There are roughly twenty platforms a serious enterprise will consider for agent deployment in 2026, and they fall into three camps with three different lock-in profiles. The choice is less about features than about which set of switching costs you are willing to live with.
The three camps
The first camp is ecosystem-native platforms: agent-building infrastructure that lives inside an existing application platform. Salesforce Agentforce, Microsoft Copilot Studio, ServiceNow Now Assist, and (depending on how you count) Workday's agent suite all live here. They are powerful where you already are. They are difficult to escape.
The second camp is cloud-provider platforms: agent infrastructure tied to a hyperscaler. Google's Vertex AI Agent Builder (rebranded as Gemini Enterprise Agent Platform at Cloud Next 2026), AWS Bedrock AgentCore, and the Azure Foundry/Agent Framework path are the three you will compare. The lock-in here is to the cloud, not the application — and most enterprises are already locked-in to one cloud anyway.
The third camp is horizontal automation platforms with AI capabilities: n8n (open source, self-hostable), Zapier Agents, Make.com, and UiPath at the RPA-extending end. Their lock-in is the lowest because the surface area of integration is smaller and more standardised — but they also offer less in exchange. This is the right camp for tech-forward teams that prize sovereignty.
Ecosystem-native platforms
Salesforce Agentforce hit $500M ARR growing 330% year-over-year in Q4 2025 — Salesforce's fastest-growing product ever, with 8,000+ deals closed and nearly half of the Fortune 100 using Data Cloud + AI. Pricing is typically $0.10/action or $125–$550 per user per month. The Atlas Reasoning Engine powers autonomous decisions inside CRM workflows. The honest caveat: Agentforce's value depends on having clean, unified data in Salesforce Data Cloud first. Organisations that have not solved their CRM data problem will not solve it by buying agents.
Microsoft Copilot Studio integrates natively with Microsoft 365, Azure, and Power Platform, accessing SharePoint, OneDrive, Teams, and Outlook via Microsoft Graph without separate data migration. At Ignite 2025, Microsoft declared the "Agentic Enterprise" era. The honest caveat: Copilot Studio is at its best when your authoritative data already lives in Microsoft tenancies. If your tenancy is split, the platform's value drops.
ServiceNow Now Assist ranked first in Gartner's 2025 Critical Capabilities for AI agents. The AI Agent Orchestrator plans, reasons, and coordinates multiple agents end-to-end, and the AI Control Tower provides centralised governance even across competing platforms (including Microsoft Copilot Studio integration). Subscription revenue of $3.47B in Q4 2025 reflects sustained AI platform momentum.
Horizontal automation
n8n is self-hostable, open source, and supports LangChain integration with 70+ AI nodes. It charges per execution, not per action — which produces approximately 80–90% cost reduction versus Zapier at high volume. It is the right choice for teams with engineering capacity, data sovereignty requirements, or anyone who has been bitten by per-action pricing escalation. The trade is operational cost: you run it.
Zapier has 8,000+ app integrations and is the easiest tool in the field to operate. It charges per task, which means each step costs. It is the right choice for non-technical teams with moderate volume. It becomes the wrong choice — economically — at high volume, and the migration off it once you are deep can be painful.
Make.com sits between the two: visual canvas, 2,000+ apps, charges per operation. The right balance for mid-size teams that want a visual builder without n8n's operational burden, and at substantially lower cost than Zapier at scale.
UiPath is the right starting point for organisations with a heavy RPA estate, where the existing process-mining and bot infrastructure is the natural foundation for an agent layer. The advantage is that the surrounding governance — bot inventory, RPA exception handling, change control — is already built. The disadvantage is that pure-AI agents are not UiPath's centre of gravity, and the platform's evolution is one to watch rather than to bet on.
Lock-in: the question that survives the demo
The demo is always good. The question that survives the demo is: can you export your agents, their configurations, and their training data? Ecosystem-native platforms are designed for maximum integration with their host ecosystems — which also maximises switching cost. The pattern is reliable: lower initial integration costs, higher long-term lock-in costs.
The mitigations a sophisticated team applies: design agents against standardised interfaces (MCP, A2A) where the platform supports them; keep prompts model-agnostic where possible; maintain training data and eval datasets in vendor-neutral formats; build a thin abstraction layer over the platform-specific APIs your agents use most. None of these prevent lock-in. They reduce the cost of the day you decide to move.
For each platform you are seriously considering: if you decide in two years to move off it, what does the migration look like? Whose problem is it? What does it cost? If the answer is "we'd rebuild from scratch", you are buying more lock-in than the demo suggests. That may still be the right trade. Just buy it deliberately.
The next chapter takes the platform decision and asks the harder question: does the agent pay back? The arithmetic is mostly written by the cost stack, and the cost stack is mostly invisible.