Est.

How Big Companies Choose AI Agent Platforms

Platform choice decides who controls your data, workflows, and costs.

Publication desk · · 9 min read
Cover illustration for “How Big Companies Choose AI Agent Platforms”
AI Commoditization · November 11, 2026 · 9 min read · 1,967 words

Choosing an AI agent platform is a decision about who controls a company’s workflows, data, and future costs. For enterprises looking for cost optimization software, the best fit connects the systems where money is spent, enforces approval controls, and shows that savings exceed operating costs.

Why choosing an AI agent platform is structurally different from buying enterprise software

In a little over two weeks in late 2026, Google, OpenAI, and Workday each told enterprise customers that their agents should live on its platform. Each meant something different by that claim. The speed of the announcements exposed the contest underneath them. Vendors were competing for the control point where a company decides which model does a task, what an agent may touch, and who gets paid.

Conventional software procurement compares features, price, and integration effort on the assumption that switching later will be expensive but possible. Agent platforms accumulate dependencies that reach deeper into how work gets done. Teams learn a platform’s orchestration logic, agents follow its permission model, and business context settles into its data layer. Decision history may live nowhere else. Switching then means more than replacing software, because the company must recover the knowledge and operating rules it has built around it.

How the five dominant platform bets divide the market

The market makes more sense as five bets about where control belongs than as a ranking of interchangeable products. Google Cloud describes Agent Gateway as “an AI network firewall” through which traffic between agents and outside systems passes, enforcing the organization’s policies. Workday puts control in a different place. It opens its HR and finance systems to other companies’ agents on the condition that they obey Workday’s permissions, betting that its data and rules are harder to replace than a model.

Anthropic’s bet is to supply many environments independently of owning one. Claude runs inside Google’s agent and Microsoft’s Copilot, while Anthropic also sells Managed Agents and backs a services company with Blackstone, Hellman & Friedman, and Goldman Sachs. Distribution is the source of leverage here. Kimi takes a different approach with open-weights models that customers can run locally on their own hardware. That gives buyers a path to data sovereignty rather than requiring them to put their agents inside another vendor’s managed environment.

The fifth bet opens the model layer while keeping the business platform around it proprietary. SAP and Salesforce illustrate this approach, in which models can change but the data, context, and orchestration remain tied to the platform. Model choice does not remove that dependency. A buyer comparing only model quality or price can therefore miss the more durable commitment being made around the model.

What governance and approval architecture actually requires at enterprise scale

Governance is a condition of production eligibility, especially in regulated industries. An agent that makes autonomous, multi-step decisions across live systems needs audit trails, explainability, and data provenance built into the product. These cannot be postponed until after deployment.

Kore.ai’s enterprise guide identifies built-in audit logs, permission frameworks, context-aware risk scoring, compliance reporting, and guardrails as capabilities that in-house teams cannot safely replicate at production scale. Without them, the enterprise inherits the work of building its own governance infrastructure. That costs more and makes deployment take longer. It also creates gaps that an audit can expose, even when the agent itself appears to perform its task well.

Approval controls set the rules before an action, while audit logs record what happened afterward. Reading a system and changing it are separate permissions, and nothing should change without a human approval step. A log is not a substitute for permission. Platforms that cannot enforce that separation leave the enterprise to add it on top.

The same control problem spreads beyond one deployment when departments buy AI tools on their own that do some of the same work. Multiple contracts create duplicate spending, weaken vendor leverage, and spread audit exposure across separate systems. Marketing labels offer little protection against this fragmentation. A chatbot or RPA script relabeled as an agent still needs to show branching decisions, tool use, context retention, and failure recovery before buyers can trust what it claims to do.

Why integration depth determines whether an agent delivers value or just runs in isolation

Once a platform passes the governance test, its usefulness depends on what it can reach. An agent confined to one system will tend to find the problems that system’s users already know about. The gaps between systems are less visible. Purchasing may see one number while support and engineering see another, leaving costs that no single team is responsible for recovering.

Salesforce Agentforce’s value narrows outside the Salesforce ecosystem, while SAP-heavy or mixed environments face additional integration work. Microsoft Copilot Studio has a different advantage. Its native presence in Teams and SharePoint, integration into Dynamics 365, and access through Microsoft Graph suit Microsoft-stack enterprises, but that fit does not extend cleanly to ERP-led environments.

The lesson for buyers is to test connections to their actual systems of record, including ERP, HRIS, CRM, finance, and support. A connector claim is only the starting point. Buyers need to check which connections are pre-built and which need custom engineering with its own budget and staff. For a system requiring a new connection, a platform should be able to build it from that system’s documentation and verify it in the enterprise’s own environment. Otherwise, the unfinished work on the connection also limits the business case.

Data ownership and model flexibility as long-term cost and control variables

Integration establishes what an agent can reach today, but ownership determines what the company retains over time. Business context, decision history, and learned patterns are among the most valuable outputs of an agent deployment. Yet platform contracts often leave that value with the vendor. The company puts more money at risk as more of its work relies on knowledge it cannot easily take elsewhere.

Model flexibility addresses part of that risk, but not all of it. By mid-2026, Microsoft and AWS had each broken their model exclusivity arrangements, with Claude available inside Microsoft 365 Copilot and AWS Bedrock hosting OpenAI models. The model layer was becoming more interchangeable. Data, orchestration logic, business context, and agent runtime could still remain proprietary, so a choice of models did not necessarily offer a choice of platforms.

For buyers, model portability should let them switch models without losing what the system has learned. It also creates room to route simpler queries to cheaper models as the market changes. Deployment choices belong in the same negotiation. Enterprises should require their own cloud, their own servers, or a vendor-neutral cloud with client data never mixed, plus a decision trace covering the evidence, reasoning, approval, action, and result for each agent action.

Sovereignty makes the trade-off especially clear. Kimi’s local open-weights approach lets customers choose where the model runs, but they must provide the infrastructure and keep up with model maintenance. The exit test is blunt. If the company stops using the platform tomorrow, it should keep its data, decision history, and company knowledge in a form usable independently of the vendor’s proprietary format.

Why does total cost of ownership bear almost no relationship to license price?

An agent’s purchase price reveals little about the cost of operating it. A single request can expand into model calls for planning, retrieval, tool use, and retries, so the economics differ from a simple question-and-answer system. Each additional step can add inference costs. So a workflow that looks cheap in a demonstration may cost very different amounts to run in production.

Agentforce’s conversation-based pricing and Copilot Studio’s credit-based pricing tie spending to activity. Greater containment can increase the volume handled by the system, expanding the bill even as fewer requests reach people. License-free software avoids only one part of this equation. LangGraph still requires engineering headcount to deliver production-grade governance, reliability, and observability.

A sound ownership model must account for inference and token consumption, data pipelines, integration, human oversight, security middleware, correction, monitoring, and implementation. Ongoing AI and people costs belong there too. Governance requires a continuing budget because oversight and regulatory exposure create work during setup and ongoing operation. A lower license price may simply move spending into these less visible categories.

To prevent that shift from becoming a surprise, buyers should agree on the baseline, the savings definition, and the measurement owner before work begins. Estimates must stay separate from results. Time saved must also stay separate from cash saved, because freeing someone’s time does not by itself remove an expense.

How to measure real business outcomes rather than adoption metrics

User counts, agent invocations, and containment rates describe activity. A CFO or board needs to see cash recovered, revenue protected, or cost removed in the same currency as the original business case. Adoption can support that explanation. It cannot replace evidence that the company is better off after paying to run the system.

Operational measures become useful when paired with a financial outcome. IBM’s AskHR illustrates that framing by pairing containment with a reduction in HR operating costs against a starting baseline. Klarna’s assistant offers another example, with work done more efficiently and lower operating expenses in the same reporting period. These operational claims are useful as examples when they connect changes in work to changes in cost.

That connection depends on conditions that platform demonstrations often leave unresolved. Data quality gaps, unclear ownership of edge cases, and missing governance infrastructure can undermine deployment and make its outcomes difficult to measure. Responsibility must be explicit before work starts. A sound measurement agreement sets out the goal, the fee or cost, and how savings will be calculated, while counting implementation, AI, human review, correction, and monitoring together.

The comparison also needs an alternative to immediate action, whether that is waiting, gathering more evidence, or changing nothing. Without that comparison, doing something can seem valuable just because it was done. Build-versus-buy follows the same logic. Building fits when the agent’s logic is a competitive differentiator and the organization can staff a permanent agent operations team. Buying fits a high-impact process that is not differentiating, avoiding the expense of creating a platform where ownership offers little competitive return.

What a rigorous platform evaluation looks like in practice

A rigorous evaluation turns these criteria into a sequence of gates, starting with governance and approval architecture, followed by integration depth, then data ownership and model flexibility. Total cost follows those checks. Outcome measurement completes the evaluation by establishing whether the proposed platform can support a defensible business case before budget is committed.

The test should use real workflows that need branching decisions, tools across multiple systems, context kept from one step to the next, and recovery from failure. Feature checklists cannot demonstrate that behavior. The results also need a written baseline agreement covering the goal, cost, and savings calculation, so success is not defined after the deployment has already consumed money.

Platform fit then becomes more specific than a general claim to be best. Agentforce, with its Atlas Reasoning Engine and Einstein Trust Layer, is the first-choice fit for CRM-native workflows where Salesforce is the system of record. Copilot Studio fits Microsoft enterprises with employee-facing IT, HR, and knowledge workflows. These recommendations depend on the surrounding systems. Neither removes the need to test connections beyond its native ecosystem.

For regulated organizations that need a voice-first contact center, Cognigy (NICE) and Kore.ai offer on-premises and private cloud deployment with governance built in. Organizations that want model sovereignty and freedom from a vendor’s data layer can look at open-weights deployment or LangGraph instead, though they must take on the engineering work. Full control is not free of operating work. Finally, multi-year commitments need performance gates and terms for leaving the contract, because what platforms can do and what they cost change on six-to-twelve month cycles.

More in AI Commoditization