Enterprise AI World 2025, co-located with KMWorld 2025, offered a clear signal this year: the era of “drop a chatbot on the intranet and call it transformation” is over. The conversations shifted toward AI that sits inside real work—capturing tacit knowledge, restructuring workflows, building organizational memory, and reframing what humans are actually good for.
Across the keynotes and sessions, three patterns kept repeating:
- AI is moving from content generator to decision partner and team member.
- Knowledge, structured, contextual, and connected, is becoming the real platform.
- Policy and practice, including leadership and workforce design, are now the rate-limiting steps, not models.
Underneath those patterns, each talk brought its own angle, from conversational surrogate agents and quantum-inspired succession planning to graph-powered RAG and agent ecosystems spanning Amazon, Google, and Microsoft.
Pattern 1: AI as Collective Intelligence, Not Just Automation
David Baltaxe from Unanimous AI began his talk with a simple observation: organizations still treat people like data points rather than data processors. Polls, surveys, and forms strip away the very thing that makes a workforce valuable: its ability to think together in real time.
The company’s Thinkscape® product, which employs its Hyperchat AI™ and Swarm AI ® technologies, uses “conversational surrogate agents” embedded in small groups to scale discussion. Agents listen to breakout conversations, extract arguments and rationales, and share them with their peers in other groups. The agents intentionally look for conflict and opposing views, not consensus, and then feed those back into the rooms to keep thinking sharp. The result is not a giant webinar, but a hundred-person conversation that still feels like a five-person working session.
That same theme of human-plus-AI thinking together surfaced in the presentation by Microsoft’s Ross Smith, titled “Deploying AI in the Organization.” Smith has built “Calliope,” a generative AI muse that acts as rehearsal partner, adversary, and advisory council. He uses it to simulate contentious meetings, rehearse presentations, and run scenario-style debates among a synthetic board of roughly fifty “voices” drawn from philosophy, psychology, business, and literature.
Calliope isn’t there to replace judgment. It compresses hours of reading and internal debate into minutes of dialogue so humans arrive at the room more prepared.
Lee Rainie’s work at Elon University added another layer. His research tracks how AI is changing human traits and behavior. Experts he surveyed see AI as likely to enhance curiosity, creativity, and decision-making, but as a net negative on deeper capacities like critical thinking, metacognition, empathy, and moral judgment. The irony: the traits organizations say they want to protect are precisely the ones most at risk of atrophy if they hand too much thinking to AI.
Taken together, these talks point to a simple design principle: treat AI as a catalyst for richer human interaction, not a substitute. Build systems that surface disagreement and nuance, systems that demand reasons and rationales, not just checkboxes and click-throughs.
Pattern 2: From LLMs to Agents
Several sessions made a sharp distinction between large language models and agents. The panel with leaders from AWS, Legion, and Feith Systems hammered this home. An LLM is one component of an agent, responsible for language and reasoning. The agent itself wraps that model with memory, tools, policies, permissions, and audit trails.
That distinction matters because organizations keep buying “chatbots” and wondering why they don’t see value. Generic Q&A interfaces without a specific job usually become just another SaaS cost center. The panelists argued that real wins come from tightly scoped agentic workflows aligned to hard costs, such as shortening a 27-day process to nine hours, cutting overtime, or eliminating backlogs, not from generic assistants floating in a browser tab.
My own session on “The Future of Work in a World of AI Agents” offered a map for this shift. I framed agents across a spectrum of agency, from minimal (scripts, RPA) to collective (multi-agent ecosystems). My classification schema walked the audience through agents along a continuum of increases in autonomy, reasoning, memory, learning, and collaboration as systems move from simple automations to coordinating swarms of agents across domains.
My “Big 3” slide showed that Amazon (Nova/Bedrock), Alphabet (Gemini/Vertex), and Microsoft (Magma/Azure/Copilot) are converging on remarkably similar agent stacks. Each offers:
- Pre-built agents (e.g., Amazon Q, Gemini Code Assistant, Dynamics 365 Agents)
- Agent dev environments (Bedrock Agents, Vertex Agent Builder, Copilot Studio)
- Marketplaces and orchestrators
- Early moves toward interoperability through A2A (agent-to-agent communication) and MCP-style tool/context layers
The strategic implication: organizations can’t just pick a model anymore; they’re picking an agent ecosystem.
Cohere’s Martin Kon delivered an opening keynote titled, “Unlocking Enterprise Value for Knowledge Work,” that highlighted necessary constraints on how organizations should adopt those ecosystems without getting lost in the hype. He argued that the real economic transformation will come from enterprise AI, not consumer chatbots, and laid out a pragmatic path:
- Build excellent search and retrieval across existing systems, inside security and sovereignty boundaries.
- Teach AI to use existing tools and systems rather than rebuilding them.
- Only then move to true agents orchestrating multi-step workflows.
That methodical path lines up well with the panel’s plea to avoid “death by a thousand POCs” and instead push at least one use case all the way to production scale to build institutional muscle.
Ross Smith’s “6Ds” model added a deployment rhythm: Discover, Design, Develop, Diagnose, Deploy, Detect/Monitor. His insistence on responsible AI reviews, inclusive early adopters, and synthetic transactions to catch drift reflects a maturing discipline that sees agents as long-lived systems, not throwaway experiments.
Pattern 3: Knowledge as Infrastructure—Graphs, RAG, and Tacit Capture
As agents start to impinge on operating systems and operating models, it’s important to refocus on the core elements of knowledge management, which many organizations have given short shrift over the years. For AI to work in enterprises, it requires enterprise data to consume and incorporate into its models. Many failures in AI don’t arise from flaws in how AI works, but from the messiness of the enterprise content exposed as ingestion pipelines that return poor results that make it hard for end users to build trust.
Zorina Alliata, Principal AI Strategist at Amazon, and Theresa Minton-Eversole, Project Manager, Net Impact, positioned knowledge graphs as organizational memory, a way to encode entities and relationships so AI can reason with context, not just text strings. Their framework distinguished the following knowledge types:
- Persistent knowledge: Manuals, slide decks, videos—relatively easy to ingest.
- Transient knowledge: Meetings, chats, e-mails—captured increasingly by assistants.
- Tacit knowledge: The intuition and shortcuts of experts—still the hardest part.
Their tacit capture case study was refreshingly concrete: record a senior operator working for a full day, then use Gemini 2.5 to interpret the video, extract decision rules, and auto-draft training materials. The expert doesn’t have to become a writer; their behavior becomes the raw data.
They applied that same approach to succession planning. Using a graph of 187 employees and 300 projects, they compared classical Louvain community detection with a quantum-inspired algorithm. The algorithms disagreed on successors for roughly 90% of employees; human reviewers consistently preferred the quantum-inspired matches, which often surfaced cross-department candidates the old methods missed.
Andreas Blumauer from Graphwise extended the graph story. He argued that LLMs and vector RAG are not enough for high-stakes use cases like compliance and technical knowledge management. The answer, in his view, is Graph RAG—retrieval augmented generation using a semantic layer that understands entities and relations.
His case study with an engine manufacturer showed accuracy jumping from about 30% to 80% when a modest knowledge graph was added to the mix. EY’s global rollout of a knowledge graph for 300,000 employees underscored that large enterprises are already betting on semantic backbones to improve reuse and reduce hallucinations.
Blumauer also reframed the often-overlooked human role he called taxologists, a conflation of taxonomists and ontologists, who design the initial 1% of domain models that power the automated 99% of graph growth. His “two-flywheel” diagram links a technical inner loop (data scientists iterating on models and graphs) with a business outer loop (executives focused on ROI, trust, and resilience). Knowledge scientists sit between the two, translating between semantics and the execution of strategy.
Alliata’s buy-vs-build analysis mirrored that thinking. Building an AI-enabled knowledge graph platform can run $500k–$2M and take 12–24 months; buying a platform lands closer to $50k–$300k in the first year, with trade-offs in customization and lock-in. In a market changing this fast, she recommended hybrid models, modular architectures, and managed cloud services to keep options open.
Pattern 4: Culture, Leadership, and the Emergent Meritocracy
The strongest undercurrent across Enterprise AI World wasn’t model talk—it was anxiety and opportunity around the workforce.
Rainie’s data showed that 57% of US adults already use language models, with the primary use case tilted toward personal enrichment and social interaction rather than business productivity. That lands AI in a strange place: a tool that is simultaneously intimate and invisible, shaping cognition even when people don’t label it as AI.
He framed this as an “intimacy pivot”: systems moving from optimizing for engagement to optimizing for companionship and dependency. That has significant implications for work as agents become ever-present colleagues, not just tools.
My own “agent-human work experience” section tried to anticipate what that looks like day-to-day: digital colleagues on teams; persistent multi-threaded collaboration; work happening 24/7 as agents prepare briefs and draft decisions overnight; new power dynamics as those who orchestrate agents gain leverage over those who don’t. My slide titled “What gets done while you sleep” lists the kind of preparatory work agents will take on, such as triaging mail, managing orders within supply chains, assembling slide decks, and simulating strategic options.
Ross Smith and the implementation panel both spoke about a flattening and elevating of roles. Routine tasks disappear; expectations rise. People are being pushed to do overnight what once took weeks, with AI as the justification. That creates stress, but also the outline of a new meritocracy: those who learn to work with agents, designing prompts, critiquing outputs, and connecting systems, become disproportionately valuable.
The panel also laid responsibility firmly at leadership’s feet. Leaders are being asked to be AI champions, yet many remain cautious, waiting for clarity, rather than effectively navigating uncertainty. Meanwhile, AI budgets demand visible action. Their advice:
- Frame AI initiatives around concrete business problems and hard costs.
- Skip generic innovation theater; go directly to a business unit with a painful workflow and fix it.
- Make HR, Learning & Development, and Knowledge Management central to transformation rather than back-office spectators.
Smith added a less punitive twist via gamification: develop internal leaderboards, certifications, and play-based reinforcement to make experimentation with AI less threatening and more rewarding.
So What Should Organizations Actually Do?
The conference didn’t offer one blueprint. It did, however, sketch a set of converging practices that sensible organizations can adopt now.
- Stop treating people as rows in a dataset. Employ systems like thinkscape® that use AI to scale deliberation, not just data collection. Build in mechanisms that surface disagreement, not just average it away.
- Treat agents as long-lived products, not experiments. Use frameworks like the 6Ds, clear OKRs, and robust monitoring. Start with one high-value workflow, run it to production scale, gather scar tissue, and then replicate.
- Invest in a semantic backbone. Taxonomies, ontologies, and knowledge graphs are not optional for serious AI. They are the substrate that enables Graph RAG, cross-silo retrieval, and governance. Hire or grow taxologists and knowledge scientists who can sit between data science and business execution.
- Use AI to inexpensively capture tacit knowledge. Use multimodal models to turn real work like video, screen recordings, and conversations into structured insights. Let experts do the work while AI observes and drafts. Reserve scarce human time for validation, not authorship.
- Differentiate generic AI from “alpha-generating” AI. Accept that generic features will be bundled into productivity suites and SaaS. Focus custom investments where proprietary data and workflows create enduring advantage.
- Design for the emergent meritocracy. Explicitly plan for new roles around agents, from orchestration and monitoring to ethics and governance. Build learning paths and incentives so the people closest to the work become AI-literate co-designers, not passive recipients.
- Plan for intimacy and dependency. As agents become ever more embedded in daily life, actively protect critical thinking, metacognition, and moral judgment. Measure them. Train for them. Do not assume they survive by default.
Enterprise AI World 2025 didn’t resolve the open questions about jobs, agency, or the long arc of automation. It did something more pragmatic: it showed how quickly AI is moving from novelty to infrastructure, from chatbots at the edge to agents in the middle of every important workflow.
Organizations now face a choice. They can keep adding bots to websites and running small, disconnected pilots. Or they can acknowledge that AI is becoming part of the fabric of knowledge, work, and leadership, and start redesigning that fabric with intent, before someone asks an agent do it for them.
His work has been featured in Harvard Business Review, Fast Company, Wired, NASA Ask, and dozens of other magazines and websites. His analysis on the future of work can be found at SeriousInsights.net. He is also the author of Understanding Artificial Intelligence, Cyberlife, Rethinking Smart Objects, and Empower Business with Generative AI.
Rasmus regularly speaks on the future of work at events such as Comic-Con International, Wondercon, Impact, Enterprise AI World, Worktech, CLO Symposium, KMWorld, AAAI, Computers in Libraries, Microsoft’s Building the Future, Educause, Expo Capital Humano, Devlearn, Internet Librarian, CAMEX, EduComm, and Sourcing Summit Europe.
As an Affiliate Instructor, Rasmus teaches scenario planning at the University of Washington and is an Associate Adjunct Professor at Bellevue College.
