The Gist
What happens when an AI agent pulls from outdated or conflicting company systems? It doesn’t reconcile the conflict — it delivers whatever it can access, with full confidence, even if the data is wrong. What’s actually slowing down enterprise AI adoption? Unanswered governance questions — data access limits, authorization, logging and accountability — not a lack of AI tools. What six governance elements does Liferay say must exist before deploying AI agents? Access control, a single source of truth, audit trails, human review checkpoints, a clear escalation path and model-agnostic architecture.
Imagine a customer contacts a company through a chatbot embedded on its website to check on a claim. The agent gives a confident answer, but then the customer logs into their portal and sees something different. They call support and hear a third version. That’s what happens when information comes from different systems, updated at different times, with no single source of truth. Adding AI into the mix only automates and accelerates the confusion.
The pressure to adopt AI and achieve results with it is intense. We’re all bombarded with promises about AI’s potential to finally solve longstanding problems. It can, but not if we don’t lay the right groundwork.
In the rush to show adoption numbers, many organizations are skipping the conversation that actually determines whether AI improves the customer experience or just makes existing problems worse.
Enterprise AI Agents Are Already Handling High-Stakes Customer Decisions
AI agents in enterprise environments today aren’t just chatbots that answer basic questions. They’re handling customer service triage, answering account-specific questions about benefits and claims, guiding users through self-service portals, automating compliance review steps and managing multilingual publishing workflows.
My company uses AI to allow field marketers to generate content and help our support team triage requests. We’ve encouraged our teams to innovate with AI for both internal and customer-facing use cases. I’ve seen first-hand how consequential the outputs of those AI tools are, and I know we aren’t outliers in that regard.
An agent tasked with answering customer questions about billing, benefits or product eligibility can have a real impact on revenue and customer experience. When a customer receives an incorrect answer about their coverage or their account balance, it comes across like a broken promise. In highly regulated industries, compliance and reputation are on the line. As organizations deploy agents more widely, the stakes only get higher.
Related Article: How CX Leaders Should Split Work Between Humans and AI
Why Do AI Agents Amplify Broken Customer Experiences Instead of Fixing Them?
Because an agent tasked with billing, benefits or eligibility questions doesn’t reconcile conflicting data sources — it delivers whatever it can reach, and a wrong answer there reads to the customer as a broken promise, with real revenue and compliance consequences.
What Happens When AI Agents Inherit a Fragmented Data Environment
Before deploying AI, organizations need to understand that, unless it’s been trained to do so, an AI agent doesn’t reconcile conflicting information. It delivers whatever it can access and moves on. If your customer portal hasn’t been updated since a policy change last quarter and your support documentation is running two versions behind, the agent will reflect that.
On top of that, AI will deliver its response with authority. That leads to one of two outcomes. The customer takes the AI at its word and then is surprised when it turns out their claim actually wasn’t being processed, and had been flagged for an error. Or, the AI output is so off that the customer recognizes it immediately. In both instances, they lose trust. Regaining that trust is hard, and it completely offsets any productivity gains you gained from deploying AI in the first place.
What Happens When an AI Agent Delivers Wrong Information With Total Confidence?
Customers either believe the wrong answer and get blindsided later, or the answer is obviously off and they lose trust immediately — either way, any productivity gain from deploying the agent gets wiped out.
Why Governance Gaps — Not AI Capability — Are Stalling Enterprise Adoption
As I talk to both peers and customers about AI adoption, I’m noticing a trend. Even though they have access to powerful AI tools, they’re realizing that they haven’t answered the governance questions that responsible deployment requires.
What data can an agent access, and what data should it never touch? When an agent acts on a user’s behalf, whose authorization covers that action? What gets logged? Who reviews AI outputs in high-stakes interactions before they reach a customer? What happens when an agent produces an incorrect or harmful response, and who is accountable for that outcome?
We went through our own version of this when we started expanding AI adoption across teams. We built internal guardrails such as training protocols, clear accountability structures and review checkpoints before we deployed AI tools broadly. The principle we kept returning to was that if you use AI to produce something and then communicate it, that output is yours. The model can’t be accountable, but people can.
What Governance Questions Must Enterprises Answer Before Deploying AI Agents?
Who’s authorized when an agent acts on a user’s behalf, what data it can never touch, what gets logged, who reviews high-stakes outputs and who’s accountable when the agent gets it wrong. The author’s operating principle is that if you use AI to produce something and communicate it, that output is yours.
Six Infrastructure Requirements for Trustworthy Enterprise AI Agents
The following table highlights the most important lessons, actions and strategic considerations emerging from these six infrastructure requirements for trustworthy enterprise AI agents.
RequirementWhat It PreventsWho Should Own ItAction StepAccess controlAgents surfacing data a customer isn’t authorized to seeIT/SecurityRestrict agent access to the same permissions framework as authenticated usersReliable source of truthAgents scaling stale or inconsistent data across channelsData/OpsGovern and continuously update customer-facing information before deploying agentsAudit trailsUntraceable errors and no visibility into agent failuresEngineering/ComplianceLog every AI interaction for accountability and improvementHuman review checkpointsHigh-risk responses reaching customers uncheckedRisk/ComplianceDefine review triggers by risk level and context before launchEscalation pathCustomers stuck with no way to reach a humanCustomer SupportBuild a clear, fast handoff from agent to humanModel-agnostic architectureGovernance breaking when the underlying model changesPlatform/EngineeringBuild controls at the platform layer, independent of any one modelView All What Six Infrastructure Elements Make Enterprise AI Agents Trustworthy?
Access control tied to existing user permissions, a governed and current source of truth, full audit trails, defined human review checkpoints by risk level, a real escalation path to a human and model-agnostic architecture that doesn’t tie governance to one vendor’s model.
Why Governance Is the Prerequisite for AI Agents to Scale
I want to push back on a framing I hear often in conversations about enterprise AI. There’s a perception that investing in AI governance slows things down. What we’ve done at my company, and what I’m seeing in the market, is the opposite. Those who took governance seriously from the outset are able to implement AI more quickly and with fewer bumps in the road.
Looking ahead to the next couple of years, if you want to have the most capable, trusted AI agents, then you have to treat governance as a prerequisite today. For the sake of your investment, and most importantly, customer trust, it’s worth it.
Learn how you can join our contributor community.