Summary
Every AI system you deploy carries failure points you probably haven't mapped. The risk surface of AI has expanded at each stage of its evolution — from narrow models with a clean lifecycle, to generative systems with layered decision-makers, to agents connected to databases, CRMs, the internet, and other agents, where every node is a place things can go wrong. The consequences are already public: Deloitte made headlines in October 2025 for an Australian government report containing AI-hallucinated information, then again a month later for the same failure in a report for the Canadian government; a major law firm submitted a legal brief with fabricated citations; and HR software company Eightfold faces litigation over allegedly biased hiring algorithms. This article maps the hidden points of failure across the AI lifecycle and explains why agentic AI multiplies them.
What Did the AI Risk Landscape Look Like Before?
Narrow AI — the model that tells you whether a bone is broken in an X-ray, scores a resume, or recognizes faces in photos — had a relatively clear lifecycle: concept, design, build, test, validate, deploy, monitor. The points where an ethical, reputational, or legal error could enter were bounded and, for the most part, knowable in advance.
That clarity is gone. The risk landscape is now diabolically complex, and organizations built for the narrow-AI era aren't prepared for it.
Where Are the Failure Points in Generative AI?
Generative AI splits the risk chain across multiple decision-makers:
Foundation model companies (OpenAI, Anthropic, Google) make consequential choices about training data, fine-tuning, and system behavior baked into the base model before your organization ever touches it.
Internal data science teams add layers of customization — fine-tuning, RAG architectures, adapters — each introducing new ways to introduce error.
Everyday users prompt the system in countless ways across countless roles and contexts, generating a near-infinite space of inputs no policy can fully anticipate.
Each of these is a point where hallucination, bias, or privacy violation can enter — and each has produced real-world consequences:
Deloitte (October 2025): a report commissioned by the Australian government contained AI-hallucinated information, with payment disputes making headlines.
Deloitte (November 2025): a different member firm repeated the failure in a report for the Canadian government.
A major law firm: submitted a legal brief to a judge containing AI-hallucinated information.
Eightfold: the HR software company faces litigation over potentially biased hiring algorithms.
These cases are documented and tracked by academic institutions studying AI accountability. Harvard's Berkman Klein Center examines how AI-based systems can "exacerbate existing inequalities and biases, or even create new ones" and researches governance mechanisms to prevent it (Berkman Klein Center, Ethics and Governance of AI), while Oxford Internet Institute's Programme on the Governance of Emerging Technologies investigates the legal and ethical risks of machine learning systems as they're deployed at scale (Oxford Internet Institute, GET Programme).
Why Does Agentic AI Multiply the Failure Points?
Agentic AI is, in essence, an LLM connected to lots of digital tools: the internet, enterprise databases, CRMs, other pieces of enterprise software, other AIs (narrow or generative), and even other agents.
Every one of those connections is a node of interaction — and every node is a potential point of failure where biases, hallucinations, or privacy violations can enter. Agents are so complicated that, without proper governance, they're practically guaranteed to go off the rails: they're intricate pieces of machinery, and machinery that isn't properly developed and overseen fails.
Notably, full-blown agentic AI failures have not yet appeared in the news at the scale organizations claim to be deploying agents. The most plausible explanation: people don't know how to create AI agents in a way that is governable, so they aren't actually doing it at the scale they claim. The failure points exist; the deployments, for now, are smaller than the announcements.
How Can Enterprises Map Their Own Failure Points?
A practical starting sequence for any organization:
Inventory every node. For each AI system, list the model provider, every customization layer, every tool connection, and every user population.
Identify who makes decisions at each node. Foundation model company, internal data scientists, end users, and — with agents — the agent itself.
Articulate the nightmares at each node. What's the worst ethical, reputational, or legal outcome this node could produce?
Score and prioritize. Rate each nightmare by likelihood and severity to focus resources where risk concentrates.
Design context-specific controls. Generic procedure lists won't fit: of 20 generic controls applied to any given agent, 10 won't be relevant and 10 needed controls won't be on the list.
For enterprises benchmarking their approach, Stanford HAI's AI Index provides global data on AI capability, investment, and policy — essential context for understanding how quickly the risk surface is expanding (Stanford HAI, 2026 AI Index Report).
Key Takeaways
Narrow AI had bounded, mappable failure points; generative and agentic AI do not.
Generative AI risk spans foundation model companies, internal customizers, and everyday prompters.
In agentic AI, every tool connection — databases, CRMs, other agents — is a point of failure.
Real-world failures are already public: hallucinated government reports, fabricated legal citations, biased hiring algorithms.
Generic control lists fail because agent risks are context-specific; govern at the node level.
FAQ
What are the main points of failure in generative AI?
Three layers: decisions made by foundation model companies, decisions made by internal teams customizing the model (fine-tuning, RAG, adapters), and the countless ways users prompt the system across roles and contexts.
Why is agentic AI riskier than generative AI?
Agents are LLMs connected to external tools — databases, CRMs, the internet, other AIs, and other agents. Each connection is a node of interaction where failure can occur, and agents act with autonomy, so failures can propagate without human review.
What happens when an AI system fails?
Depending on the failure point, an organization may face hallucinated content in public deliverables, discriminatory outcomes in hiring or lending, privacy violations, lawsuits, regulatory fines, and lasting reputational damage.
Have major companies already experienced these failures?
Yes. Deloitte delivered AI-hallucinated content in reports for the Australian (October 2025) and Canadian (November 2025) governments; a major law firm submitted a brief with fabricated citations; and HR software firm Eightfold faces litigation over allegedly biased algorithms.
How should an organization start mapping its AI risks?
Inventory every node of each AI system, identify the decision-makers at each node, articulate the worst-case outcome at each point, score by likelihood and severity, and design context-specific controls rather than applying generic checklists.
To learn more about mitigating corporate risks, listen to this episode of The AI Table Podcast and review our Agentic AI Implementation Guide.
