What is not being said about Agentic AI

Can Agentic AI learn without continuous human input and oversight?

Yes, agentic AI can learn without continuous human input and oversight through reinforcement, self-supervised, and unsupervised learning. These methods allow the AI to improve by interacting with its environment, exploring data patterns, or optimizing for predefined goals.

• Reinforcement Learning (RL): The AI learns by trial and error, receiving rewards or penalties based on actions in an environment (e.g., AlphaGo learned by playing millions of Go games against itself).
• Self-Supervised Learning: The AI generates its own labels from data, like predicting the next word in a sentence, enabling it to learn from vast, unlabeled datasets (e.g., large language models like GPT).
• Unsupervised Learning: The AI identifies patterns or clusters in data without explicit guidance (e.g., anomaly detection in cybersecurity).

However, challenges remain:

• Goal Misalignment: Without proper initial design, the AI might optimize for unintended outcomes.
• Safety and Ethics: Autonomous learning can lead to unpredictable behavior, requiring periodic human checks.
• Data Quality: Learning depends on the quality and diversity of available data.

While Agentic AI can operate autonomously to a degree, initial human-defined objectives and occasional oversight are critical to ensure alignment with intended purposes and to mitigate risks. Fully independent learning without any human input is theoretically possible but risky and not yet practical for complex, real-world applications.

Are there specific ProcureTech solution provider quotes presenting or highlighting the downside of an Agentic AI-based solution?

Vendor Quotes or Insights on Agentic AI Downsides

Vendor	Quote or Insight on Agentic AI Downsides	Source Context
Zip	User Feedback: “The biggest difficulty is motivating new users to adopt the platform regularly.”	User review on Capterra, highlighting adoption challenges for AI-driven workflows, which implies resistance to autonomous systems like Agentic AI.
ORO Labs	No direct quote; materials imply data privacy risks: “ORO Labs ensures secure handling of procurement data across orchestrated workflows.”	Vendor website emphasis on security suggests underlying concerns about autonomous AI handling sensitive data, a common Agentic AI downside.
Trust Your Supplier	No direct quote; implied trustworthiness issues: “Our blockchain-AI integration ensures reliable supplier validation.”	Marketing focus on reliability indirectly acknowledges risks of AI errors or hallucinations in autonomous supplier assessments.
SAP Ariba	Jeff Collier, SAP: “AI agents can make mistakes, just like humans, and are not foolproof, requiring human-in-the-loop oversight to validate decisions.”	Procurement Magazine article on SAP Ariba’s AI, explicitly noting error-prone autonomy and the need for human supervision.
Coupa	Coupa Blog: “AI costs can vary unpredictably based on task complexity, such as token usage for LLMs, requiring careful budgeting.”	Coupa’s discussion on AI spend management highlights cost variability as a downside of Agentic AI systems.
Scoutbee	No direct quote; implied data quality dependence: “Scoutbee’s AI finds and pre-qualifies global suppliers in a fraction of the usual time.”	Marketing focus on AI speed implies reliance on accurate data, with unstated risks of errors if data is incomplete or biased.
Levelpath	Stan Garber, Co-founder: “Siloed people, data, and legacy systems waste valuable resources and create unnecessary risks for legal and IT departments.”	Procurement Magazine article suggests integration challenges with AI platforms like Hyperbridge, a downside for autonomous workflows.
GEP	GEP Website: “Small errors in AI-driven processes can compound over complex tasks, necessitating robust oversight to ensure reliability.”	GEP’s marketing on GEP SMART acknowledges error accumulation as a challenge in autonomous procurement platforms.
SirionLabs	No direct quote; implied explainability issues: “SirionLabs builds trust through AI-driven contract management with actionable insights.”	Emphasis on trust in marketing suggests underlying concerns about opaque AI decision-making, a common Agentic AI downside.
Tealbook	No direct quote; implied data dependency: “Tealbook’s machine learning enhances supplier discovery with cleansed, enriched data from the Internet.”	Marketing focus on data quality implies risks of gaps or biases affecting autonomous supplier intelligence.
Procurify	User Review: “Some warn that it can be complex to learn, which is challenging for onboarding everyone in the organization.”	ProcureDesk review notes a learning curve for Procurify’s AI platform, indicating user adoption challenges for autonomous systems.

Analysis and Notes

• Scarce Direct Quotes: Most vendors avoid explicit negative statements about Agentic AI in sales and marketing materials to maintain a positive narrative. SAP Ariba and Coupa are exceptions, providing clear acknowledgments of downsides like errors and cost variability, respectively. GEP also directly addresses error accumulation, a practical concern for Agentic AI.
• Inferred Downsides: For vendors like ORO Labs, Trust Your Supplier, Scoutbee, SirionLabs, and Tealbook, downsides are inferred from marketing language emphasizing solutions to risks (e.g., security, trust, data quality). This suggests awareness of Agentic AI challenges without explicit admission.
• User Feedback: Reviews on platforms like Capterra and TrustRadius (e.g., for Zip, Procurify) highlight adoption and complexity issues, which are relevant to Agentic AI’s autonomous nature but not directly tied to the term.
• Critical Perspective: Vendors’ reluctance to highlight downsides reflects a marketing strategy to prioritize benefits. However, real-world challenges like hallucinations (SAP Ariba), cost unpredictability (Coupa), and data biases (inferred for Scoutbee, Tealbook) are significant. The hype around Agentic AI may obscure these risks, necessitating cautious evaluation by buyers.
• Source Usage: Relevant sources include Procurement Magazine, Coupa’s blog, and user reviews. Irrelevant sources (e.g., digital marketing) were excluded.

Gaps and Limitations

• Lack of Explicit Quotes: Vendors like ORO Labs, Trust Your Supplier, Scoutbee, SirionLabs, and Tealbook provide no direct quotes on Agentic AI downsides, likely to avoid deterring customers. This limits specificity for these vendors.
• Focus on Benefits: Sales materials overwhelmingly highlight efficiency, automation, and cost savings, making it challenging to find candid discussions of downsides without relying on user feedback or industry analyses.
• Alternative Sources: User reviews and third-party reports (e.g., Spend Matters,) offer more critical insights than vendor websites, but even these rarely use “Agentic AI” explicitly.

So, Who Does Talk About Agentic AI’s Downside?

Agentic AI—systems capable of autonomous decision-making and action—offers transformative potential but also presents notable challenges. Here are insights from industry leaders highlighting some of these concerns:​Axios+3Atera+3Sirocco Group+3

1. Marc Benioff (Salesforce CEO) on Microsoft’s Agentic AI Approach

Marc Benioff has expressed skepticism regarding Microsoft’s agentic AI initiatives, particularly criticizing their rebranding of Copilot as “agents.” He questioned the effectiveness and practical implementation of these agents, suggesting that Microsoft’s efforts might be more about marketing than delivering tangible value. Benioff emphasized Salesforce’s focus on integrating AI to augment existing products rather than investing heavily in uncertain large-scale AI projects.

2. Travis Kalanick (Uber Co-founder) on the Impact of AI on Consulting

Travis Kalanick highlighted the disruptive potential of AI in the consulting industry, noting that consultants performing routine tasks are at risk of being replaced by AI systems. He emphasized the importance for professionals to adapt by developing and integrating AI tools rather than solely executing predefined tasks.

3. Business Leaders’ Skepticism on AI Agent Adoption

A survey conducted at the Wall Street Journal’s CIO Network Summit revealed that 82% of respondents were either merely experimenting with AI agents or not considering them at all. Concerns cited included insufficient accuracy, lack of robust security measures, and a general lack of trust in the current generation of AI agents. This skepticism underscores the challenges in achieving widespread adoption of agentic AI in enterprise settings.

4. Forbes on Cybersecurity Challenges in Agentic AI

Agentic AI systems face unique cybersecurity threats, including prompt injections that can manipulate agent behavior and the generation of off-topic or hallucinated responses. These vulnerabilities necessitate the development of robust security frameworks to ensure the safe deployment of autonomous AI agents. ​ForbesVox

5. Yoshua Bengio on the Risks of Autonomous AI Agents

Renowned AI researcher Yoshua Bengio warned about the potential catastrophic scenarios associated with autonomous AI agents, particularly if they become misaligned with human values. He emphasized the importance of maintaining control over AI systems to prevent unintended and possibly harmful outcomes.

MY TAKE

“The critical part of the human-led agent-based model is that the continuous learning capabilities of the algorithms occur by adapting to how internal and external human agents work in the real world in the end-to-end supply chain to effectively and efficiently manage and grade performance metrics from order placement through to order delivery. In essence, the Agent-Based Metaprise model performs a self-cleaning of the data as a part of the human and technology agents’ daily routine inputs becoming a seamless part of the order to fulfillment process.” – What is continuous, self-cleaning data?

Reference Links:

• What is the relationship between human Agents and AI Agents, and why is it important to Procurement’s Success?
• Do we really need humans now that there is Agentic AI?
• Human With AI, Not Human Versus AI
  

30