Follow ZDNET: Add us as a preferred source on Google.

ZDNET’s key takeaways

• Most agentic AI deployment failures are not AI failures – they’re architectural failures.
• The 12 rules of agentic AI for successful enterprise transformation are vendor-neutral and agnostic.
• Most AI pilots focus on capability and speed – and skip the hard work of earning trust from the business. 

A recent Salesforce study found that more than half of US desk workers consider themselves AI skeptics, while people in emerging economies are more trusting of AI. 

The American AI skepticism goes beyond job losses. US desk workers are concerned about employee experience, lack of training, and readiness to adopt AI technologies. The top three reasons for an unsuccessful AI tool or pilot among US workers include generic outputs, insufficient training, and low trust in outputs. 

Also: US workers are the world’s biggest AI skeptics – and it’s not just about job loss

The lack of trust in agentic AI pilots and transformational efforts extends further, with many studies pointing to higher failure rates of production deployments of AI agents. 

Accenture’s latest research finds that companies must demonstrate sustained early wins from AI investments to build momentum. The key is shifting from siloed AI to systemic AI. The research found that successful agentic AI projects require strong data foundations using clean data to deliver the right context, investments in governance and semantically consistent data, which requires a modern AI-enhanced cloud stack, AI guardrails, and redesigned workflows.

More than half of agentic AI adopters cite data quality and retrieval issues as deployment barriers, according to a survey of chief data officers by Informatica. 

Requirements for true agentic AI transformation 

Although there have been many documented stories of agentic AI adoption in the enterprise, with mentions of high rates of pilot and production failures, many AI agent deployments are successful. Over 80% of US government agencies already use AI agents. A new survey finds that most government leaders believe that by 2030, the public sector will consist of humans and AI agents working together. According to IDC research focused on public-sector readiness, agentic AI is no longer in the experimental phase for government; it is a leadership mandate.

Also: Moving from AI pilots to business-wide value requires a superhighway – how to ramp up

Salesforce has learned invaluable lessons on successful agentic AI production deployments. With over 20,000 agent AI production deployments, Salesforce has identified many common mistakes, including overreliance on language models, reliance on encoding policies rather than complex prompting logic, and poor context engineering. But the most important lesson is this: With traditional software, 90% of the work is complete before launch. But with AI agents, 90% of the work comes after they are deployed in production, including managing and improving them. 

True agentic AI transformation in business does require rules that businesses must follow to ensure an intelligent, scalable, and trustworthy system of outcomes. 

John Taschek, executive vice president and chief market strategy officer at Salesforce, has been researching and developing a set of rules to benchmark the critical capabilities AI agents need to deliver successful production deployments. Taschek’s research included observations across thousands of agentic AI deployments, engagements with industry analysts, senior executives, board members, and a community of agentic AI trailblazers. 

The 12 rules of agentic AI 

Developed by Taschek, the 12 rules of agentic AI for successful enterprise transformation are vendor-neutral and agnostic. Taschek was inspired by a set of principles proposed by computer scientist Dr. Edgar F. Codd in 1985, called Codd’s 12 Rules for true relational database management systems.

Adherence to the 12 rules of agentic AI must be evidence-based with documented capabilities, technical artifacts, third-party analysis, earning commentary, or verified implementation outcomes. The evidence must be current and inclusive of the most recent set of capabilities. The evidence must also be architecture-led instead of simple messaging.

Also: AI agents are getting their own search engine

The rules also support an outcome-aware model where evaluations can distinguish between technical possibilities versus deployment capabilities, customer adoption, and measurable business impact. And lastly, the rules and the overall framework must also be risk-aware, able to identify failures, implementation and governance gaps, and customer-reported challenges. Here are the 12 rules of agentic AI: 

Foundation – system of data/context

Rule 1. Unified data lineage: Every piece of data must have a traceable history — where it came from, how it changed, and who’s allowed to use it. No mystery data feeding your agents.

Rule 2. Grounded real-time data access. Agents must work with live data, not stale snapshots. Acting on outdated information is a design flaw, not just an inconvenience.

Rule 3. Semantic metadata: Agents need to understand the meaning of data, not just the raw values. “At-risk customer” or “qualified account” must be formally defined — not guessed by the model.

Core – system of agency

Rule 4. Observability / behavioral traceability: Every decision an agent makes should be logged and explainable. You need to be able to look back and understand why it did what it did.

Rule 5. Continuous adversarial validation: Constantly test agents against edge cases, bad inputs, and adversarial scenarios — not just at launch, but ongoing. Think of it as a permanent red-team exercise.

Also: AI engineer vs. forward deployed engineer: Which role delivers the most business value?

Rule 6. Multi-step reasoning/goal decomposition: Agents must be able to take a complex goal, break it into steps, and execute — adapting if things change along the way, and not just following a script.

Rule 7. Hybrid deterministic governance: AI reasoning is probabilistic, but some rules cannot be bent. Legal, financial, and safety guardrails must be hard-coded — the agent should be architecturally incapable of violating them.

Operations – system of work

Rule 8. Agnostic orchestration: Agents from different vendors and models need to coordinate without custom plumbing for every pairing. Avoid lock-in at the orchestration layer.

Rule 9. Human-agent synergy/empathy mandate: Agents should collaborate with humans, not replace them. When confidence is low or emotional context is detected, hand off gracefully — with full context, not a cold transfer.

Rule 10. Sovereign agency: The enterprise stays in control — data residency, model choice, identity, and policy. External agents get scoped, auditable access only. Nothing is trusted by default.

Also: Why AI tokens will send your enterprise cloud bill sky-high again

Rule 11. Outcome-based parity: Measure agents by business outcomes (revenue influenced, issues resolved, time saved), not by how many tasks they complete. The bar is real-world impact.

Apex – system of engagement

Rule 12. Trusted agency: The highest-weighted rule. Agents earn the right to act through:

• Algorithmic fairness – no bias across protected groups.
• Toxicity and content safety – content screening before delivery.
• Consent and data permissions – honoring what customers agreed to.
• Hallucination prevention – no confabulation in high-stakes contexts.
• Explainability – anyone (regulator, customer, advisor) can understand why.
• Stakeholder value – outcomes must benefit customers, not just the enterprise.
• Vendor accountability – liability is pre-assigned, not negotiated after an incident.

Applying these rules before and after production

Most agentic AI pilot failures are not AI failures; they are architectural failures — teams trying to build systems of engagement without a complete foundation. The single most common failure is due to AI agents launched on top of messy, siloed, or stale data. Without unified data (rule 1), the agent cannot trace what it’s acting on. Without real-time access (rule 2), the agent makes decisions on outdated snapshots. And without semantic metadata (rule 3), the agent does not understand what the data means. This is why so many AI agent pilots look great in controlled environments but fail the moment they face production data. 

Also: How AI agents will transform your customer service – despite 3 hurdles

When an agent AI pilot produces the wrong answer or a weird answer, teams discover they have no visibility into why. Nobody can answer what happened (rule 4) without observability and behavior traceability – what you need to debug, defend, or improve. Pilots fail not because AI was wrong, but because it was opaque. Pilots are validated on clean, representative data in controlled settings. They rarely face adversarial inputs, edge cases, or bad actors (rule 5). Continuous adversarial validation is skipped because it feels like extra work. Demos usually show single-step tasks. Real enterprise work is multi-step and ambiguous. When the AI agent hits a genuine multi-step challenge (rule 6) — dependencies, context shifting, competing signals — it either fails silently or requires constant human babysitting. 

We often see no guardrails until there’s an incident. Teams will skip hybrid deterministic governance (rule 7) because it slows things down. They rely on the model to “know” what not to do. Then the AI agent approves something it shouldn’t, or violates a policy. Governance is added reactively after the incident — far more costly than building it in from the start. Successful AI agent production deployments require agents to work with other agents and humans (agnostic orchestration, rule 8) — human-agent synergy (rule 9). 

AI pilots also often use vendor-hosted models without thinking through data residency, access controls, or who owns what. Sovereign agency (rule 10) concerns — especially in regulated industries – surface late, triggering legal and procurement reviews that freeze or kill production deployments. When AI agents are in production, business leaders must be able to measure business impact before and after AI deployments. Without outcome-based parity (rule 11), the case for scaling agentic AI deployment is a gut feel, not a number. Budget holders ask: “What did we actually accomplish?” and there’s no answer. 

Also: The autonomous business is coming. Here’s why that shift is good news for professionals

And lastly, AI production deployments fail because trust was never earned. Most pilots focus on capability and speed – and skip the hard work of fairness testing, consent enforcement, hallucination prevention, and explainability. When something goes wrong, there’s no trust architecture to fall back on. One bad output in a regulated or customer-facing context ends the program entirely. 

The 12 rules of agentic AI pyramid does not work upside down. The agentic AI pilots and production deployments that succeed treat data quality, governance, and human collaboration as prerequisites — not afterthoughts.