Why Your AI Platform Should Work With Any Model or Tool

Vendor lock-in raises AI costs and strands institutional knowledge.

Publication desk · · 9 min read
Cover illustration for “Why Your AI Platform Should Work With Any Model or Tool”
Agentic AI Containment · November 9, 2026 · 9 min read · 1,954 words

Agentic workflow automation for enterprise needs a platform that can work across business systems, use different models and tools, and keep decision history when those choices change. Locking the platform to one vendor gives that vendor control over costs, capabilities, and institutional knowledge. The best architecture keeps those assets under the enterprise’s control and uses routing, spending limits, and human approval to govern what agents do.

Why AI inference costs have become a board-level problem in 2026

Uber used up its whole AI coding-tools budget for the year within four months. Microsoft had started cutting Claude Code licenses across one of its major divisions by June 30, 2026. As AI moves from experiments into production, inference has overtaken training as the dominant cost category, growing with each new workload and user. Platform architecture now shapes an expense that reaches well beyond the engineering budget.

Agentic workflows make that expense harder to predict. A chatbot query might need one model call, while an agentic workflow can set off sub-agents, tool calls, and repeated reasoning loops before it finishes a task. EY research shows that modern agentic systems cost vastly more per interaction than simple 2023-era linear workflows. The bill follows behavior as well as user numbers. Giving the same employees more capable automation can therefore increase consumption without any change in headcount.

Finance teams were built to manage project-based budgets, while inference spending changes with every task. Procurement can negotiate a price, but it cannot control how often an agent calls a model or whether that model is unnecessarily expensive for the work. That mismatch makes this a governance problem. Financial evaluation now includes both what an AI subscription costs and what the platform allows a completed workflow to consume.

What lock-in actually costs when the model beneath your platform changes

NexGen Manufacturing spent $315K migrating 40 AI workflows to a new platform after Builder.ai’s collapse. The migration consumed three months of engineering time, during which customer-facing AI features were degraded or unavailable. The subscription was only part of the exposure. A multi-provider abstraction layer would have prevented that migration cost by keeping the workflows from depending on a single provider.

Enterprise migration projects can cost hundreds of thousands of dollars and take multiple months of engineering work. Across dozens of workloads, exposure can reach into the millions. At the agentic layer, total ownership costs can therefore be a multiple of the advertised subscription price, because replacing the platform means more than buying another license.

Lock-in also accumulates across vendors through multiple commitments. Each dependency reduces the freedom to negotiate, replace a service, or adopt a better model without disturbing existing work. If a stack is poorly prepared, a model change brings months of regression risk. A well-prepared one can make it a routing-rule change instead, leaving the workflow independent of the provider that performs the inference.

The strongest objection is straightforward. Why plan to switch after choosing the best model on offer? Because model leadership changes within months, while workflows and contracts can outlast that lead. A platform tied to one model makes the enterprise wait for its vendor’s upgrade cycle, even when another model offers a better fit for the task. The original choice can be sound and the dependency still be costly.

What enterprises lose beyond money when their AI knowledge is tied to one vendor

Approved workflows, correction history, and domain-specific routing rules are assets of the business. They record what the business has learned. When that knowledge lives only inside a vendor’s infrastructure, switching platforms risks losing more than a connection to a model. The enterprise can lose the accumulated context that made its automation useful, forcing the new platform to start without the same memory of past decisions.

Decision history also has a separate governance role. Finance and compliance teams need to reconstruct what evidence was used, what was approved, and what changed. Without those traces, an outcome may be difficult to explain in an audit or board report, even if the workflow appeared to succeed. A result alone is not enough evidence. The platform must preserve the reasoning and approvals that make the result defensible.

Model choice belongs within that retained knowledge and under the business’s control. Schneider Electric’s Chief AI Officer has publicly described a policy of using the smallest model that clears the bar for each solution. That policy requires a platform capable of acting on it. Hardwiring a single model prevents the business from applying its own judgment about what each task needs.

When the central platform cannot do what departments need, shadow AI has more appeal. Departments procure their own tools, spreading learned context across more systems and making costs harder to see. Fragmentation then undermines the case for further investment. Poor visibility into AI spending is causing enterprises to delay or cancel projects, sacrificing the productivity gains those projects were meant to deliver.

How model-agnostic routing works and what it recovers in practice

Model-agnostic routing puts an internal interface between workflows and model providers. Requests pass through that interface, which directs each task to a model suited to its cost, quality, and speed requirements. The cheapest adequate model gets the work. Classification, extraction, and summarization often do not need a flagship model, so using one by default pays for capability the task does not require.

RouteLLM, developed by researchers from UC Berkeley, Anyscale, and Canva and presented at ICLR 2025, provides specific evidence for this approach. It trained a router on human preference data and substantially reduced benchmark cost while preserving most of flagship model quality. Production teams report bill reductions of 40 to 85 percent from intelligent routing, with rule-based decisions adding negligible latency, though those results are not guaranteed for every workload.

Routing is only one source of savings. Prompt caching cuts out the cost of using the same context again, while batch APIs cut costs for work that can wait for a response. Combined on stable, recurring workloads, these techniques can bring effective per-call cost to roughly a quarter of the on-demand rate. Writer’s research also found that redesigning its harness cut token spending considerably at steady accuracy. Changing the vendor is not the only way to improve the economics.

But it is more useful to measure the cost of each task completed successfully. A resolved ticket, closed invoice, or merged pull request connects spending to an outcome that finance can recognize. Cheap tokens do not establish that the work succeeded. Routing therefore needs governance that can track both what a workflow consumed and what it actually accomplished.

The governance layer that makes model flexibility work at enterprise scale

Without controls, the freedom to use different models can also lead to sprawl. Good governance covers five dimensions, starting with enforcement at the inference layer. Budget caps and routing decisions apply before requests reach a model. Control starts before the money is spent. This gives the enterprise a way to act on its cost policy during execution.

The second dimension is per-request attribution. Each call needs identity, team, model, and environment metadata so spending can be assigned to the work that produced it. The third is agent-level protection, including loop detection and per-task budget limits. Autonomous agents can trigger hundreds of inference calls within one workflow. Circuit breakers stop that consumption from compounding unchecked, while attribution makes legitimate consumption explainable.

The fourth dimension covers GPU and compute management for self-hosted workloads, where appropriate sizing and autoscaling prevent idle capacity from quietly accumulating costs. The fifth is visibility across providers. Enterprises that run workloads across several clouds and model providers need attribution that brings all the parts of the bill together. These controls need rules above individual department budgets. Otherwise, governance defaults to whoever funds each tool, reinforcing the fragmentation that obscures total spending.

A read-only starting point separates permission to inspect information from permission to change it, allowing governance to be established before a workflow enters production. A human should have to give explicit approval for changes at each step. Audit logs and decision history can remain in the customer’s own cloud or on-premise deployment. That control must survive a model swap, so changing the underlying model does not require rebuilding workflows or surrendering institutional knowledge.

What the tool landscape actually covers — and what it leaves to the platform

Enterprise AI cost tools cover different parts of the bill, including infrastructure, SaaS spending, and inference. Most enterprises need coverage at all three layers, but the tools are not interchangeable. CAST AI targets Kubernetes and the infrastructure beneath it. Zylo and Tropic address software and AI spending. Those roles differ from controlling which model handles a request and how much an agent consumes while completing it.

Within the inference layer, observability and enforcement also differ. Langfuse provides open-source, token-level tracing, while LangSmith gives visibility into what models are doing rather than serving as an enforcement platform on its own. Helicone combines observability with gateway-level features, including rate limiting, caching, and routing controls. Visibility explains activity. Enforcement can constrain it. A comparison that treats every tool as merely an AI cost dashboard misses that practical distinction.

These tools can show how infrastructure is used, what software costs, or what models do, with gaps in their coverage of money that falls between departments because each system holds only part of the picture. An inference trace is not a cross-system investigation. The platform still needs to bring together business context from across those systems and keep the evidence behind what it finds.

For a buyer asking how AI agents work across enterprise resource planning, customer relationship management and human resources systems, that distinction is central. Access to several models does not establish access to several business systems, and neither establishes permission to change them. The platform’s cross-system scope and its approval boundaries are separate capabilities to evaluate. Good governance starts read-only, requires explicit approval for changes, and keeps audit logs and decision history in the customer’s hands. Model flexibility then operates within those boundaries.

How can you tell whether a platform really delivers model flexibility?

A vendor’s claim of model flexibility is testable before a contract is signed. Which models can the platform actually run, and what happens when the active model changes? The vendor needs to explain whether workflows stay intact when the model changes, as well as how many model names the product lists. A rebuild is still a migration. Multiple providers on a marketing page do not prove that the architecture separates workflows from those providers.

After the swap, can the customer still use and control approved decisions, correction history, and routing rules, or do they stay behind with the vendor? The answer exposes an exit cost that subscription pricing does not show. Buyers can also ask whether read and write permissions are separate, and whether every change requires explicit approval. Those questions distinguish access to information from authority to act on it.

Operational evidence is just as important as portability. Can the platform attribute every inference call to a team, workflow, and business outcome, with controls operating before the invoice arrives? Dashboards that show spending after it happens help explain the bill but cannot stop a runaway agent from running it up. Does the platform also work across systems, or only inside one? That answer establishes whether it can investigate gaps between departments.

Finally, the buyer can ask what the vendor will commit to measuring. A 72-hour findings guarantee is a concrete delivery promise, but a finding is not yet a defensible saving. Before work begins, the goal, baseline, savings method, measurement owner, and time period should be agreed in writing. What evidence will demonstrate a successfully completed task, and who will stand behind the calculation of what it cost?

More in Agentic AI Containment