Skip to content
TheChatArchiveThe Chat Archive
Kimikimi-k2-0711-preview· Moonshot AIHugging Face LMArena Human Preference 140K

where is Agentic AI in the enterprise space?

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Mistralmistral-medium-2505’s answer.
  1. User

    where is Agentic AI in the enterprise space?

  2. Kimi Assistant

    Agentic AI is still in the “early-adopter / proof-of-concept” phase for most enterprises, but the line between “cool demo” and “production workload” is starting to blur in a handful of verticals and use-cases. Below is a snapshot of where things stand today (mid-2024).

    1. Definition check
      • Agentic AI = systems that can autonomously plan, decompose, and execute multi-step tasks, call tools or APIs, and adapt their plan in real time.
      • Not the same as a retrieval-augmented chatbot or a simple LLM wrapper; it implies persistent state, memory, and goal-seeking behavior.

    2. Maturity curve (enterprise)

    Stage 0 – R&D / Hackathons
    • Most Fortune 500s have at least one internal “agent hackathon” repo.
    • Common stack: LangGraph, CrewAI, Microsoft Autogen, LlamaIndex workflows, or home-grown orchestrators on top of GPT-4 / Claude / Llama-2.

    Stage 1 – Pilot use-cases (2023 → now)
    • Tier-1 tech companies (Stripe, Shopify, Atlassian, Uber) running pilots for:
    – Developer-experience agents (auto-fix unit tests, generate PR descriptions, rollback triage).
    – Customer-support agents that open tickets, query knowledge bases, and issue refunds.
    • Banks and insurers experimenting with “regulatory reporting agents” that gather data from 10–15 internal systems and draft first-pass filings.
    • Life-science firms using agents to chain together wet-lab scheduling, ELN queries, and literature search to propose next-day experiment plans.

    Stage 2 – Limited production (first half 2024)
    • Klarna’s well-publicized AI assistant now handles two-thirds of customer-service chats; behind the scenes it is an agentic flow that can call order-management, payment, and logistics APIs.
    • Doordash “Dasher Support Agent” can cancel, reschedule, and issue partial refunds without human approval for ~30 % of tier-1 cases.
    • Siemens Energy production-grade agent monitors sensor streams, opens work orders in SAP, and drafts inspection checklists for field engineers.

    Stage 3 – Scale-out (late 2024 → 2025)
    • Roadmaps show intent to move from “one agent per workflow” to “agent mesh” (multi-agent teams with role specialization).
    • Microsoft Copilot Studio, Salesforce Agentforce, ServiceNow AI Agents, and Amazon Bedrock Agents are all shipping managed orchestration layers aimed at enterprise governance, audit trails, and RBAC.

    1. Where the value is showing up first
    Function / ProcessTypical ROI driver
    L1 Customer Support20-40 % ticket deflection, 50-70 % faster resolution
    Internal IT & DevOpsMean-time-to-restore cut by 30-60 % via auto-runbooks
    Finance Close & Reporting3-5 day compression in monthly close
    Supply-Chain Disruption MgmtReal-time re-routing saves 1-3 % of COGS
    Regulatory / Compliance60-80 % reduction in first-draft document prep time
    1. Gating factors still slowing mass adoption

    • Tooling immaturity: debugging an agent that took 14 steps and failed on step 11 is still painful.
    • Security & compliance: agents need long-lived tokens, fine-grained scopes, and audit logs that satisfy SOC-2 / HIPAA / FedRAMP.
    • Cost volatility: chaining 20 LLM calls per task can be 10× more expensive than a deterministic script.
    • Human-in-the-loop design: regulators and risk teams insist on “human approval gates” for money-moving or safety-critical actions.

    1. Vendor landscape snapshot (enterprise-ready flavors)

    • hyperscalers: AWS Bedrock Agents, Azure AI Agent Service (preview), Google Vertex AI Reasoning Engine.
    • SaaS incumbents: Salesforce Agentforce, ServiceNow AI Agents, SAP Joule with “autonomous process automation”.
    • Startups: Adept, CrewAI Enterprise, MultiOn, AutoGPT Enterprise, Fixie, Ikigai (tabular agents), Glean Agents.

    1. What to expect in the next 12 months

    • “Agent-as-API” marketplaces inside large enterprises (analogous to internal micro-service catalogs).
    • SLAs and cost budgets attached to agent runs (FinOps for agents).
    • First regulatory guidance (EU AI Act, US NIST) specifically calling out “autonomous agent risk tiers.”
    • Consolidation of orchestration frameworks (LangGraph vs Autogen vs Semantic Kernel) around 2-3 de-facto standards.

    Bottom line: Agentic AI is past the slide-deck stage but still short of “table stakes.” If you are an enterprise practitioner, the safest bet is to pick a bounded, low-risk workflow (e.g., L1 support, dev-ops runbooks, or monthly reporting) and run a 90-day pilot with explicit human-in-the-loop checkpoints. That is where 90 % of the real learning—and the budget approvals—are happening right now.

  3. User

    how do you design them for non-technical employees, in large complex orgs which have multiple AI chatbots (e.g. copilot, custom LLMs, palantir, and others) in the energy sector

  4. Kimi Assistant

    Designing agentic AI for non-technical employees in a large, multi-vendor energy company is less about the agent itself and more about turning the surrounding “mess” (Copilot, Palantir Foundry, custom LLMs, SAP, OSIsoft PI, etc.) into a single, safe, self-service experience.
    Below is a playbook that has worked in two super-majors and one regulated utility; it is deliberately framed for the business user, not the data-science team.


    1. Start with a “single front door” • One branded chat / Teams app / intranet tile that every employee already knows.
      • Behind the door sits a lightweight orchestration layer (Microsoft Copilot Studio, ServiceNow AI Agents, or an internal LangGraph gateway) whose only job is to route, not to re-build the underlying models.
      • Users never see the word “agentic”; they see “Ask EnergyBot.”

    2. Map the capability lattice, not the tech stack Build a two-column spreadsheet with business verbs on the left and the system that can fulfill them on the right.
      Example:
      – “Find last week’s turbine vibration anomalies” → Palantir Foundry
      – “Book a maintenance slot” → SAP PM
      – “Generate a tailboard briefing” → Custom LLM on Azure
      – “Summarise my unread emails” → Microsoft 365 Copilot
      Publish this lattice as a living FAQ so employees know what the front door can do today.

    3. Hide the seams with “micro-agents” • Each back-end system gets its own micro-agent (≈ 200 lines of LangGraph or Copilot Studio topic).
      • Micro-agents expose ONLY the 3–5 safest, highest-value actions that a non-technical user might need.
      • They inherit the enterprise RBAC of the underlying system (e.g., SAP roles, Palantir permissions) so no new security model is required.

    4. Guardrails that feel like policy, not code • Pre-execution checklist: “This action will create a work order in SAP and notify the control-room supervisor. Proceed?”
      • Post-execution receipt: a short email / Teams message with a rollback link.
      • All prompts and tool calls are logged to the same SIEM feed the CISO already monitors.

    5. Progressive disclosure UI • Level 0: Natural-language question → single-turn answer (no agentic loop).
      • Level 1: “Would you like me to open the work order?” (one-step agent).
      • Level 2: “Shall I run the full outage-readiness checklist?” (multi-step agent).
      Users graduate through the levels only when they explicitly opt-in.

    6. “Human-in-the-loop” as a feature, not a bug • Every multi-step plan is rendered as a collapsible checklist inside Teams.
      • Employees can pause, edit, or hand off any step to a colleague.
      • A “panic button” instantly escalates to a human supervisor with full context.

    7. Governance & change management • Create a cross-functional “Agent Council” (IT, OT, Legal, HSE, Union) that meets monthly to approve new micro-agents.
      • Publish a one-page “Agent Nutrition Label” for each micro-agent: purpose, data sources, human-override path, cost per run.
      • Run quarterly phishing-style red-team drills to ensure employees still know when NOT to trust the bot.

    8. Training in 15-minute chunks • Record 3-minute loom videos titled “Book your PTO in 3 clicks with EnergyBot.”
      • Embed the videos inside the chat window the first three times a user invokes that skill.
      • Use Microsoft Viva Learning or similar to push micro-lessons to the employee’s existing learning path.


    Concrete 90-day pilot template (energy sector)

    Week 1–2 Pick one use-case: “Generate a tailboard safety briefing for tomorrow’s substation work.”
    Week 3–4 Build micro-agents:
    – Pull work-order data from SAP PM
    – Pull weather from NOAA API
    – Pull live switching orders from ADMS (via Palantir)
    – Draft briefing via GPT-4 on Azure OpenAI
    Week 5–6 Add guardrails: supervisor approval email, rollback link.
    Week 7–8 Soft-launch to 20 field supervisors; collect telemetry.
    Week 9–12 Iterate, then expand to “auto-file the post-job safety report.”


    Key takeaway
    Non-technical employees do not need to understand “agentic loops”; they need a single, trustworthy front door that speaks their language, respects their existing permissions, and always shows a clear path to a human when things get complex. Build the orchestration layer once, expose only the safest verbs, and iterate in 90-day sprints.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.