OpenAI is the best foundation for frontier models, built-in tools, an Agents SDK, tracing, evaluations, and workspace agents. Microsoft wins inside Microsoft 365, Dynamics, Power Platform, and Azure. Salesforce Agentforce is strongest for Customer 360 workflows. Google Vertex AI and AWS Bedrock provide cloud-native model choice, while UiPath leads when agents must coordinate with enterprise robots.
Platform comparison
| Platform | Best fit | Building approach | Pricing dimensions | Primary drawback |
|---|---|---|---|---|
| OpenAI | Custom knowledge and action agents using OpenAI models/tools | Responses API, Agents SDK, built-in tools, workspace agents, Frontier | Model tokens, tools, storage, seats, enterprise agreement | Custom production controls still require engineering |
| Microsoft Copilot Studio/Azure | Microsoft-centered employee and business agents | Low-code studio, connectors, Power Platform, Azure agent frameworks | Copilot Credits/messages, seats, Azure usage, connectors, capacity | Licensing and overlapping products are complex |
| Salesforce Agentforce | Sales, service, commerce, and employee agents on CRM data | Topics/instructions, actions, Flows, Data Cloud, platform security | Flex Credits/actions, conversations, employee licenses, platform products | Expensive when data and process are not already in Salesforce |
| Google Vertex AI | Cloud developers building multimodal, data-grounded agents | Vertex AI Agent Builder, Gemini, search/data tools, Google Cloud | Model tokens, grounding/search, compute, storage, cloud services | Product map and consumption require cloud expertise |
| Amazon Bedrock Agents | AWS organizations wanting model choice and managed orchestration | Bedrock models, Agents, Knowledge Bases, Guardrails, AWS services | Model inference, knowledge/vector services, Lambda/API, storage | Integration and observability span many AWS services |
| UiPath | Enterprises combining agents, RPA robots, documents, and people | Agent builder/orchestration, robots, Integration Service, AI Trust Layer | Platform/agent units, users, robots, LLM calls, document/API capacity | Quote complexity and center-of-excellence overhead |
OpenAI: strongest general agent foundation
OpenAI’s Responses API combines model output with tools such as web search, file search, computer use, and custom functions. The Agents SDK supports orchestration and handoffs, while tracing and evaluations help developers inspect runs and test behavior. Business workspace agents can be shared across an organization with connectors and company context. Frontier extends the enterprise story toward deploying and managing agents across existing systems.
The platform suits proprietary workflows spanning more than one CRM or cloud application. Strong models, multimodal input, structured output, hosted tools, and libraries reduce orchestration code.
Pricing includes model input, cached input, output, built-in tool calls, file/vector storage, computer use or other resources, and external services. ChatGPT Business or Enterprise seats are separate from API usage unless a contract states otherwise. Frontier is enterprise-priced. Run real tasks because a long agent loop can call the model and tools repeatedly.
The weakness is that an API does not provide the complete operating model. Teams must build authentication, authorization, business validation, approval, retry, idempotency, monitoring, data retention, and incident response around the agent.
Our pick: OpenAI for custom cross-system agents with a capable engineering team
Microsoft: best for Microsoft 365 and business applications
Copilot Studio lets makers build agents using topics, generative orchestration, knowledge, connectors, tools, Power Automate, and Dataverse. Microsoft 365 Copilot can expose employee-facing agents in the daily productivity environment. Azure services support deeper development, models, data, security, and observability.
The advantage is identity and distribution. An agent can use Entra ID, appear in Teams or Microsoft 365, retrieve permitted SharePoint content, call Dynamics or Power Platform actions, and operate under tenant governance. Employee scenarios may be included under particular Microsoft 365 Copilot licenses when the authenticated user and agent meet current rules.
Copilot Studio meters many activities in Copilot Credits under current licensing. Different features consume different rates; official 2026 documentation, for example, lists computer-use agent actions at five Copilot Credits and premium voice at much higher rates. Azure model and service usage, premium connectors, Dataverse, hosted capacity, and Microsoft 365 Copilot seats can add cost.
The ecosystem is powerful but difficult to price. Ask Microsoft or a qualified partner to map the exact scenario, identities, actions, environments, and expected usage.
Salesforce Agentforce: best customer-data platform
Agentforce works with Salesforce objects, Data Cloud, knowledge, Flows, Apex or APIs, and platform security. Agents can answer product questions, qualify leads, prepare seller work, resolve cases, update records, and invoke approved business processes.
The product’s main advantage is proximity to customer state. A service agent can see the case, entitlement, account, order, and previous interactions rather than retrieving them through a custom integration. Salesforce administration, permission sets, audit, sandbox, and its partner ecosystem support enterprise deployment.
Pricing may use Flex Credits for actions, conversation-based structures, or employee-agent licenses. Salesforce defines an action as a function such as updating a record, summarizing a case, answering a question, or executing a prompt or flow. One customer outcome can consume several actions. Data Cloud, Service/Sales Cloud, integrations, and consulting may exceed the agent credit line.
Agentforce is a poor first choice when customer data is fragmented outside Salesforce. Fix identity, knowledge, permissions, and process ownership before buying autonomy.
Amazon Bedrock Agents: best for AWS model choice
Amazon Bedrock offers managed access to models from Amazon and third parties, plus Agents, Knowledge Bases, Guardrails, evaluation and related services. An agent can interpret a goal, retrieve knowledge, and invoke action groups backed by Lambda or APIs. AWS identity, networking, encryption, logging, and data services support enterprise architectures.
Model choice is the differentiator. A team can select among supported model families for cost, latency, region, or task rather than committing every agent to one provider. Bedrock is compelling for an organization whose data lake, applications, and security controls already live in AWS.
Costs include model inference, knowledge-base ingestion and retrieval, vector database, Lambda, API Gateway, storage, logs, networking, and other services. Provisioned throughput, batch modes, or caching may change unit economics. Cost allocation tags and per-agent telemetry are essential.
The drawback is assembly. A robust agent may span IAM, KMS, CloudWatch, Lambda, Step Functions, OpenSearch or another vector store, and application code. That flexibility requires ownership.
UiPath: best for agent-plus-robot operations
UiPath’s platform coordinates agents with attended or unattended robots, Integration Service, document processing, queues, orchestration, testing, and human approval. It is uniquely strong when an agent needs to interpret an exception and a reliable robot must operate a legacy Windows, mainframe, Citrix, or ERP interface.
AI Trust Layer and governance functions help centralize model access. Existing UiPath customers can reuse bot components rather than rebuilding every action as an API.
Licensing uses Unified Pricing or Flex structures and may involve Platform Units, Agent Units, users, robot capacity, managed-model LLM calls, documents, and integration calls. Official documentation shows different consumption for UiPath-managed and customer-managed models. Obtain a transaction-level quote.
UiPath is overkill for a startup connecting three SaaS applications. It earns its cost in regulated, high-volume estates with legacy automation and operational support.
Evaluate with one job, not a demo
Choose a job with 100 representative cases, such as preparing an account brief, resolving a standard support request, or processing a document exception. Define sources, allowed tools, expected outcome, approval, and prohibited actions.
Score task success, factual grounding, source citation, tool choice, parameter accuracy, permission compliance, escalation, latency, and cost. Include contradictory knowledge, missing data, tool failure, prompt injection, other languages, and transaction limits. Re-run after changing model, prompt, tool, or data.
Price median and 95th-percentile successful runs. Include loops, retries, escalations, model tokens, tools, vector search, third-party APIs, platform credits, infrastructure, seats, implementation, evaluations, monitoring, and human review.
Security requirements
Give each agent a dedicated service identity and least-privilege tools. Validate every action server-side. Treat retrieved pages, emails, documents, and CRM notes as untrusted content. Require approval for money, access, external messages, legal commitments, publication, or deletion. Log sources, tool arguments, results, approval, cost, and errors; set step, time, spend, and record caps plus a kill switch.
Verdict
OpenAI provides the best general developer platform for building capable custom agents. Microsoft is the best enterprise choice for Microsoft 365 and Power Platform, Salesforce for Customer 360 workflows, Google for multimodal Google Cloud systems, AWS for model choice in an AWS estate, and UiPath for combining agents with enterprise RPA.
Select the platform nearest to the authoritative data and approved actions. Run a 100-case evaluation and full cost model before granting write access. Model quality attracts attention; permission design, evidence, and operating discipline determine whether the agent belongs in production.