GS
Most discussion of federal AI moves between two surfaces. The policy layer — OMB Memorandum M-24-10,[1] the NIST AI Risk Management Framework,[2] the executive orders — and the model layer, where the public conversation tracks which foundation models agencies selected, which vendor's chatbot landed in pilot, which contract just got awarded. The layer between them — where AI actually runs when it runs — gets almost no attention. It is the integration layer, and the platforms operating it are the most important federal technology category nobody is writing about.
Where federal AI actually runs
The technical architecture of a federal AI deployment is mostly not the model. It is the integration platform that lets the model touch anything.
Federal agencies do not run AI on greenfield infrastructure. They run AI inside organizations with decades of accumulated systems — legacy ERP, 30-year-old custom applications, identity infrastructure that predates cloud, records environments built before the iPhone. An agentic AI system in a federal context has to reason across those systems, take actions inside them, retrieve and update records held inside them, and produce auditable trails of everything it did. The foundation model does none of that. The model is a reasoning engine that produces tokens. The middleware is what turns those tokens into actions in a federal environment.
The category name for this middleware is iPaaS — integration platform as a service. Boomi, MuleSoft, Workato, Informatica, and a handful of competitors. Each is a complete platform for moving data between systems, managing API contracts, transforming records as they flow, orchestrating workflows that cross system boundaries. Each has been quietly added to the federal vendor mix over the last several years as agencies have figured out that integration is a binding constraint on every modernization effort.
Federal AI press coverage and federal AI engineering effort optimize for different layers of the stack.
The bottom of the stack does the work. The top of the stack gets the headlines.
What iPaaS solves that models can't
Foundation models, even the best ones, are stateless reasoning engines. Federal AI deployments require something fundamentally different: persistent, governed, auditable interaction with operational systems. Four capabilities matter at federal scale, and the model doesn't deliver any of them.
The first is system connectivity at federal breadth. A federal AI agent in production may need to read from a Documentum content repository, write to an Oracle HR system, query a mainframe-backed records system, post to a ServiceNow ticket, and notify Microsoft 365 — all within a single workflow, all governed by different identity systems, all auditable to different standards. The integration layer is what makes any of this possible.
The second is state management. Foundation models hold no state between calls. Federal workflows are stateful — a benefits eligibility process unfolds over weeks; a procurement spans months. The middleware holds the workflow state the model needs to act consistently across that timeline.
The third is governance and audit. Every action an agent takes in a federal environment must be auditable, reversible, and explainable. The model produces a decision; the middleware records the decision, captures the context, ties it to the operator's identity, and makes the whole chain queryable months later when an inspector asks. Without that layer, the agent is unauditable, which means it is unauthorized.
The fourth is federal-grade controls. The middleware is where FedRAMP boundaries get enforced, where data classification rules are applied, where rate limits and circuit breakers prevent runaway agents from creating thousand-record errors. The model has no awareness of any of this; the integration platform is where federal control surfaces live.
The vendor landscape nobody is mapping
The iPaaS market in federal is real, growing, and almost completely unwritten about by the policy and industry press. A handful of vendors have built substantive federal practices. Most agencies have already made a primary platform commitment — or are about to.
No single platform dominates federal integration. The platform decisions being made now will lock agencies into ten-year vendor relationships.
The leader differs by sector — and most agencies haven't picked yet.
Logic Apps / Power Platform
Salesforce
webMethods / App Connect
Workato, SnapLogic, Tibco, custom
The footprint patterns differ by sector. Defense workloads tend toward platforms with strong on-premise and air-gapped deployment options. Federal civilian agencies are more cloud-native and weight FedRAMP-High coverage heavily. Federal infrastructure operators (postal, transit, logistics) prioritize integration with legacy ERP and complex workflow orchestration. No single platform dominates across all three sectors, which means the federal iPaaS market is genuinely contested — and the platform decisions being made now will lock agencies into ten-year vendor relationships.
"The federal AI procurement that matters most is not the foundation-model award. It is the integration-platform commitment underneath, which costs less to procure and matters more to outcomes. The asymmetry of attention to cost reveals where the conversation is broken."
The cost inversion most procurements miss
Federal AI procurement narratives center on model cost. Foundation models are visible, branded, and easy to compare. Integration platforms are commodity-named, less branded, harder to compare, and usually scoped as a line item buried inside a larger SOW. The cost reality is the inverse of the visibility.
Foundation models are 8% of total federal AI program cost. Everything else is 92%.
The most visible line item is the smallest. The largest line item — integration — is rarely scoped on its own.
Across federal AI deployments, the foundation model is rarely more than 10% of total program cost. Integration, content and data preparation, governance instrumentation, workflow operations, and change management together account for the other 90%. Procurement organizations that anchor evaluation around model selection are optimizing 8% of the program against 92% of the risk.
The corollary is that vendor differentiation that lives at the model layer transfers poorly to federal outcomes. A vendor with marginally better benchmark scores on a foundation model but no federal integration practice will lose to a vendor with adequate model selection and deep iPaaS muscle. Federal AI procurement is becoming a middleware procurement with a model attached.
What this changes
The pattern reshapes how federal technology leadership should be thinking about AI program design through 2027 and into the next budget cycle:
- The model is not the procurement. Foundation-model selection is necessary but increasingly commoditized. The integration platform commitment is where the real ten-year decision is being made — vendor lock-in, architectural reversibility, operational tempo, and audit posture all come down to the iPaaS choice, not the model choice.
- Vendor evaluations led by model benchmarks are evaluating the wrong layer. Replace or supplement model-benchmark scoring with iPaaS depth-of-practice scoring: federal-specific connectors built, FedRAMP boundary handling, audit trail completeness, operations runbook maturity. A vendor with strong middleware credentials and a competent model partner will outperform a vendor with the reverse.
- The integration layer is where federal AI fails when it fails. Most federal AI pilots that stall do not stall at the model. They stall at the integration handoff between the agent and the real systems the agent needs to touch. Where pilots stall is where the next vendor selection should focus, not where the press release focuses.
- The platform choices being made now will compound for a decade. Federal agencies that pick an iPaaS partner well in 2026 will run their next decade of AI programs on that platform's connectors, audit infrastructure, and operational patterns. Agencies that pick poorly will pay for the mistake in every subsequent AI initiative.
A federal AI agent doesn't act on the model. It acts across the systems the middleware connects.
Eight federal system surfaces. One agent. The lines between them are the platform decision.
The decision
The federal AI conversation worth having in 2026 is not about which model an agency chose. It is about which integration platform that agency chose, and what that platform is going to make possible — or impossible — over the next ten years. The agencies that recognize the middleware layer as the binding constraint are making the decision deliberately. The agencies that don't are making it accidentally, through a sequence of small procurement decisions that quietly add up to a platform commitment. Both groups end up with a primary iPaaS platform; only one group chose it.[4]
