Pretxt Emerges From Stealth With $5 Million Seed Round
Pretxt has come out of stealth with $5 million in seed funding for software that supervises AI agents against company policy while they work.
Pretxt, a start-up that supervises autonomous AI agents while they work, has come out of stealth with $5 million in seed funding from Deciens Capital, Howdy Partners, Service Provider Capital and NSV Wolf Capital.
The company is launching what it calls AI Supervision, software that checks each action an agent takes against an organization’s own policies as a task runs, and steers the agent back if it strays. For banks and large companies, the pitch is AI agent supervision compliance teams can evidence: a control that operates at the moment an agent acts, rather than a review of logs afterward.
What the seed round pays for
Pretxt said it will use the money to establish runtime supervision as a dedicated layer of enterprise AI. It has not published a lead investor, a valuation or a customer list, and its website names no design partners, only example use cases in customer support, marketing, financial services and healthcare.
Brian Soby is co-founder and chief executive, and Aaron Costello is co-founder and chief technology officer. Its AI agent supervision is aimed at high-stakes workflows, particularly in regulated industries where an agent’s error can carry compliance consequences.
The problem Pretxt targets is behavior that is wrong without being malicious. An agent can skip a mandatory step, invent an exception, make a commitment it has no authority to make or ignore an approval requirement. A support agent that offers an unauthorized refund, or a financial services agent that bypasses an identity check, causes damage with no attacker involved.
“You don’t need a prompt injection attack for an AI agent to do something it shouldn’t.”
Brian Soby, Co-founder and CEO, Pretxt
How runtime AI agent supervision works
Pretxt connects to an existing agent, which keeps its job, and teams then write policies and limits in plain language. The platform turns those into rules and a readable plan, which teams test against their own example situations before approving it.
In production, the supervision layer assesses behavior in the context of each task. Compliant interactions continue after millisecond-scale checks, while actions that breach an organization’s boundaries are caught and resolved before they reach a customer or a system of record. Pretxt also learns from live agent behavior, flags gaps in the guidance and suggests refinements, but humans review and approve every change before it goes live.
The design choice that separates Pretxt from many guardrail products is that it does not use a large language model as the judge of each action. The company said this removes latency, cost and inconsistent decisions. Its website puts the added delay at about 50 milliseconds per decision, against more than three seconds for an LLM judge, and says identical requests produce identical outcomes.
In testing on Zapier’s AutomationBench, which mirrors real-world support workflows, Pretxt cut policy-violating agent actions from almost one in five to zero. That is a vendor test on one benchmark of support tasks, not a result from a bank’s production environment, so compliance teams will want to see the AI agent supervision reproduced on their own workflows.
Where AI agent supervision compliance fits the rulebook
Pretxt does not claim to map to a specific regulation, and its materials do not mention the EU AI Act or US model risk guidance. The comparison is nonetheless one buyers will make, because each framework asks firms to show control over automated decisions while they are in use.
Article 14 of the EU AI Act, Regulation (EU) 2024/1689, requires high-risk AI systems to be designed so that people can oversee them effectively while they are in use. A runtime layer that stops or redirects an agent and routes policy changes to a human approver is the kind of mechanism a firm could point to in that conversation, although whether a given agent is high-risk under the Act depends on its use, not on the tooling around it.
In the US, the Federal Reserve and the Office of the Comptroller of the Currency set out expectations for model risk management in SR 11-7, including ongoing monitoring and effective challenge of models. Banks that already run a model risk management program will have to decide whether agents are inventoried and validated as models, and where a runtime control sits in that framework.
Senior manager accountability regimes raise a related question of who answers when an agent acts outside its authority. A control that records what an agent was told to do, what it did and when it was redirected gives the accountable executive a clearer trail, though it does not move responsibility from the executive to the software.
How the product differs from existing tools
Pretxt’s argument is that security, governance, guardrails and observability each cover part of the risk without confirming that an agent is completing its assigned task to business or regulatory standards. Some guardrails halt a prohibited action and leave the process stuck awaiting human intervention, while monitoring tools show only what has already happened.
That places Pretxt next to, rather than inside, the established governance, risk and compliance stack. Those platforms hold policies, attestations and evidence for the board, whereas Pretxt acts on the transaction in flight. Whether buyers treat that as a new category or as a feature of existing platforms will depend on customer evidence the company has yet to publish.
Other start-ups are approaching agent risk from different angles. Opsin raised $21 million for enterprise AI agent protection, and Cygrid raised €2.66 million to control which data agents can reach. Pretxt’s focus is narrower: whether the agent followed the process, not what it can access.
What compliance buyers will ask next
Pretxt sits in the same wave of tools as Norm AI and Arva AI, which put agents to work inside compliance functions. As those agents take on screening, onboarding and case review, firms will need a way to show that each one stayed inside approved procedure, which is the gap Pretxt says it fills.
Co-founder Aaron Costello said the platform is meant to spare teams from learning every agent’s internals.
“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.”
Aaron Costello, Co-founder and CTO, Pretxt
That plain-language approach lowers the barrier for compliance staff to write AI agent supervision rules, but it also moves the weight onto policy quality. The rules a supervision layer enforces are only as complete as the standards written into it, which is why Pretxt’s gap-spotting step and its human approval of changes matter to auditors.
The next evidence to watch is whether Pretxt names its first regulated customers and publishes results from production deployments, which would show how its checks perform against the policies of a bank or insurer rather than a benchmark.

The go to weekly newsletter for compliance, risk, and governance professionals.
Trusted by 10,000+ industry leaders for authoritative RegTech intelligence, delivered weekly.
Join them today.
