The enterprise software giants made their bets months apart, but the timing created the same headache. SAP said its AI Agent Hub would hit general availability around mid-2026. Oracle's NetSuite added MCP support this past March. Suddenly, IT leaders had vendor-blessed AI agents capable of autonomously navigating the guts of ERP systems—purchase orders, financial close processes, supply chain reconfigurations—with no standardized way to prove those agents wouldn't, say, hallucinate a fraudulent invoice or leak customer data during a routine query.
A three-person Y Combinator startup thinks it has the answer, or at least an answer. TrustAI, which emerged from the accelerator's Summer 2026 batch, launched a compliance platform around July 21 designed specifically for AI agents running on sensitive business systems. The pitch: pre-deployment assessments that map agent behavior to ISO 42001, NIST AI RMF, and the EU AI Act's AIUC-1 framework, alongside a cost-optimization router that claims to slash model inference costs by about a third.
The timing isn't accidental. The EU AI Act's 2026 enforcement deadlines for high-risk systems are approaching, and a Cloud Security Alliance research note from April flagged that most production agents lack the audit-quality evidence trails regulators will soon demand. TrustAI is wagering that enterprise buyers need that evidence before their agents go live in S/4HANA or Oracle Cloud ERP. Not after an incident makes the rounds on LinkedIn.
What They're Actually Selling
TrustAI ships two modules, though "Security" feels like an understatement for what the first one does. It runs 51 preregistered tests across six risk categories: hallucinations, data privacy, security, robustness at scale, permission compliance, and something the company calls "correctness and control." Each assessment spits out an 18-page report with pass/fail verdicts, 95% confidence intervals on red-team attack success rates, and a cross-walk to major compliance frameworks.
The sample report TrustAI published—hypothetical agent, real methodology—shows a system that passed 10 of 13 controls but failed the "security gate" threshold. The attack success rate's upper confidence bound exceeded 5.0 percent. The report flags which NIST AI RMF measures and OWASP LLM Top 10 controls the agent violates, then ties failed tests to specific framework requirements. For a compliance officer prepping for an EU AI Act audit, that mapping is the entire point.
The second module, TrustRouter, is an OpenAI-compatible gateway that routes each inference request to the cheapest model meeting a quality threshold. TrustAI's internal benchmark claims 32.4% cost reduction while retaining 100.2% of baseline quality, with routing decisions clocking in at 0.01 milliseconds at both the 50th and 95th percentiles.
The company hasn't published third-party validation of those numbers, and no external verification is available for these internally generated claims.
The ERP Gambit
TrustAI's integrations page lists SAP S/4HANA, SAP ECC, Oracle Cloud ERP, and NetSuite as built-in connectors, alongside ITSM tools like Jira and ServiceNow. The platform can assess agents running on a vendor's native MCP hub or through TrustAI's own gateway. Either way, the company says, it applies the same risk model and can link go/no-go decisions directly into a ServiceNow ticket or Jira issue.
That ERP-first positioning tracks with what's happening in the market, or at least what the big vendors are promising. SAP's Joule assistants—announced at Sapphire 2026 and rolling out across supply chain and finance modules later in the year—represent the first wave of vendor-supported autonomous agents in core transactional systems. Oracle and NetSuite are pursuing similar strategies, though with the typical variance in messaging and timelines that makes covering enterprise software a special kind of endurance sport.
For IT leaders, the question isn't whether to deploy these agents; the vendors are baking them into the roadmap regardless. The question is how to govern them without derailing the projects their CFOs have already budgeted for.
Who's Building This
Hannah Chung, TrustAI's CEO, studied computer science and economics at MIT before stints at Virtu Financial and the World Bank. Medha Venkatapathy, the CTO, also went to MIT, focusing on physics and computer science. She competed on the U.S. Physics Olympiad team and worked on LLM inference optimization research with Jacob Andreas at MIT. It's a technical founding team tackling a technical problem, though the startup's messaging emphasizes compliance outcomes over model architecture—probably the right call when you're selling to CISOs.
The team is small. Three people as of early August, per Y Combinator's company directory. TrustAI hasn't disclosed customers, revenue, or a funding round beyond the standard YC batch participation. The company is incorporated as TrustAI AI Limited in London, with operations listed in San Francisco. Standard startup geography for a compliance play that needs to straddle EU and U.S. regulatory regimes.
A Market That Got Crowded Fast
TrustAI is entering a space that filled up with remarkable speed in 2026. Companies like Trust3 AI, Rein Security, Aguardic, and TrustLogix (which confusingly also markets a "TrustAI" feature) all launched agent governance platforms between April and July. Most emphasize runtime monitoring and continuous compliance.
TrustAI differentiates by focusing on pre-deployment assessments—catching issues before an agent touches production data—though its FAQ mentions options for continuous re-evaluation when agents change. Whether that distinction holds up as products evolve is anyone's guess.
The academic community is working the same problem from another angle. Papers presented at June's FAccT conference and posted on arXiv describe formal verification methods for agent policy compliance and runtime safety checks. Whether the market consolidates around commercial platforms, open-source tooling, or vendor-native governance features in SAP and Oracle's own products remains an open question. History suggests the answer will be "all of the above," uncomfortably.
Why the Urgency
The MIT AI Agent Index 2025, presented at FAccT in late June, found that many production agents lack published safety or compliance frameworks—a finding that surprised approximately no one who's worked in enterprise IT. A Forrester Wave report on governance platforms, released around the same time, noted the industry is shifting from periodic audits to continuous controls monitoring. Meanwhile, the EU AI Act's August 2026 deadlines are creating urgency for anyone deploying high-risk AI systems in Europe.
TrustAI's bet is that enterprises will pay for deterministic, framework-mapped evidence before deploying agents, rather than scrambling to generate audit trails after something goes sideways. The company's test catalog includes crosswalks to ISO 42001, NIST, MITRE ATLAS, and the EU's AIUC-1 standard. For a CISO explaining to auditors why the finance agent has write access to accounts payable, that documentation could be worth the spend. Or at least worth a pilot project.
Whether the platform gains traction depends on factors TrustAI hasn't disclosed: pricing, early customer adoption, and whether its Router's cost savings hold up under independent testing. The company also faces a classic startup challenge—it needs enterprises to adopt agent governance practices before they'll buy agent governance platforms. Right now, many IT organizations are still figuring out if they need Joule assistants at all, let alone how to assess them.
But the market TrustAI is addressing is real, even if the contours are still forming. SAP and Oracle are shipping autonomous agents into ERP systems. Regulators are demanding evidence of AI safety controls. And someone has to build the tools that generate that evidence.
For a three-person YC startup, it's a narrow opening. But perhaps—if they move fast enough, and if enterprises decide they care about compliance before they care about cost—a durable one.
