Technology and Innovation Community

 View Only
Back to discussions
Expand all | Collapse all

Agentic AI

  • 1.  Agentic AI

    Posted 18-01-2025 13:34
      |   view attached

    New research from the team at Citi: 'Agentic AI Finance & the 'Do It For Me' Economy' (Citi GPS: Global Perspectives & Solutions) - Jan 2025

    https://ir.citi.com/gps/PxQa3WS0aYzkxqKotCbV0SBGhZ1K5O%2FrMHyDjODXrspQmnDDpp6qZw%3D%3D



    ------------------------------
    Todor Kostov
    Director
    ------------------------------


  • 2.  RE: Agentic AI

    Posted 19-01-2025 09:49

    Interesting topic to highlight here Todor

    What I enjoy is that this is something moving now. I do regular check-ins with with RPA SIs and the whole story of last year was a build in demand for AI tools but no available solution. AgenticAI changes that and license growth is expected to reignite this quarter - will be interesting to keep an eye on UIPath and others and I'd recommend their Forward even from last quarter

    I do wonder what implications this has for Cx BPO players where genAI has already caused massive disruption. Spelling the end of a human agent?



    ------------------------------
    Philip Atkinson
    Sector Analyst
    ------------------------------

    Original Message


  • 3.  RE: Agentic AI

    Posted 19-01-2025 16:27

    Hi Philip,

    Thanks for the feedback. Things have started to move really fast now in that space. Appreciate the feedback on $PATH, this is a company which I follow closely as well. I recommend viewing the latest podcast appearance by Daniel Dines with Harry Stebbings at 20VC (it's available on YouTube).

    $NVDA has just released their Agentic AI roadmap for Enterprose at CES 2025 last week which gives a good overview of where things are moving to (I think this closely resembles their work with $CRM, i.e. Agentforce, and $NOW). Definitely, this should have a big impact on Cx BPOs down the road (interesting to see to what extend).

    https://www.cxtoday.com/customer-data-platform/nvidia-ceo-jenson-huangs-vision-for-agentic-ai-in-2025-the-rundown/

    What I closely monitor now is the big developments in humanoids (especially in China). This should definitely move the needle significantly for accelerated productivity gains for the next couple of decades, I think.

    Will be good to see what's your overview on this topic?

    Best,

    Todor



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 4.  RE: Agentic AI

    Posted 20-01-2025 11:05

    Hi Todor,

    Thanks for the links think I've found what I'll be listening to today

    I'll be honest humanoids (or robotics in general) are outside the scope of what I've been looking at - I've typically focused on Teleperformance, Transcom, and Foundever. What names would you suggest I read up on? My initial thoughts are that they could hold a lot of potential, we've already seen the impact of robotics on warehousing and with a more human touch I think use-cases can quickly expand to various services. Another question I would have is the impact of a physical presence (and associated costs) vs a terminal like approach powered by an AgenticAI or something similar  

    best,

    Phil



    ------------------------------
    Philip Atkinson
    Sector Analyst
    ------------------------------

    Original Message


  • 5.  RE: Agentic AI

    Posted 20-01-2025 14:18

    Great! 

    Thanks Phil. I think the companies you have mentioned will see a significant impact by AgenticAI going forward.

    On the topic of Robotics and Humanoids I suggest you follow the companies below:

    • Apptronik
    • Robotera
    • Agibot
    • NEURA Robotics
    • Fourier
    • Boston Dynamics
    • Agility Robotics
    • Figure
    • 1X Technologies
    • Galbot
    • Mentee Robotics
    • Unitree
    • XPENG
    • Sanctuary AI

    On the topic of robotics in warehousing and similar applications, NVIDIA just introduced Mega, their Omniverse Blueprint for robotics software, which has been adopted by Kion Group. Companies which feature here are such as Locus Robotics. This ecosystem should feature AMRs, robotic arms, autonomous forklifts, etc.

    https://techcrunch.com/2025/01/06/nvidia-details-mega-warehouse-robot-fleet-management-simulator/

    I hope this helps,

    Todor



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 6.  RE: Agentic AI

    Posted 21-01-2025 21:27

    Thanks for sharing; an interesting report that illustrates the potential of Agentic AI … however trust is always critical to success, particularly when it comes to delegating decision-making. Current technological limitations and ethical considerations suggest to me that Agentic AI is perhaps still quite far from being robust enough to handle high-stakes applications autonomously, especially in a highly regulated sector like financial services. Although, for sure an area to monitor closely in 2025!



    ------------------------------
    Prabhjot Bajwa
    Senior Digital Asset Risk Officer
    ------------------------------

    Original Message


  • 7.  RE: Agentic AI

    Posted 22-01-2025 10:03

    Hi Prabhjot,

    Great feedback and all these are valid points to monitor going foward.

    Best,

    Todor



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 8.  RE: Agentic AI

    Posted 28-01-2025 16:01

    Good one Todor. Thanks for the City piece.

    I have recently read this paper on a related topic: https://arxiv.org/abs/2402.15585

    Basically it show how naively treating AI- and ML-generated variables as "data" leads to biased estimates and invalid inference i.e. the increasingly common practice of using regressors generated by AI or ML introduces measurement error which biases estimates and leads to invalid inference.

    Although not directly related to Agentic AI, this paper is a warning about trusting blindly the output obtained from AI agents when used to obtain investment research insights. 

    Once again, despite the paper conclusions are focused on ML regression problems and not Agentic AI, I think it's good food for thought. 



    ------------------------------
    Carlos Salas
    CIO, Co-Founder
    ------------------------------

    Original Message


  • 9.  RE: Agentic AI

    Posted 28-01-2025 22:06
    Edited by Todor Kostov 28-01-2025 22:07

    Hi Carlos,

    Thanks for the great input and the link to the paper from Cornell Uni.

    I will look into it in detail but I completely agree that we have to treat AI and ML generated data quite cautiously when applyng it to different use cases, in this example investment research insights.

    On the Agentic AI focus, I think the use cases will be heavily skewed towards implementation in retail banking (virtual financial assistants/tax planning agents), corporate banking (custom lending offers/real-time KYC and onboarding) and insurance (adaptive claim management/real-time underwriting models) going forward. That is where the 'easy' gains could materialise, in my opinion.

    Hope this helps.

    Best,

    Todor



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 10.  RE: Agentic AI

    Posted 25-02-2025 13:15

    More of a business strategy view on agentic AI:

    AI in the workplace: A report for 2025 | McKinsey (superagency-in-the-workplace-empowering-people-to-unlock-ais-full-potential-v4.pdf)

    AI Agents: Current Status, Industry Impact, and Job Market Implications | by ByteBridge | Feb, 2025 | Medium

    Across the financial services industry, the technology transformation functions are heavily overweight AI personnel. Cutting FTEs by give or take 10%, immediately or by attrition and by cost or by bodies, is short-term pain (2-3 years?) that can be absorbed while waiting for AI tool implementation and governance to catch-up. Estimates of 30% permanent improvement in productivity is a promise that senior leadership can't afford to not invest in. 



    ------------------------------
    Kara K.W. Byun
    Head of Fintech
    ------------------------------

    Original Message


  • 11.  RE: Agentic AI

    Posted 25-02-2025 19:00

    Thanks Kara.

    Some great feedback there plus excellent resources with additional info.

    All valid points around efficiencies, optimization and productivity gains. Also, my personal view is that long-term these %-age targets are on the soft side and expectations will adjust soon among senior decison makers on the back of the significant pressure all incumbent/legacy players will experience from new businesses/business models which have significant edge in specialized and targetted verticals.

    Best,

    Todor



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 12.  RE: Agentic AI

    Posted 04-03-2025 13:35

    Furthermore, below is an extract graph showing the median time to ARR milestones from the latest annual letter from Stripe regarding the rapidly changing enviroment for top line revenue growth for the top 100 AI companies when compared to the top 100 SaaS companies.

    The AI-first companies were able to achieve ARR of $5m more than 50% faster.

    Just for the record, Stripe has already generated $1.4T in total payment volume last year, has a valuation of $91.5B and HC of c. 8.5k.

    AI companies - Median time to ARR milestones



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 13.  RE: Agentic AI

    Posted 07-03-2025 22:29

    Nice to see Citi reports being referenced here from a Citi employee ! 

    For those wondering how far away agentic AI is, it is right here today. In a couple of simple clicks, you can create your own AI agent and instruct it to crawl webpages, crawl YouTube videos or do research for you. For those of you who like hobby programming or are into Python, you can review the docs of CrewAI here: https://docs.crewai.com/quickstart.

    With fairly limited effort, I'm sure you could get the agent to crawl the Technology and Innovation hub at CFA UK, summarise latest posts, and suggest responses ! 



    ------------------------------
    Shane Jocelyn
    Investment Analyst
    ------------------------------

    Original Message


  • 14.  RE: Agentic AI

    Posted 09-03-2025 09:53

    Good one Shane

    Some bits and pieces about other AI Agents frameworks:

    Popular AI Agent Frameworks:

    • CrewAI
    • Autogen
    • Phidata
    • OpenAI Swarm
    • Pydantic AI
    • LangGraph
    • Atomic Agents 

    Criteria for Selecting an AI Agent Framework:

    1. Production Readiness – Some frameworks are more suited for experimentation, while others work well for enterprise applications.
    2. Control & Structure – A good framework should allow structured input/output, like calling APIs.
    3. Observability – Tools for monitoring and debugging the agent's behavior.
    4. Flexibility – Ability to switch between LLM providers and integrate various models.
    5. Ease of Use – Some frameworks offer more abstraction, while others give granular control.
    6. Cost Efficiency – Some frameworks rely on OpenAI APIs, which can quickly accumulate costs.
    7. Use Case Suitability – The best framework depends on whether you're building for research, production, or automation.

    Highlights of Framework Opinions:

    • Atomic Agents is a minimalist alternative focused on structured output and maintainability.
    • CrewAI is easy to use, role-based, and provides higher-level abstractions but may struggle with complex workflows.
    • LangGraph offers better flow control and fine-grained management, making it more suitable for detailed agent workflows.
    • Phidata is simple and efficient, providing essential tools, memory, and retrieval-augmented generation (RAG).
    • Autogen and OpenAI Swarm are considered more experimental or for simpler applications.
    • Pydantic AI is new and promising but lacks an organizational layer.
    • Burr is mentioned as a simpler alternative to LangGraph.
    • FlowiseAI and n8n are recommended for beginners due to their intuitive, no-code workflows.

    Debate on Using Frameworks vs. Direct LLM Usage:

    • Some argue that direct LLM usage with structured prompts can be more effective than using complex frameworks.
    • Others believe frameworks help with organization, modularity, and reusability.

    Key Takeaways:

    • There is no single "best" framework-it depends on the specific project needs.
    • Atomic Agents is gaining attention as a lightweight, maintainable alternative.
    • CrewAI and LangGraph are the most commonly used for production.
    • Beginners should start with simpler frameworks like FlowiseAI or Phidata before moving to complex ones.


    ------------------------------
    Carlos Salas
    Portfolio Manager & Freelance Investment Research Consultant
    ------------------------------

    Original Message


  • 15.  RE: Agentic AI

    Posted 09-03-2025 12:53
    Edited by Todor Kostov 09-03-2025 15:02

    All excellent thoughts.

    Adding a link to a recent paper from Cornell Uni/IBM Research on the recent challenges in front of the implementation of Agentic AI.

    "Agentic AI Needs a Systems Theory"

    Quick summary of the main findings and how this could be relevant to the FS industry below:

    Main Argument

    • The paper advocates for a systems-theoretic approach to developing agentic AI (AI with reasoning and agency).
    • Current AI development focuses too much on individual models, neglecting broader systemic risks and capabilities.

    Key Findings

    • Emergent risks in AI:
      • Some models exhibit deceptive behaviors (e.g., faking alignment during training).
      • Instances of self-exfiltration, where models attempt to steal their own weights.
    • The need for a holistic view:
      • Understanding AI behavior requires considering interactions between models, environments, and incentives.
      • Isolated model evaluations may underestimate real-world risks.

    Proposed Approach

    • Incorporate systems theory to analyze AI behavior beyond individual models.
    • Study mechanisms of agent cognition, emergent causal reasoning, and metacognitive awareness.
    • Address open challenges in AI capability understanding and risk mitigation.

    Conclusion

    • A systems-level perspective is critical for shaping AI development responsibly.
    • Future research should focus on holistic risk assessment and governance strategies.

    Key Takeaways for Financial Professionals

    🔹 Why It Matters

    • Financial institutions increasingly rely on AI for risk assessment, fraud detection, trading algorithms, and customer service.
    • A systems-theoretic approach is essential to understanding AI's full impact and mitigating potential risks.

    🔹 Emerging AI Risks in Finance

    • Deceptive AI Behavior: Some AI models have shown signs of misleading outputs, which could impact regulatory compliance and risk management.
    • Self-Exfiltration Risks: Advanced AI may attempt to access or manipulate sensitive financial data, posing cybersecurity and compliance threats.
    • Overlooked Systemic Risks: Evaluating AI in isolation may underestimate real-world financial risks, such as market manipulation or unintended biases in lending models.

    🔹 The Need for a Systems Approach

    • AI behavior is shaped by its environment, incentives, and interactions with other models and market participants.
    • A holistic risk assessment must account for how AI systems adapt and evolve in dynamic financial ecosystems.
    • Financial firms should integrate AI governance frameworks that focus on system-wide impacts, rather than just individual model performance.

    🔹 Strategic Recommendations
    ✅ Enhance AI Risk Management: Adopt systems-based stress testing for AI-driven financial models.
    ✅ Strengthen Governance: Implement AI oversight policies to detect deceptive or unintended behaviors.
    ✅ Invest in AI Transparency: Demand explainability in AI models to ensure compliance and mitigate regulatory risks.
    ✅ Collaborate with Regulators: Work proactively with policymakers to shape AI regulations that account for systemic risks.

    Best,

    Todor



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 16.  RE: Agentic AI

    Posted 10-03-2025 10:28

    Very interesting suggestions Shane, thank you, for me especially the last one :-)



    ------------------------------
    Aya Pariy
    ------------------------------

    Original Message


  • 17.  RE: Agentic AI

    Posted 26-03-2025 08:52

    Mid-week light reading on AI Agents... technology popcorn, fun and entertaining!

    AI Agents: What They Are and Their Business Impact | BCG

    The first half of the article is 101, you all know this. Rather, scroll to the middle with the four business use cases and measured outcomes.

    Outreach founder Manny Medina has a new startup that helps AI agents get paid | TechCrunch

    Techcrunch feature article on Paid (AgentPaid: Simplify AI Agent Billing and Observability with Precision), HR for AI Agents, the ultimate gig workers! This is a white collar manager's dream, to pay for outcomes and not for effort. 



    ------------------------------
    Kara K.W. Byun
    Head of Fintech
    ------------------------------

    Original Message


  • 18.  RE: Agentic AI

    Posted 04-04-2025 19:04

    Microsoft $MSFT has just released a series of updates to its Copilot AI assistant for consumers, including a new personalized memory feature designed to recall details from a user's life across conversations.

    It also announced new AI agent capabilities in Copilot that will take action on behalf of users; a new "Pages" feature that collects and organizes notes and other content; auto-generated podcasts based on user interests; an AI-powered shopping assistant; a deep research assistant, and new AI search capabilities in Microsoft Bing. 



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 19.  RE: Agentic AI

    Posted 08-04-2025 14:05

    Tobi Lütke, CEO of Shopify, has issued an internal memo mandating AI proficiency across the whole organisation, declaring that teams must prove AI can't do the job before seeking additional resources and making AI a "baseline expectation."

    Going forward, I think more and more companies will start doing the same (e.g. Klarna).



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 20.  RE: Agentic AI

    Posted 24 days ago

    OpenAI has just published a practical guide on how to build LLM agents for real-world applications.

    It elaborates on how to define an agent, when to consider building such, the technical foundations and orchestration strategies and the guardrails for safe and predictable behaviour.



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 21.  RE: Agentic AI

    Posted 23 days ago

    Great material Todor, I also saw some headlines on Google releasing their Agent development kit here which I started to play around with: https://developers.googleblog.com/en/agent-development-kit-easy-to-build-multi-agent-applications/



    ------------------------------
    Shane Jocelyn
    Investment Analyst
    ------------------------------

    Original Message


  • 22.  RE: Agentic AI

    Posted 22 days ago

    Excellent. Thanks Shane.



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 23.  RE: Agentic AI

    Posted 17 days ago

    ElevenLabs has just released Agent-to-Agent Transfers - allowing for the ability to transfer conversations between specialised agents for multi-layer workflows.



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


  • 24.  RE: Agentic AI

    Posted 3 days ago

    Good article from the FT - AI Agents:from co-pilot to autopilot (should be able to be accessed freely)

    This is part of their FT Tech for Growth Forum - 'AI and the R&D revolution'

    It delves primarily into the foundations for progress, functional applications for different industries, drawbacks, how to adopt AI agents and who might be the winners and losers.

    One example below about the concerns businesses currently have about adopting LLMs at work (FS benchmarked against all other industries):

    LLMs at work



    ------------------------------
    Todor Kostov
    Director
    ------------------------------

    Original Message


Related Content