Observe ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- Most agentic AI deployment failures will not be AI failures – they’re architectural failures.
- The 12 guidelines of agentic AI for profitable enterprise transformation are vendor-neutral and agnostic.
- Most AI pilots give attention to functionality and velocity – and skip the exhausting work of incomes belief from the enterprise.
A latest Salesforce study discovered that greater than half of US desk employees take into account themselves AI skeptics, whereas folks in rising economies are extra trusting of AI.
The American AI skepticism goes past job losses. US desk employees are involved about worker expertise, lack of coaching, and readiness to undertake AI applied sciences. The highest three causes for an unsuccessful AI instrument or pilot amongst US employees embrace generic outputs, inadequate coaching, and low belief in outputs.
Additionally: US workers are the world’s biggest AI skeptics – and it’s not just about job loss
The dearth of belief in agentic AI pilots and transformational efforts extends additional, with many research pointing to greater failure charges of manufacturing deployments of AI brokers.
Accenture’s newest research finds that corporations should display sustained early wins from AI investments to construct momentum. The secret’s shifting from siloed AI to systemic AI. The analysis discovered that profitable agentic AI tasks require sturdy information foundations utilizing clear information to ship the best context, investments in governance and semantically constant information, which requires a contemporary AI-enhanced cloud stack, AI guardrails, and redesigned workflows.
Greater than half of agentic AI adopters cite information high quality and retrieval points as deployment limitations, based on a survey of chief information officers by Informatica.
Necessities for true agentic AI transformation
Though there have been many documented tales of agentic AI adoption within the enterprise, with mentions of excessive charges of pilot and manufacturing failures, many AI agent deployments are profitable. Over 80% of US authorities businesses already use AI brokers. A brand new survey finds that the majority authorities leaders imagine that by 2030, the general public sector will encompass people and AI brokers working collectively. Based on IDC research targeted on public-sector readiness, agentic AI is not within the experimental part for presidency; it’s a management mandate.
Additionally: Moving from AI pilots to business-wide value requires a superhighway – how to ramp up
Salesforce has realized invaluable lessons on profitable agentic AI manufacturing deployments. With over 20,000 agent AI manufacturing deployments, Salesforce has recognized many widespread errors, together with overreliance on language fashions, reliance on encoding insurance policies reasonably than complicated prompting logic, and poor context engineering. However crucial lesson is that this: With conventional software program, 90% of the work is full earlier than launch. However with AI brokers, 90% of the work comes after they’re deployed in manufacturing, together with managing and bettering them.
True agentic AI transformation in enterprise does require guidelines that companies should comply with to make sure an clever, scalable, and reliable system of outcomes.
John Taschenkgovt vice chairman and chief market technique officer at Salesforce, has been researching and creating a algorithm to benchmark the vital capabilities AI brokers have to ship profitable manufacturing deployments. Taschek’s analysis included observations throughout hundreds of agentic AI deployments, engagements with business analysts, senior executives, board members, and a neighborhood of agentic AI trailblazers.
The 12 guidelines of agentic AI
Developed by Taschek, the 12 guidelines of agentic AI for profitable enterprise transformation are vendor-neutral and agnostic. Taschek was impressed by a set of rules proposed by pc scientist Dr. Edgar F. Codd in 1985, known as Codd’s 12 Rules for true relational database administration programs.
Adherence to the 12 guidelines of agentic AI should be evidence-based with documented capabilities, technical artifacts, third-party evaluation, incomes commentary, or verified implementation outcomes. The proof should be present and inclusive of the latest set of capabilities. The proof should even be architecture-led as a substitute of easy messaging.
Additionally: AI agents are getting their own search engine
The foundations additionally help an outcome-aware mannequin the place evaluations can distinguish between technical prospects versus deployment capabilities, buyer adoption, and measurable enterprise influence. And lastly, the foundations and the general framework should even be risk-aware, in a position to determine failures, implementation and governance gaps, and customer-reported challenges. Listed below are the 12 guidelines of agentic AI:
Basis – system of knowledge/context
Rule 1. Unified information lineage: Every bit of knowledge should have a traceable historical past — the place it got here from, the way it modified, and who’s allowed to make use of it. No thriller information feeding your brokers.
Rule 2. Grounded real-time information entry. Brokers should work with dwell information, not stale snapshots. Performing on outdated data is a design flaw, not simply an inconvenience.
Rule 3. Semantic metadata: Brokers want to grasp the which means of knowledge, not simply the uncooked values. “At-risk buyer” or “certified account” should be formally outlined — not guessed by the mannequin.
Core – system of company
Rule 4. Observability / behavioral traceability: Each resolution an agent makes must be logged and explainable. You want to have the ability to look again and perceive why it did what it did.
Rule 5. Steady adversarial validation: Continuously take a look at brokers towards edge circumstances, unhealthy inputs, and adversarial eventualities — not simply at launch, however ongoing. Consider it as a everlasting red-team train.
Additionally: AI engineer vs. forward deployed engineer: Which role delivers the most business value?
Rule 6. Multi-step reasoning/objective decomposition: Brokers should have the ability to take a posh objective, break it into steps, and execute — adapting if issues change alongside the way in which, and never simply following a script.
Rule 7. Hybrid deterministic governance: AI reasoning is probabilistic, however some guidelines can’t be bent. Authorized, monetary, and security guardrails should be hard-coded — the agent must be architecturally incapable of violating them.
Operations – system of labor
Rule 8. Agnostic orchestration: Brokers from completely different distributors and fashions have to coordinate with out customized plumbing for each pairing. Keep away from lock-in on the orchestration layer.
Rule 9. Human-agent synergy/empathy mandate: Brokers ought to collaborate with people, not change them. When confidence is low or emotional context is detected, hand off gracefully — with full context, not a chilly switch.
Rule 10. Sovereign company: The enterprise stays in management — information residency, mannequin selection, identification, and coverage. Exterior brokers get scoped, auditable entry solely. Nothing is trusted by default.
Additionally: Why AI tokens will send your enterprise cloud bill sky-high again
Rule 11. Consequence-based parity: Measure brokers by enterprise outcomes (income influenced, points resolved, time saved), not by what number of duties they full. The bar is real-world influence.
Apex – system of engagement
Rule 12. Trusted company: The best-weighted rule. Brokers earn the best to behave via:
- Algorithmic equity – no bias throughout protected teams.
- Toxicity and content material security – content material screening earlier than supply.
- Consent and information permissions – honoring what clients agreed to.
- Hallucination prevention – no confabulation in high-stakes contexts.
- Explainability – anybody (regulator, buyer, advisor) can perceive why.
- Stakeholder worth – outcomes should profit clients, not simply the enterprise.
- Vendor accountability – legal responsibility is pre-assigned, not negotiated after an incident.
Making use of these guidelines earlier than and after manufacturing
Most agentic AI pilot failures will not be AI failures; they’re architectural failures — groups attempting to construct programs of engagement with out a full basis. The one commonest failure is because of AI brokers launched on prime of messy, siloed, or stale information. With out unified information (rule 1), the agent can’t hint what it is performing on. With out real-time entry (rule 2), the agent makes selections on outdated snapshots. And with out semantic metadata (rule 3), the agent doesn’t perceive what the info means. For this reason so many AI agent pilots look nice in managed environments however fail the second they face manufacturing information.
Additionally: How AI agents will transform your customer service – despite 3 hurdles
When an agent AI pilot produces the improper reply or a bizarre reply, groups uncover they don’t have any visibility into why. No one can reply what occurred (rule 4) with out observability and habits traceability – what you must debug, defend, or enhance. Pilots fail not as a result of AI was improper, however as a result of it was opaque. Pilots are validated on clear, consultant information in managed settings. They not often face adversarial inputs, edge circumstances, or unhealthy actors (rule 5). Steady adversarial validation is skipped as a result of it looks like further work. Demos normally present single-step duties. Actual enterprise work is multi-step and ambiguous. When the AI agent hits a real multi-step problem (rule 6) — dependencies, context shifting, competing alerts — it both fails silently or requires fixed human babysitting.
We frequently see no guardrails till there’s an incident. Groups will skip hybrid deterministic governance (rule 7) as a result of it slows issues down. They depend on the mannequin to “know” what to not do. Then the AI agent approves one thing it should not, or violates a coverage. Governance is added reactively after the incident — way more expensive than constructing it in from the beginning. Profitable AI agent manufacturing deployments require brokers to work with different brokers and people (agnostic orchestration, rule 8) — human-agent synergy (rule 9).
AI pilots additionally typically use vendor-hosted fashions with out pondering via information residency, entry controls, or who owns what. Sovereign company (rule 10) issues — particularly in regulated industries – floor late, triggering authorized and procurement opinions that freeze or kill manufacturing deployments. When AI brokers are in manufacturing, enterprise leaders should have the ability to measure enterprise influence earlier than and after AI deployments. With out outcome-based parity (rule 11), the case for scaling agentic AI deployment is a intestine really feel, not a quantity. Finances holders ask: “What did we really accomplish?” and there is not any reply.
Additionally: The autonomous business is coming. Here’s why that shift is good news for professionals
And lastly, AI manufacturing deployments fail as a result of belief was by no means earned. Most pilots give attention to functionality and velocity – and skip the exhausting work of equity testing, consent enforcement, hallucination prevention, and explainability. When one thing goes improper, there is not any belief structure to fall again on. One unhealthy output in a regulated or customer-facing context ends this system totally.
The 12 guidelines of agentic AI pyramid doesn’t work the wrong way up. The agentic AI pilots and manufacturing deployments that succeed deal with information high quality, governance, and human collaboration as stipulations — not afterthoughts.
