Platforms Clamp Down on Customer AI Agent Access
Major enterprise software platforms are moving to block or limit how outside artificial intelligence agents interact with their systems, setting up a structural conflict at the center of the broader push toward AI-powered automation.
As reported by The Information, platforms including Slack, Workday and LinkedIn have started restricting what customer-deployed agents can access and do inside their environments, citing data privacy, system stability and competitive concerns. The restrictions are arriving just as model providers and tooling companies roll out agent capabilities built for exactly the kind of cross-platform work these platforms want to slow down.
Platforms Are Tightening Access
The cutbacks are showing up across major platforms in rapid succession. According to Reuters, Salesforce tightened third-party access to Slack data in mid-2025, leaving outside applications rate-limited and blocked from storing historical messages long term. Enterprise tools built on that access, like internal copilots, knowledge search products and workflow automation, lost the data pipelines that made them useful.
The pattern extends beyond Slack. As covered by TechCrunch, Meta updated WhatsApp’s business API terms to ban general-purpose AI chatbots from the platform entirely, a policy that took effect in January and shut out assistants from companies including OpenAI and Perplexity.
Meta framed the decision as an API design issue. The company said that its business messaging layer was built for company-to-customer conversations, not as a distribution channel for third-party AI products, and the chatbot use case placed a load on its systems while falling outside the intended design. Meta AI is now a general-purpose assistant operating on a platform with more than 1 billion monthly active users, according to CBNC.
The operational risks for enterprises are not theoretical. As reported by VentureBeat, Google cut off access to its Antigravity coding platform for users who had connected OpenClaw, an open-source AI agent, through standard login integrations. Those users had been routing a disproportionate volume of Gemini token requests through the connection, overwhelming the platform for other customers.
Agents Can’t Wait for Platforms to Catch Up
While platforms tighten their perimeters, the agent infrastructure market is moving in the opposite direction. Arcade.dev’s ToolBench, a benchmarking platform for Model Context Protocol servers, scores and grades the quality of AI agent integrations across external services. The agent-to-platform connection layer is becoming an infrastructure to standardize.
As Andreessen Horowitz covered, modern AI products almost always depend on incumbents’ data accessed via APIs, and typical AI applications need information stored in systems of record to automate workflows. That dependency runs both ways. Agents need platforms for data, context and execution authority. Platforms need agents to keep workflows active and justify the value of their data assets to customers.
A similar analogy playing out in banking makes the stakes concrete. As reported by CNBC, JPMorgan processed 1.89 billion data requests in a single month from FinTech aggregators, with internal documentation showing just 6% of those API calls were tied to active customer transactions.
JPMorgan proposed fee schedules that could cost aggregators like Plaid up to $300 million annually and attributed $50 million in annual fraud losses to transactions initiated through data aggregators. Plaid ultimately agreed to pay for access but didn’t disclose the amount of the deal, according to Bloomberg.
The precedent is direct: Data pipelines built on free, open access can be repriced or restricted overnight, and enterprises running agents against platform APIs at scale are exposed to the same dynamic.
The larger stakes here are not whether platforms block agents outright. They are over who writes the terms on which access continues. Platforms are unlikely to revoke agent access entirely. Agents drive engagement, automate workflows and justify the ongoing relevance of platform data assets.
The more probable outcome is a bifurcated access model: human users on one pricing or access structure, AI agents on another. An agent querying Slack data thousands of times per hour is not the same as an employee running a search.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
The post Platforms Clamp Down on Customer AI Agent Access appeared first on PYMNTS.com.