Off-the-shelf AI agents fail because workflows are unique

At a glance

What this is: Opnova argues that generic AI agents break down in real businesses because workflows, approval logic, and data context vary too much for plug-and-play deployment.

Why it matters: IAM, IGA, PAM, and AI governance teams need to treat agent deployment as a workflow and identity design exercise, not a simple model selection problem.

Context

Off-the-shelf AI agents often fail for a simple reason: business workflows are not interchangeable. What looks like a standard process usually hides local approval chains, data exceptions, and system-specific rules that generic automation does not understand.

In identity terms, the problem is less about model capability and more about governed access to the right systems, at the right time, with the right context. For agentic AI programmes, modelling, grounding, and fine-tuning are closer to identity integration work than to software installation.

Key questions

Q: How should teams govern AI assistants, workflows, and autonomous agents differently?

A: Teams should govern them by runtime behaviour, not by model family. Assistants need strong prompt and response controls, triggered workflows need untrusted-input screening and narrow tool scope, and autonomous agents need separate identities, scoped delegation, and traceability across each decision. A single AI policy rarely fits all three.

Q: Why do AI agents create more governance risk than ordinary integrations?

A: AI agents can connect quickly, run continuously, and accumulate broad permissions across multiple services. That combination makes ownership blur and scope drift more likely, so the real risk is not the tool itself but the uncontrolled access path it creates across enterprise systems.

Q: What are the signs that an AI agent is too broadly scoped?

A: Frequent exception handling, repeated human correction, misuse of adjacent data sources, and inconsistent outcomes across similar cases are strong indicators. When the agent needs constant override to complete common tasks, the problem is usually not model quality alone, but a governance boundary that is too loose for the workflow.

Q: Should organisations customise AI agents before production use?

A: Yes. Production use should follow modelling, grounding, and controlled fine-tuning, because those steps convert a generic system into one that fits the actual business process. Without them, the agent may appear productive while silently breaking process rules that matter for security, compliance, and operational reliability.

Technical breakdown

Why workflow modelling matters for AI agent identity

Workflow modelling is the step where an organisation maps the real process, not the idealised one. For AI agents, that means capturing approval thresholds, exception paths, human review points, and downstream system dependencies before the agent is allowed to act. Without that map, the agent may follow the happy path but fail on the cases that actually matter to the business. In practice, this is where agent identity governance begins: the agent needs a defined operating boundary, not just a prompt and a tool list.

Practical implication: model the business process before granting any agent production access.

Grounding determines whether the agent can act on real context

Grounding is the process of connecting the agent to the organisation’s own data, terminology, policies, and operational signals. A generic model may know how invoices usually work, but it does not know your vendor naming, internal approval hierarchy, or product-specific exception handling. That gap matters because an agent without grounding will make confident but context-blind decisions. For identity teams, grounding is the difference between a tool that can call systems and a governed actor that understands which records and rules it is allowed to use.

Practical implication: restrict agent access to the data sources and policy context it has actually been grounded in.

Fine-tuning creates bespoke behaviour, but not autonomy

Fine-tuning adapts the model to an organisation’s patterns, language, and edge cases. It improves fit, but it does not by itself create safe governance or independent judgement. The article’s core point is that customisation is required because no two businesses run the same process, not because the agent suddenly becomes a trusted employee. For practitioners, this means fine-tuning should be treated as one layer in a broader control stack that includes authorisation, observability, and review of what the agent is permitted to change.

Practical implication: pair fine-tuning with explicit authorisation and monitoring controls before scaling use.

NHI Mgmt Group analysis

Off-the-shelf AI agents expose an assumption collapse in workflow governance: the belief that process intent is stable enough to provision once and reuse everywhere. That assumption was designed for bounded, repeatable operations, but it fails when each business function contains local exceptions, implicit approvals, and hidden data dependencies. The implication is that agent governance must move from generic deployment thinking to process-specific identity design.

Customisation is not a nice-to-have for agentic AI, it is the control plane: the article shows that modelling, grounding, and fine-tuning are the means by which an agent becomes safe enough to operate in a live environment. This is a governance pattern, not a feature checklist. Practitioner teams should treat every new workflow as a separate authorisation and integration decision, not as a reusable template.

Anthropomorphism creates governance debt: naming agents like people encourages organisations to grant them human-like trust without human-like accountability. That shortens review discipline and obscures the fact that the system is still executing through machine identity, not judgment. The practical consequence is that identity programmes need language discipline as much as access discipline.

The strongest control is not broader access, but narrower context: the article makes clear that agents fail when they are exposed to too much business ambiguity. The right boundary is not “let the agent learn everything,” but “let the agent act only inside the workflow it was modelled for.” That aligns with least privilege for non-human identities and should shape how AI agents are onboarded.

AI agent governance now sits at the intersection of NHI, IAM, and process ownership: this is not purely an AI problem and not purely an access problem. It is a joint accountability problem where domain owners define the workflow, IAM defines the access boundary, and security validates the operating assumptions. Teams that split those responsibilities will keep producing brittle pilots instead of durable automation.

From our research library:

• 62% of organisations are at least experimenting with AI agents, but only 23% have scaled one into a single business function, according to McKinsey's 2025 State of AI survey.
• Read next: Agentic AI Identity Guide

What this signals

Workflow specificity is now an identity governance issue: agents do not fail because they are AI, they fail because the access model assumes a generic process that does not exist. Programmes that treat every workflow as interchangeable will keep creating brittle permissions and hard-to-audit exceptions. The better pattern is to define the agent’s context boundary as tightly as any other privileged non-human identity.

53% of security leaders expect AI to run major portions of their infrastructure autonomously within the next three years, according to the 2026 Infrastructure Identity Survey. That expectation raises the bar for workflow modelling, because autonomy without bespoke governance simply scales mistakes faster than humans can review them.

For practitioners

• Model the real workflow before deployment Document approval thresholds, exception paths, system dependencies, and handoffs for each target process before any agent is connected to production systems.
• Ground the agent in approved data sources Limit the agent to the exact policies, records, and knowledge sources required for the workflow, and remove access to adjacent context it does not need.
• Separate human-like branding from governance decisions Do not let names, personas, or “digital employee” language substitute for defined ownership, escalation rules, and review points for the agent’s actions.
• Treat fine-tuning as a controlled change Put model updates, prompt changes, and workflow expansions through the same review discipline you would use for any production access change or process redesign.

Key takeaways

• Off-the-shelf AI agents break down when organisations assume one workflow template can fit many different approval, data, and exception patterns.
• The article’s central lesson is that modelling, grounding, and fine-tuning are governance steps, not just technical implementation details.
• Practitioners should define the agent’s operating boundary before production use, or risk scaling automation that cannot handle real business variation.

Key terms

• Workflow Modelling: The process of mapping how a business task actually works, including approvals, exceptions, systems, and handoffs. For AI agents, workflow modelling defines the boundary inside which the system can act safely, because generic prompts cannot infer local operating rules or hidden dependencies.
• Grounding: The process of tying an AI system to the organisation's actual data, terminology, and operating context so it can act with relevant awareness. For identity governance, grounding is not just about accuracy. It is about limiting the agent to the business reality it was approved to operate in.
• Fine-tuning: A model adaptation step that changes how an AI system behaves when exposed to organisation-specific data or tasks. In governance terms, fine-tuning is only useful when paired with access boundaries and testable controls, otherwise it can increase confidence without improving accountability.
• Anthropomorphism: Anthropomorphism is the tendency to attribute human traits, intentions, or emotions to a non-human system. With AI, that can cause users to over-trust a tool, read social meaning into generated language, and expect human-like memory or judgment where none exists.

Deepen your knowledge

NHI governance, agentic AI identity, and machine identity security are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are building or maturing an IAM programme, it is worth exploring.