Susan Hanley

Every claim represents someone waiting for an answer. Behind that answer is a claims team working through documents, policy details, tasks, and decisions. That’s why I’m excited about Appian Connected #Claims 2.0. It brings those pieces together so claims professionals can spend more time moving claims forward. Here’s what stands out to me: 🔎 A connected workspace Give adjusters, examiners, and supervisors a shared view of claim information, documents, tasks, and communications—from intake through settlement. 🤖 AI assistance built around everyday work Assess loss descriptions against policy coverage, with additional AI modules to summarize documents and comments or ask questions about claims data. Help teams get oriented faster and focus their expertise where it matters. ⚙️ Workflows that can adapt An embedded Control Panel gives business users no-code tools to configure workflows, with conditional tasks and flexible assignments to support different claim circumstances. 🧩 Capabilities that fit the business Extend the solution with modules for Automobile, Home Owners, Life, and Workers Compensation, plus APIs for claim creation and document uploads. For me, the opportunity is straightforward: less time gathering context and coordinating handoffs, and more time making informed decisions and helping policyholders. 💬 Where does your claims process lose the most time today—finding information, coordinating work, or adapting to change? Learn more: https://lnkd.in/e9zd2gAz | https://lnkd.in/eRVvqrE6 #Appian #ConnectedClaims #Insurance #ClaimsManagement #AI Aidan Reilly Isaac Dweck Skandha Gopalan Yin He

13

1 Comment

The most-shared AI launch this week isn't a smarter chatbot. It's a model that refuses to write. Production AI agents don't need more prose.. They need thousands of small judgments per minute; is this ticket urgent, which model handles this, is this command dangerous, does this passage answer the query? We've been sending every one to a frontier LLM. Wait for tokens. Parse text. Retry on shape errors. Pay for output we didn't need. JEV flips the interface. Code keeps the branching. The model supplies judgment code can't calculate. Define the answers - choice, score, yes/no get back a typed decision with calibrated probability. 📌 JEV vs LLM ↓ → Interface JEV: typed answer from a schema you define LLM: free-form text you parse → Latency JEV: single-pass, parallel questions per request LLM: token-by-token, sequential → Uncertainty JEV: calibrated probability on every answer LLM: needs a separate judge model → Failure mode JEV: confidently wrong within its schema LLM: hallucinates outside any schema → Pricing JEV: input only $0.042 / M tokens, output free LLM: input + output 📌 How to use it in your agent: 1. Find one decision going to a frontier LLM - routing, risk gating, judge, reranking. 2. Write the rubric before calling the model. 3. Shadow-mode it. Log confidence and outcome. 4. Set thresholds auto-act high, escalate mid, review low. 5. Automate the safest branch. Expand after numbers hold. Frontier LLMs plan, write, reason. JEV-class models route, gate, score, verify. That's the two-model stack production agents needed. If your agent runs every decision through a frontier model, JEV isn't your problem. Your architecture is. Which decision node in your agent should've been a classifier six months ago? 👇 📘 Grab your copy → The Human-Agent Orchestrator: https://amzn.to/4gEIzmv Save 💾 React 👍 Share ♻️ - follow for daily AI agent breakdowns

1

As part of building AI tools at the Program Integrity Alliance, we need more control over document ingestion. We apply community-informed taxonomies for better filtering, image processing, so that this important data can be included, and deep linking to the exact pages - features not offered by public search engines. A great side effect of this AI work is being able to also ship our GovQuery search engine, which has been extended this week to include Executive Orders for the last 6 presidents. It's been great to see people start using this resource! Especially in the case where the original content has been taken offline due to federal defunding. When building AI solutions, the work to streamline data processing can lead to more basic tools like search, which can be very useful on their own. More to come soon!

3

Happy New Year! I'm excited to share something I've been working on for a while: my new book on modernizing legacy technology systems with AI. I've worked for over 20 years across federal, state, and local government. At every level, I've seen agencies struggle to modernize legacy systems. These efforts almost always end in failure, and it took me a while to understand why. The system is set up to encourage large, big-bang modernization projects that focus almost exclusively on one thing: replacing old code with new code. Requirements documents and architecture diagrams get created along the way, but they exist solely to support that conversion process. If these projects finish successfully (and that's a really big if), the new code becomes the only source of truth for how the system works. The legacy clock is reset and the countdown to the next crisis begins. But there is now a better way. Behind all the hype about artificial intelligence over the past few years, a real opportunity has emerged. AI tools can enable a fundamentally different approach to legacy modernization, one that addresses the risks that have contributed to so many past failures. My book, "The SpecOps Method," describes this new approach. It shares the same DNA with practices like behavior-driven development and incorporates more recent advances in specification-driven development using AI tools. The book provides a conceptual framework for understanding why this approach works where others have failed, along with practical guidance for getting started on your own own projects. https://lnkd.in/g6AYDFue I hope you'll read it. I hope at least some of it will help anyone working to support or modernize legacy systems. I hope you'll share lessons learned so that we can finally break free from from the legacy cycle.

51

12 Comments

This Salesforce piece on the “last‑mile challenge” of enterprise AI really resonated with me. While consumer AI adoption has exploded, most organizations are still struggling to move from impressive pilots to AI that actually works at scale in day‑to‑day operations. What stood out is the acknowledgment that *agentic AI is as much a change‑management challenge as it is a technical one*. Enterprise adoption slows not because the models don’t work, but because scaling AI requires: ☑️internal alignment and executive sponsorship, ☑️redesigned processes and operating models, and ☑️trust, governance, and clarity around how humans and AI work together. We’ve seen this pattern before with email, cloud, and other transformative technologies - the real work happens in the “last mile,” where behavior, mindset, incentives, and capability must shift alongside the tech. If organizations want AI to deliver real value, change management can’t be an afterthought. It needs to be embedded from day one: preparing leaders, enabling teams, redesigning workflows, and addressing the human questions of trust, accountability, and adoption. Technology may unlock possibility - but *people and change unlock impact*. 🔗 Worth a read: How Enterprise AI Will Overcome the Last Mile https://lnkd.in/duzTUw_a #ChangeManagement #EnterpriseAI #AIAdoption #FutureOfWork #HumanCenteredAI #DigitalTransformation

5

Improve prediction accuracy, visualize trends, reduces manual analysis in Opex and Revenue planning, leveraging embedded AI in EPM cloud.

2

Most people look at federal procurement as a mess of paperwork and bureaucratic limbo. Author sees it differently: it’s the primary engine for turning priorities into reality. The problem isn't just the rules; it’s also a leadership gap.

14