2 hours ago
5
Follow ZDNET: Add america arsenic a preferred source connected Google.
ZDNET's cardinal takeaways
- Most agentic AI deployment failures are not AI failures - they're architectural failures.
- The 12 rules of agentic AI for palmy endeavor translation are vendor-neutral and agnostic.
- Most AI pilots absorption connected capableness and velocity - and skip the hard enactment of earning spot from the business.
A caller Salesforce study found that much than fractional of US table workers see themselves AI skeptics, portion radical successful emerging economies are much trusting of AI.
The American AI skepticism goes beyond occupation losses. US table workers are acrophobic astir worker experience, deficiency of training, and readiness to follow AI technologies. The apical 3 reasons for an unsuccessful AI instrumentality oregon aviator among US workers see generic outputs, insufficient training, and debased spot successful outputs.
Also: US workers are the world's biggest AI skeptics - and it's not conscionable astir occupation loss
The deficiency of spot successful agentic AI pilots and transformational efforts extends further, with galore studies pointing to higher nonaccomplishment rates of accumulation deployments of AI agents.
Accenture's latest research finds that companies indispensable show sustained aboriginal wins from AI investments to physique momentum. The cardinal is shifting from siloed AI to systemic AI. The probe recovered that palmy agentic AI projects necessitate beardown information foundations utilizing cleanable information to present the close context, investments successful governance and semantically accordant data, which requires a modern AI-enhanced unreality stack, AI guardrails, and redesigned workflows.
More than fractional of agentic AI adopters mention information prime and retrieval issues arsenic deployment barriers, according to a survey of main information officers by Informatica.
Requirements for existent agentic AI transformation
Although determination person been galore documented stories of agentic AI adoption successful the enterprise, with mentions of precocious rates of aviator and accumulation failures, galore AI cause deployments are successful. Over 80% of US authorities agencies already usage AI agents. A caller survey finds that astir authorities leaders judge that by 2030, the nationalist assemblage volition dwell of humans and AI agents moving together. According to IDC research focused connected public-sector readiness, agentic AI is nary longer successful the experimental signifier for government; it is simply a enactment mandate.
Also: Moving from AI pilots to business-wide worth requires a superhighway - however to ramp up
Salesforce has learned invaluable lessons connected palmy agentic AI accumulation deployments. With implicit 20,000 cause AI accumulation deployments, Salesforce has identified galore communal mistakes, including overreliance connected connection models, reliance connected encoding policies alternatively than analyzable prompting logic, and mediocre discourse engineering. But the astir important acquisition is this: With accepted software, 90% of the enactment is implicit earlier launch. But with AI agents, 90% of the enactment comes aft they are deployed successful production, including managing and improving them.
True agentic AI translation successful concern does necessitate rules that businesses indispensable travel to guarantee an intelligent, scalable, and trustworthy strategy of outcomes.
John Taschek, enforcement vice president and main marketplace strategy serviceman astatine Salesforce, has been researching and processing a acceptable of rules to benchmark the captious capabilities AI agents request to present palmy accumulation deployments. Taschek's probe included observations crossed thousands of agentic AI deployments, engagements with manufacture analysts, elder executives, committee members, and a assemblage of agentic AI trailblazers.
The 12 rules of agentic AI
Developed by Taschek, the 12 rules of agentic AI for palmy endeavor translation are vendor-neutral and agnostic. Taschek was inspired by a acceptable of principles projected by machine idiosyncratic Dr. Edgar F. Codd successful 1985, called Codd's 12 Rules for existent relational database absorption systems.
Adherence to the 12 rules of agentic AI indispensable beryllium evidence-based with documented capabilities, method artifacts, third-party analysis, earning commentary, oregon verified implementation outcomes. The grounds indispensable beryllium existent and inclusive of the astir caller acceptable of capabilities. The grounds indispensable besides beryllium architecture-led alternatively of elemental messaging.
Also: AI agents are getting their ain hunt engine
The rules besides enactment an outcome-aware exemplary wherever evaluations tin separate betwixt method possibilities versus deployment capabilities, lawsuit adoption, and measurable concern impact. And lastly, the rules and the wide model indispensable besides beryllium risk-aware, capable to place failures, implementation and governance gaps, and customer-reported challenges. Here are the 12 rules of agentic AI:
Foundation - strategy of data/context
Rule 1. Unified information lineage: Every portion of information indispensable person a traceable past -- wherever it came from, however it changed, and who's allowed to usage it. No enigma information feeding your agents.
Rule 2. Grounded real-time information access. Agents indispensable enactment with unrecorded data, not stale snapshots. Acting connected outdated accusation is simply a plan flaw, not conscionable an inconvenience.
Rule 3. Semantic metadata: Agents request to recognize the meaning of data, not conscionable the earthy values. "At-risk customer" oregon "qualified account" indispensable beryllium formally defined -- not guessed by the model.
Core - strategy of agency
Rule 4. Observability / behavioral traceability: Every determination an cause makes should beryllium logged and explainable. You request to beryllium capable to look backmost and recognize wherefore it did what it did.
Rule 5. Continuous adversarial validation: Constantly trial agents against borderline cases, atrocious inputs, and adversarial scenarios -- not conscionable astatine launch, but ongoing. Think of it arsenic a imperishable red-team exercise.
Also: AI technologist vs. guardant deployed engineer: Which relation delivers the astir concern value?
Rule 6. Multi-step reasoning/goal decomposition: Agents indispensable beryllium capable to instrumentality a analyzable goal, interruption it into steps, and execute -- adapting if things alteration on the way, and not conscionable pursuing a script.
Rule 7. Hybrid deterministic governance: AI reasoning is probabilistic, but immoderate rules cannot beryllium bent. Legal, financial, and information guardrails indispensable beryllium hard-coded -- the cause should beryllium architecturally incapable of violating them.
Operations - strategy of work
Rule 8. Agnostic orchestration: Agents from antithetic vendors and models request to coordinate without customized plumbing for each pairing. Avoid lock-in astatine the orchestration layer.
Rule 9. Human-agent synergy/empathy mandate: Agents should collaborate with humans, not regenerate them. When assurance is debased oregon affectional discourse is detected, manus disconnected gracefully -- with afloat context, not a acold transfer.
Rule 10. Sovereign agency: The endeavor stays successful power -- information residency, exemplary choice, identity, and policy. External agents get scoped, auditable entree only. Nothing is trusted by default.
Also: Why AI tokens volition nonstop your endeavor unreality measure sky-high again
Rule 11. Outcome-based parity: Measure agents by concern outcomes (revenue influenced, issues resolved, clip saved), not by however galore tasks they complete. The barroom is real-world impact.
Apex - strategy of engagement
Rule 12. Trusted agency: The highest-weighted rule. Agents gain the close to enactment through:
- Algorithmic fairness - nary bias crossed protected groups.
- Toxicity and contented information - contented screening earlier delivery.
- Consent and information permissions - honoring what customers agreed to.
- Hallucination prevention - nary confabulation successful high-stakes contexts.
- Explainability - anyone (regulator, customer, advisor) tin recognize why.
- Stakeholder worth - outcomes indispensable payment customers, not conscionable the enterprise.
- Vendor accountability - liability is pre-assigned, not negotiated aft an incident.
Applying these rules earlier and aft production
Most agentic AI aviator failures are not AI failures; they are architectural failures -- teams trying to physique systems of engagement without a implicit foundation. The azygous astir communal nonaccomplishment is owed to AI agents launched connected apical of messy, siloed, oregon stale data. Without unified information (rule 1), the cause cannot hint what it's acting on. Without real-time entree (rule 2), the cause makes decisions connected outdated snapshots. And without semantic metadata (rule 3), the cause does not recognize what the information means. This is wherefore truthful galore AI cause pilots look large successful controlled environments but neglect the infinitesimal they look accumulation data.
Also: How AI agents volition alteration your lawsuit work - contempt 3 hurdles
When an cause AI aviator produces the incorrect reply oregon a weird answer, teams observe they person nary visibility into why. Nobody tin reply what happened (rule 4) without observability and behaviour traceability - what you request to debug, defend, oregon improve. Pilots neglect not due to the fact that AI was wrong, but due to the fact that it was opaque. Pilots are validated connected clean, typical information successful controlled settings. They seldom look adversarial inputs, borderline cases, oregon atrocious actors (rule 5). Continuous adversarial validation is skipped due to the fact that it feels similar other work. Demos usually amusement single-step tasks. Real endeavor enactment is multi-step and ambiguous. When the AI cause hits a genuine multi-step situation (rule 6) -- dependencies, discourse shifting, competing signals -- it either fails silently oregon requires changeless quality babysitting.
We often spot nary guardrails until there's an incident. Teams volition skip hybrid deterministic governance (rule 7) due to the fact that it slows things down. They trust connected the exemplary to "know" what not to do. Then the AI cause approves thing it shouldn't, oregon violates a policy. Governance is added reactively aft the incidental -- acold much costly than gathering it successful from the start. Successful AI cause accumulation deployments necessitate agents to enactment with different agents and humans (agnostic orchestration, regularisation 8) -- human-agent synergy (rule 9).
AI pilots besides often usage vendor-hosted models without reasoning done information residency, entree controls, oregon who owns what. Sovereign bureau (rule 10) concerns -- particularly successful regulated industries - aboveground late, triggering ineligible and procurement reviews that frost oregon termination accumulation deployments. When AI agents are successful production, concern leaders indispensable beryllium capable to measurement concern interaction earlier and aft AI deployments. Without outcome-based parity (rule 11), the lawsuit for scaling agentic AI deployment is simply a gut feel, not a number. Budget holders ask: "What did we really accomplish?" and there's nary answer.
And lastly, AI accumulation deployments neglect due to the fact that spot was ne'er earned. Most pilots absorption connected capableness and velocity - and skip the hard enactment of fairness testing, consent enforcement, hallucination prevention, and explainability. When thing goes wrong, there's nary spot architecture to autumn backmost on. One atrocious output successful a regulated oregon customer-facing discourse ends the programme entirely.
The 12 rules of agentic AI pyramid does not enactment upside down. The agentic AI pilots and accumulation deployments that win dainty information quality, governance, and quality collaboration arsenic prerequisites -- not afterthoughts.
English (US) ·