The first wave of enterprise AI was generative models bolted onto applications. The second was narrow point-solution agents, capable but stranded from the rest of the stack. What’s arriving now is something else: context-aware, interoperable agentic systems grounded in the data, rules, and security models the underlying applications already hold. The products OneStream rolled out at Splash 2026 keep agents within that context while allowing them to collaborate across agentic ecosystems. General-purpose LLMs reaching in through APIs cannot match what context-bound agents return, and the gap is wider than most buyers realize.
This is what separates an “eloquent guess” — Microsoft’s Tony Surma’s phrase from the keynote — from a defensible answer. A generic LLM can summarize a contract beautifully and still get the entity hierarchy wrong, the close period wrong, or the security boundary wrong. None of those are model problems. They are context problems, and context has to come from the systems that already own it. Software vendors are starting to address this directly, pushing back on the SaaS-pocalypse narrative. LLMs cannot replicate critical business rules and compliance regimes without harvesting context from the environments that already encode them.
The piece that’s new in 2026 is the architecture underneath. MCP and the broader interoperable agent-protocol layer have matured enough that vendors can credibly promise “any agent you choose, governed by our truth.” OneStream put that architecture into shipping product at the conference, including agents that surface the context customers have spent years building into the platform.
In these architectures, applications that hold critical context to ensure LLM’s respond with correct, deterministic answers.
Agents work against the system of record, not a copy of it. Hierarchies, ownership structures, intercompany relationships, fiscal calendars, dimensionality—all of it travels with the question.
The agent inherits the user’s existing access model. A controller’s session sees what a controller is allowed to see, regardless of which front-end agent the user happens to be holding.
The agent understands process state. A variance question during close has different stakes than the same question in a what-if exercise.
Put those three on top of an interoperable protocol layer and you have something that earns the name. It’s the orchestration tier that the last several years of AI investment have been working around and that addresses users most rational concern – hallucination.
The Splash keynote was the clearest articulation of this model I’ve seen from a vendor, and it wasn’t a roadmap promise. The Sensible AI Agentic Layer was announced as generally available, with live demos of Copilot and Claude pulling income statements and drilling into variances through OneStream’s MCP tools, with results routed through the customer’s existing security and access model. Clustering Analysis went GA in the same keynote. The four native agents—Search, Finance Analyst, Deep Analysis, and the new Forecast Agent—are shipping, not previewed. An eight-year AI history in AI (ML123 in 2017, SensibleAI Forecast GA in 2022, Studio in 2025, Agents and Agentic Layer in 2026) gave the announcements the credibility keynote promises rarely earn on their own.
The customers on stage matched that posture. Cox Enterprises, Amer Sports, and Milo’s Tea described production deployments, not pilots. Cox’s next step of connecting OneStream to Dynamics and Workday transactional detail through MCP is the canonical case for agentic orchestration, and Cox discussed it as near-term, not aspirational.
Two things follow for higher education leaders.
Whether the institution is evaluating finance, HR, SIS, or research administration platforms, the question is no longer “does this vendor have AI?” It’s “does this vendor expose an agentic layer that honors our existing data, security, and workflow context, and when will it be generally available?” Anything short of GA on the agentic layer is a roadmap commitment, and we’ve all learned not to buy on promises (you are all tired of the safe harbor statements!)
A mature agentic architecture still depends on the underlying platform being configured properly, the data model being trustworthy, and the implementation partner being credible. The new capabilities raise the floor; they don’t lower the diligence when considering a new solution.
The vendors that win this cycle will be the ones that resist shipping a chatbot and calling it an agent. OneStream’s Splash 2026 was the first time I watched a vendor put real product against that bar. There will be others. The buyers who evaluate this well will be the ones who know what they’re looking at when those vendors show up. It will also inform the existing clients of vendors shipping new agentic functionality. Does it live up to these modern standards?
Originally posted by Dave Kieffer on LinkedIn. Be sure to follow him there to catch all his great industry insights.
Share Article:
© Copyright 2026, The Tambellini Group. All Rights Reserved.
Get exclusive access to higher education analysts, rich research, premium publications, and advisory services.
Weekly email featuring higher education blog articles, infographics or podcasts.
