Microsoft alone has already deployed more than 500,000 internal agents to help employees streamline workflows. Source: CIO > ABBYY in the News BSI C5: Warum das Testat für Cloud-Vertrauen so wichtig geworden ist \| safety-security.ch (German) > Processus de vérification des prêts hypothécaires grâce à l'IA documentaire et aux LLM \| Finyear (French) > The Hidden Security Risk in Document Redaction \| SecurityBrief > ABBYY in Action: Innovation Spotlight Generative AI + IDP Extracts Data from Lease Agreements with 82% Accuracy A fast-food franchise with 40,000 global locations needed expert advice to improve a tedious business process related to its commercial leasing agreements. Find out how Ashling solved this challenge for the fast-food franchise, developing a solution that leveraged ABBYY Document AI with SS&C Blue Prism technology and GPT-4 Turbo. Learn more What 75+ Developers Taught Me About LLMs and Document Processing At the ABBYY Ascend Developers Conference (DevCon) in Bengaluru, India, we ran a hands-on Prompt Activity session, and the unexpected results taught us the biggest lesson. With the same tenant, same connection, same document, and same instructions—the outcome changed from person to person. Why was that? Read the blog post to find out the answer to that question—the real learning from the session. Read now Quick Resources & Events Webinar: FlexiCapture customer? Assure it with Reveille Are you a FlexiCapture customer in the Americas? Join ABBYY and Reveille on September 23 for a live, 45-minute session covering end-to-end visibility, self-healing automation, and SLA reporting. Learn about customers who reported 95% SLA attainment and over 50% fewer trouble tickets. Even if you can't make the live event, register to receive the playback.   Register now  > ABBYY Circles: Document Fraud & KYC Modernization Calling financial services leaders in the Charlotte and Toronto area. Join ABBYY and Resistant AI for an exclusive ABBYY Circles event on document fraud detection and KYC modernization. See live demos, gain a pKYC readiness check, and leave with your own automation shortlist.   Request your seat in Charlotte (Oct. 7)  > Request your seat in Toronto (Oct. 8)  > © 2026 ABBYY Solutions Limited  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 Solutions Limited, Legal address: 25 Aphrodite Street, Office 204 (2nd floor), 1060, Nicosia, Cyprus www.ABBYY.com/company/contact-us  Director of ABBYY Solutions: Anna Terzi - Registered at Registrar of Companies, Nicosia, Cyprus HE 131582  Vat Reg. No.: CY10131582D  This email was sent to  info@abbyy.com.  Privacy Policy  You may unsubscribe from future emails at any time. Should you have any questions don't hesitate to contact us. Unsubscribe Manage preferences |
| --- |