Pretxt has emerged from stealth with $5 million in seed funding to introduce an enterprise technology platform that supervises AI agents while they perform business tasks. Deciens Capital, Howdy Partners, Service Provider Capital, and NSV Wolf Capital participated in the financing.
The company describes its product category as AI Supervision. Its runtime technology evaluates an agent’s behavior within the context of an assigned task, identifies deviations from approved business processes, and guides the agent back toward an acceptable course of action.
Pretxt also analyzes agent behavior to recommend improvements to an organization’s supervision plan. Human reviewers must approve those proposed changes before they take effect.
The company is based in Aspen, Colorado, and Dublin, Ireland. Its platform is available through an early-access program.
Supervising AI Agents During Execution
Pretxt is addressing operational failures that can occur even when an AI agent has not been compromised by an attacker.
AI agents increasingly communicate with customers, access enterprise applications, process transactions, and perform multistep business workflows. However, an agent could skip a required procedure, create an unauthorized exception, make an unapproved commitment, or proceed without obtaining a necessary authorization.
These actions can create financial, compliance, operational, and reputational consequences. Examples include a customer-service agent issuing an unauthorized refund or a financial-services agent bypassing identity-verification requirements.
Security, governance, observability, and conventional guardrails address portions of this risk, but Pretxt argues that they do not necessarily determine whether an agent is completing its assigned work according to an organization’s procedures.
Monitoring tools may identify a problem after an action has occurred, while a guardrail that blocks a prohibited action can leave the overall workflow incomplete. Pretxt instead evaluates the process while it is underway and attempts to steer an agent back toward the approved procedure.
Using Business Standards Instead Of Another AI Judge
Enterprise teams define their expectations, procedures, and operating boundaries in plain language. Pretxt then uses those standards to supervise an agent without requiring users to understand the agent’s underlying permission model, internal architecture, or tools.
The company says its platform does not use a large language model to judge another AI agent’s behavior. Pretxt believes that approach avoids unnecessary cost, latency, and inconsistent decisions.
Its technology is designed to provide fast, near-deterministic supervision at machine speed. Actions that satisfy an organization’s standards can proceed, while proposed actions that cross established boundaries are identified and addressed before reaching a customer or system of record.
Pretxt claims its approach provides supervision 60 times faster. In testing conducted with Zapier’s AutomationBench, which models real-world customer-support workflows, the company said its technology reduced policy-violating agent actions from nearly one in five to zero.
These results were reported by Pretxt and may not represent performance in every production environment.
Improving Supervision Through Operational Experience
Pretxt’s platform reviews real-world agent behavior to identify gaps in an organization’s instructions and operating boundaries.
The system can then propose refinements to the supervision plan. Enterprise users can test these proposed changes against their own examples and decide whether to approve them.
The company emphasized that this structure does not allow an agent to rewrite its own rules. Instead, it provides a human-controlled process for improving agent supervision as organizations encounter additional workflows and behavioral patterns.
Pretxt is initially targeting consequential enterprise applications, particularly in regulated industries where an AI agent’s mistake could result in financial losses, compliance violations, interrupted operations, or damage to customer relationships.
The platform could also allow companies to give agents access to tools needed to perform more substantial work, including in environments where applications do not offer sufficiently granular access controls.
Founded By Enterprise Security Veterans
Pretxt was co-founded by CEO Brian Soby and CTO Aaron Costello.
Soby has more than 25 years of experience across enterprise technology, AI, software-as-a-service security, and platform development. He previously co-founded AppOmni and served as its chief technology officer.
His earlier positions included partner at FreeFly Security, director of product security at Salesforce, lead information security engineer at MITRE, and network security engineer at Raytheon.
Costello previously served as chief of security research at AppOmni. His work included identifying more than 20 configuration risks and multiple zero-day vulnerabilities affecting Salesforce Industry Clouds.
He also discovered a vulnerability in ServiceNow AI Agents that was described as the most severe AI security flaw identified at the time. Costello’s research additionally identified more than 1,000 ServiceNow instances exposing data.
Pretxt’s distributed team includes professionals with experience at Salesforce, MITRE, SAP, Deloitte, VMware, HackerOne, and AppOmni.
KEY QUOTES:
“You don’t need a prompt injection attack for an AI agent to do something it shouldn’t. Security remains essential, but the next question is whether enterprises can trust agents to carry out real work according to their standards. You can’t supervise a workforce of AI agents with hope, a rulebook and reviews after the fact. AI needs supervision while it acts. That is the category Pretxt is here to establish.”
Brian Soby, Co-Founder and Chief Executive Officer of Pretxt
“Supervising an AI agent shouldn’t mean becoming an expert in everything it touches, or in how it’s built. Every platform has its own permission model. Every agent has its own tools and its own way of working. Nobody should have to master all of that before they can trust an agent with real work. Pretxt takes that complexity off the table. The agent already knows its job. Teams tell us, in plain language, how it should behave and where the boundaries are, and we supervise and steer the agent to stay within them. No untangling permissions, no digging into the agent’s internals. That’s a big win for trust, and for outcomes.”
Aaron Costello, Co-Founder and Chief Technology Officer of Pretxt
“Enterprises have invested heavily in tools that ask whether an agent is secure, compliant or observable. Far fewer can answer a more basic but critical question: Is the agent correctly doing the job it was given? That’s important because organizations that deploy AI own their outcomes, no matter who’s asking: regulators, insurers or customers. Most tools can only report what went wrong after the fact, when it’s too late. Pretxt intervenes while the outcome can still change to get agents back on track, creating a new category for AI supervision the market has been missing at the moment enterprises need it most.”
Chris Steffen, Vice President of Research at Enterprise Management Associates
“Our clients want to deploy agents in workflows where the business outcome matters. Pretxt gives us a way to connect written operating standards to what agents actually do in production.”
Andrew Sweet, Vice President of Emerging Technology and Strategic Accounts at EverSec
“The ROI on AI agents is limited by how much real work a company is willing to hand them. In our view, the entire AI revolution hinges on precisely this point. Pretxt expands that limit. When every action is held to an organization’s own standards as it happens, teams can move agents from simple tasks to the work that drives revenue and carries real risk. That’s a large and growing market, and Pretxt is defining it.”
Nate Pacer, Partner at NSV Wolf Capital
“AI agents are moving into workflows where mistakes have real consequences. We invested in Pretxt because we believe supervision at runtime will become a foundational layer of enterprise AI infrastructure. The opportunity is not just to identify a bad outcome after it happens, but to help organizations change the outcome while the work is still underway.”
Vishal Rana, Partner at Deciens Capital

