Governing AI Agents Starts With Deciding What Is True
A customer calls her bank's voice agent and asks to have a late fee waived because a hospital stay caused the missed payment. The agent grants the waiver, adjusts the due date, and closes the call. Three weeks later, an internal audit flags the interaction. One policy document caps hardship waivers at one per year, a compliance memo requires documentation before any waiver tied to a medical event, and branch teams have honored verbal explanations for years. The agent followed one of those rules. Nobody can say which, or whether it was right.
The instinct is to treat this as a model problem: tighten the prompt, add an output filter, or retrain on better examples. None of those fixes touch the actual gap. The organization never settled what its own rule is, so the agent had nothing definitive to follow, and the auditor had nothing definitive to check.
That is the structural reason so many agent governance efforts stall. Guardrails need rules, rules need agreement, and most enterprises have never been forced to reach that agreement in a form a machine can apply. Until AI agents started acting on their behalf, that ambiguity lived comfortably inside human judgment.
Why Improved Models Do Not Resolve the AI Agent Governance Gap
Large language models are probabilistic. Prompt instructions shape what a model is likely to do; they do not guarantee that it will do it, and they leave no reliable record of which instruction drove a given decision. That is acceptable for a drafted email, not when an agent changes an account, approves an exception, or commits the organization to an outcome a regulator may examine.
In the 2026 Gartner Head of Enterprise Architecture Role in Emerging Tech and AI Survey, roughly 69% of enterprise architecture (EA) leaders said agentic AI was already in use or in pilot at their organizations. Those same leaders named agents behaving in ways teams cannot anticipate, agents straying beyond the limits of their autonomy, and poor visibility into agent activity among the main obstacles keeping that work out of production.
The harder part sits on the other side of the equation. Symbolic, rule-based AI has always been explainable, but it depends on knowledge representations that someone has to build: what the organization treats as fact, which policies override which, and where judgment is allowed. Gartner observes that this shifts the central data challenge from volume to quality, and that it pushes business stakeholders into value judgments about which facts and policies are authoritative. In the first example, legal, compliance, and operations must decide which waiver policy is real. No vendor can make that decision for them, and it takes longer than most AI roadmaps assume.
What Governable AI Agents Are Built On
The answer pairs both approaches. In a neurosymbolic system, the neural layer handles what language models do well: understanding a caller's phrasing, interpreting documents, and recognizing patterns in messy input. The symbolic layer holds the organization's agreed rules as explicit, deterministic logic. When the agent acts, the decision passes through rules that can be inspected, and the record shows which rule applied and why. Both layers rest on a ground layer that is easy to skip: an ontology and taxonomy specific to the industry and the enterprise, defining what hardship, medical event, and verified customer mean, so every rule applies to the same concepts.
Much of what an organization knows lives outside written policy, in call recordings, recorded training, and video walkthroughs, often where informal practices like the branch teams' verbal exceptions surface. Multimodal knowledge ingestion brings those sources into the same knowledge base, so the rules teams debate reflect how work is really done. During the conversation, a caller's tone, pace, and hesitation carry information a transcript loses. An agent that interprets sentiment and emotion can slow down for a distressed customer, explain a decision more fully, or offer a person sooner. Tone shapes how the agent responds and when it brings in a human; the rules still govern the decision.
This also lets an organization set human oversight precisely. A small waiver with documentation on file may proceed without review, while a waiver tied to a medical event may always require a person to approve it. The calibration lives in the rules themselves, so it is applied consistently and can be explained after the fact.
Two practical conclusions follow, both consistent with Gartner's guidance. Codifying rules and institutional knowledge should start before any platform purchase, because building consensus is the slow part. Domain depth also matters: Gartner advises heads of EA to favor vendors whose industry expertise matches their own.
What Gartner's Neurosymbolic AI Research Signals for Buyers
Gartner's August 2026 research note, Neurosymbolic AI, a Shortcut to Agentic Systems Governance, examines how neurosymbolic approaches can support the governance of agentic systems, and expects neurosymbolic AI to become a foundation for built-in AI governance at 70% of enterprises by 2032. The note includes a table of vendors focused on neurosymbolic AI, and Openstream.ai, with its Eva platform, appears in that table, with example applications listed for conversational assistants and for the insurance, healthcare, banking, and finance sectors. The note also references Openstream.ai as an example in its discussion of multimodality, describing the company's focus on models that work with voice and video, not only text, in conversational settings. For enterprise buyers, the takeaway is the market direction: explainable, rule-constrained agent architecture is moving out of research settings and into buying criteria.
How Openstream.ai Applies Neurosymbolic Reasoning to Regulated Work
EvaTM, Openstream.ai's enterprise AI platform, applies neurosymbolic reasoning to both conversational AI experiences and multiagent systems. In the bank scenario, that means a voice agent that understands both the caller's words and the strain in her voice, then decides through the institution's codified waiver policy, recording the rule it applied, the evidence it considered, and the level of human review that rule requires. If the policy says a medical hardship waiver needs a person's approval, the request goes to a person, and the record explains why.
Openstream.ai has focused on this work in regulated industries such as banking, insurance, and healthcare, where an unexplainable decision costs the most. Eva is designed to carry an organization's own ontology, policies, and knowledge, so the rules teams agree on become the rules agents follow.
Where to Start Before Choosing a Neurosymbolic AI Platform
The decision most enterprise leaders face is not whether to adopt neurosymbolic AI next quarter. It is whether to begin the slower work now: identifying the decisions agents will make, finding where policies conflict, and getting the right owners to agree on which version governs. Organizations that wait may find the platform was never the bottleneck. Once the rules are settled, the most useful question for any vendor is a simple one: for every action an agent takes, can the system point to the rule behind it?
Gartner, Neurosymbolic AI, a Shortcut to Agentic Systems Governance, Benjamin Arnberg, 10 August 2026, ID G00861801.
GARTNER is a registered trademark and service mark of Gartner, Inc. and/or its affiliates in the U.S. and internationally and is used herein with permission. All rights reserved. Gartner does not endorse any vendor, product or service depicted in its research publications and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organization and should not be construed as statements of fact. Gartner disclaims all warranties, expressed or implied, with respect to this research, including any warranties of merchantability or fitness for a particular purpose.