Agentforce Alternatives: Every Route to AI in Salesforce

Five rules that narrow the field quickly:

1. If Salesforce should own the whole agent stack, including customer-facing and voice agents, and Data 360 is already on your roadmap, start with Agentforce.
2. If your users live in Outlook™ and Teams™ and mostly need to read Salesforce data there, evaluate Microsoft Copilot with the Salesforce connector or Sales agent.
3. If people already use ChatGPT or Claude, Salesforce Hosted MCP Servers are the quickest way to give them Salesforce access. Budget for Flex Credit usage on the Salesforce side.
4. If regulated data must be masked before any model sees it, shortlist only options that document masking: Prompt Builder used directly (not through agents), GPTfy, Oliva.AI, or your own build.
5. If the CFO needs a fixed annual number, per-user licences are easier to forecast than credits or tokens. Agentforce add-ons and Microsoft 365™ Copilot include usage in the seat price; GPTfy is per user, plus your AI provider's token charges. If you have fewer than 100 users, GPTfy's 100-user minimum rules it out.

For what Agentforce consumption actually costs, see Agentforce Flex Credits vs Conversations.

Ten options, seven criteria, one row each. Prices are list prices shown on each vendor's public page on 19 October 2026; your contract may differ. Scroll the table sideways on smaller screens.

"None documented" means the vendor's public documentation for that route does not describe masking data before it reaches the model. It does not mean the vendor has no other data protections.

Agentforce is Salesforce's own agent platform. Agents are built in Agentforce Builder and run on Salesforce-managed models. In Setup, an admin picks an org-level model option: the Salesforce default (agents created in the new Agentforce Builder use GPT-4.1), an AWS-hosted option using Claude Haiku 4.5 on Amazon Bedrock, or a Google option using Gemini 3.5 Flash. Custom actions that use a prompt template, Apex or the Models API can reference any Salesforce-managed or BYO model.

Where it wins

• First-party: one vendor for CRM, agents and models, with customer-facing agents, employee agents, Agentforce Voice and Slackbot on one price list.
• Anthropic and Amazon models run on Amazon Bedrock entirely inside the Salesforce Trust Boundary, and data sent outside it is covered by zero-retention contracts with the model providers.
• A per-user option exists: Agentforce add-ons from $125 per user per month ($150 for Industries clouds) give licensed employees unmetered agent usage.

Where it does not

• Salesforce Help states that pattern-based and field-based data masking is disabled for agents.
• Setup requires Einstein generative AI, which requires Data 360 to be provisioned and enabled. Salesforce's own pricing examples exclude Data 360 credits.
• With Flex Credits, spend follows the number of actions, so the budget depends on usage estimates.

Model detail: Which LLM does Agentforce use? Data detail: Does Agentforce require Data Cloud?.

Sources: Salesforce Agentforce pricing; Salesforce Help: select an Agentforce model option; Salesforce Developers: supported models; Salesforce Help: data masking for agents; Salesforce Help: turn on Agentforce; Salesforce Help: set up Einstein generative AI.

Prompt Builder is where admins create and manage prompt templates for Salesforce's generative AI features. Salesforce lists it for Enterprise, Performance and Unlimited Editions with the Einstein for Platform add-on, an Einstein or Agentforce for Sales or Service add-on, or Agentforce Foundations. Bring your own LLM, configured in AI Models (formerly Einstein Studio), adds models from Amazon Bedrock, Azure OpenAI, OpenAI or Vertex AI, or any endpoint through the LLM Open Connector.

Where it wins

• Masking applies: when a prompt template is used directly, Trust Layer masking runs according to your configuration. Salesforce's June 2025 blog confirms masking stays on for templates run in Prompt Builder or through Flow.
• Model choice inside Salesforce's own tooling: with BYO LLM, inference runs at your provider while requests still route through the Models API and Trust Layer.
• Salesforce's pricing page lists Prompt Builder in its free Salesforce Foundations tier.

Where it does not

• It runs prompt templates, not agents. Autonomous multi-step work comes from Agentforce, which can call a template as an agent action, and masking is disabled when it does.
• Setup follows the Einstein generative AI path, so Data 360 must be provisioned and enabled; AI Models is documented as part of Data 360.
• Usage is metered per prompt in Flex Credits, and Salesforce caps context at 65,536 tokens while masking is on.

Sources: Salesforce Help: Prompt Builder; Salesforce Help: set up Prompt Builder; Salesforce Help: generative AI usage and billing; Salesforce Help: add a foundation model (BYO LLM); Salesforce Help: AI Models; Salesforce blog: LLM data masking; Salesforce Developers: Models API data masking.

Salesforce Hosted MCP Servers let a compatible AI client (Salesforce names Claude, ChatGPT, Cursor or a custom agent) connect to an org and act on behalf of authorised users, using per-user OAuth. Salesforce announced general availability in April 2026 for Enterprise Edition orgs and above. Salesforce's setup guide connects ChatGPT through ChatGPT's developer mode.

Salesforce in Claude is a plugin that brings Salesforce accounts, opportunities and pipeline into Claude. Your Salesforce admin turns on the Salesforce MCP server for it, and it is in beta on all paid Claude plans for organisations Salesforce approves.

The Agentforce Sales app in ChatGPT is different: it is an Agentforce extension, not an alternative. It has been in open beta since December 2025 for Agentforce for Sales add-on and Agentforce 1 Edition customers, and Salesforce says data in the app is handled by the Agentforce Trust Layer.

Where it wins

• Fastest to switch on: Salesforce says setup takes less than 30 minutes.
• Users stay in the tool they already use. Salesforce in Claude only sees what each user's Salesforce permissions allow, and by default asks the user to approve each change before it is written.
• OpenAI and Anthropic both state that business-plan data is not used for training by default.

Where it does not

• Neither Salesforce nor the AI vendors document masking before the model on this route; records the client reads are processed in the AI vendor's service.
• Two bills: AI client seats, plus Salesforce usage. Hosted MCP Servers are intended for customers with Flex Credits, and Salesforce says MCP calls from registered agents will be metered as Headless Platform Interactions, with the rate listed as TBA.
• Salesforce in Claude is beta and needs the latest Sales Cloud enterprise edition and Salesforce approval.

Sources: Salesforce Developers: Hosted MCP Servers overview; Salesforce Developers: configure ChatGPT; Salesforce Developers blog: Hosted MCP Servers GA; Salesforce Help: AIforce impact; Claude Help: set up Salesforce in Claude; Claude Help: Salesforce in Claude; Salesforce: Agentforce Sales in ChatGPT; OpenAI business pricing; OpenAI enterprise privacy; Claude pricing; Anthropic privacy: commercial data.

Microsoft 365 Copilot, which Microsoft now calls Microsoft Copilot, reaches Salesforce two ways. The Salesforce CRM Copilot connector indexes Account, Contact, Opportunity, Lead, Case, Event, Task and Campaign records into Microsoft 365 so Copilot can answer from them. Sales agent works with Salesforce Sales Cloud Professional edition and above inside Outlook and Teams.

Where it wins

• AI where sellers already work: Outlook, Teams and the Microsoft 365 apps.
• The connector honours Salesforce record-level sharing rules and object permissions by default, and Sales agent applies existing CRM access controls.
• Sales agent is available as part of the Microsoft 365 Copilot licence, and Microsoft states prompts, responses and Graph data are not used to train foundation models.

Where it does not

• The connector copies and indexes Salesforce content into Microsoft 365 and Microsoft Graph, and covers eight standard objects. If an admin opts in to fields restricted by field-level security, Microsoft says the connector does not enforce that restriction in results.
• No masking before the model is documented, and Microsoft states Anthropic models are currently excluded from the EU Data Boundary.
• Copilot grounding on connector data, and Sales agent, need the Microsoft 365 Copilot add-on ($30 per user per month paid yearly, or Microsoft 365 E7) on top of a qualifying Microsoft 365 plan. With only a Copilot Studio or pay-as-you-go licence, the connector works in agents but not in Copilot itself.

Sources: Microsoft Learn: Microsoft Copilot privacy (product rename); Microsoft Learn: Salesforce CRM Copilot connector; Microsoft Learn: connector licensing; Microsoft Learn: Sales agent FAQ; Microsoft 365 Copilot pricing; Microsoft Learn: Anthropic as a subprocessor; Microsoft Learn: enterprise data protection; Microsoft Learn: Graph vs Power Platform connectors.

Copilot Studio is Microsoft's low-code agent builder. Its Salesforce connector is a Premium connector that supports SOQL and SOSL queries, record create, update and delete, and triggers. Creators pick the model per agent from a list that includes GPT models and, as generally available, Claude Sonnet and Opus models.

Where it wins

• Read and write: agents can query and change Salesforce records, not only search an index.
• Model choice inside Microsoft's platform, including Anthropic models, with external models from Anthropic, xAI or Mistral available once an admin turns them on.
• Pay-as-you-go is available at $0.01 per Copilot Credit, alongside $200 monthly packs of 25,000 credits.

Where it does not

• Microsoft's sensitive-data documentation states that data passed to generative AI and generative orchestration is not redacted.
• The connector needs Salesforce API access, which Salesforce trial accounts do not have, and is throttled at 900 calls per connection per 60 seconds.
• Credit consumption varies by agent action, so cost depends on usage.

Sources: Microsoft Learn: Salesforce connector; Microsoft Learn: select an agent model; Microsoft Learn: sensitive data; Copilot Studio pricing; Microsoft Learn: pay-as-you-go meters.

Salesforce developers can call any model API directly. A Named Credential holds the endpoint and authentication so Apex callouts do not handle secrets. MuleSoft offers an Inference connector for multiple LLM providers and an AI Chain connector for building agents in Anypoint Platform. Model providers bill per token: OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock and Google Cloud all publish per-million-token pricing.

Where it wins

• Full control over model, region, contract and data handling, with no AI licence beyond tokens.
• API data is not used for training by default at OpenAI or Anthropic, and both offer zero-retention arrangements to customers they approve.
• No Data 360 dependency for Apex callouts through Named Credentials.

Where it does not

• You build and maintain everything: prompts, masking, logging, user interface and model upgrades.
• OpenAI and Anthropic both keep API data for up to 30 days by default (OpenAI as abuse-monitoring logs) unless you are approved for a zero-retention arrangement.
• MuleSoft's Einstein AI connector routes through the Salesforce Trust Layer, so it is not an independent path.

Sources: Apex Developer Guide: Named Credentials; MuleSoft Inference Connector; MuleSoft AI Chain Connector; MuleSoft Einstein AI Connector; OpenAI API pricing; Anthropic API pricing; Azure OpenAI pricing; Amazon Bedrock pricing; Google Cloud generative AI pricing; OpenAI API data controls; Anthropic commercial terms; Anthropic privacy: API data retention; Anthropic privacy: zero-retention arrangements.

GPTfy is a 100% Salesforce-native managed package on AppExchange, with a Microsoft AppSource app for Microsoft 365. Each model is connected through a Named Credential to your own provider: OpenAI, Azure OpenAI, Anthropic Claude, Google Vertex AI and Gemini, Amazon Bedrock, Mistral, Llama, DeepSeek, or a model you host. Raw data stays in Salesforce, only masked data reaches your AI provider. Each interaction writes a Security Audit Record in Salesforce.

Where it wins

• 4 layers of data masking (field value, regex pattern, blocklist and Apex) applied before the call.
• GPTfy's AppExchange listing states no Agentforce or Data Cloud is required, and model choice is yours, including models on your existing Azure, AWS or Google Cloud contract.
• Per-user pricing with unlimited prompts: Pro $20, Enterprise $30, Unlimited $50 per user per month. AI Agents start at the Enterprise plan. GPTfy does not charge per token.

Where it does not

• A 100-user minimum per plan, so small teams will usually pay less with per-seat AI clients or consumption pricing.
• Two bills: the GPTfy licence plus your AI provider's token charges.
• GPTfy states its masking covers 16 of the 18 HIPAA PHI identifiers; it cannot mask biometric identifiers or full-face photographs.
• It is not Salesforce's agent platform. If your roadmap is built on Agentforce Builder and Agent Script, Agentforce is the direct route.
• Masking rules are configured by your admin per field, pattern and blocklist, and GPTfy states go-live in 2 to 3 weeks. The MCP route is quicker to switch on.

Head-to-head detail is on GPTfy vs Agentforce. Plans are on GPTfy pricing, and the masking layers are explained on data masking.

Sources: GPTfy pricing; GPTfy on AppExchange; GPTfy on Microsoft AppSource.

Several other managed packages run generative AI inside Salesforce. Three with public listings that document their model approach:

Ortoo Orchestrator runs case triage, lead routing, claims and email intake as workflows with specialised agents per step. It uses any LLM through your own API keys and says it works with or without Agentforce. It wins on bounded, per-work-item pricing for operational queues; it describes choosing what each step sends to the model rather than masking, and pricing is quote only.

Peeklogic AI Orchestrator offers a library of up to 40 AI models and tools, including image and video generators, with your own keys or Named Credentials. It wins on breadth of tools and published pricing ($8,000 to $14,000 per company per year, 30-day trial); it describes content moderation rather than masking, requires Sales Cloud, and some integrations run through Einstein Studio and Data 360.

Oliva.AI from LG CNS connects Salesforce to OpenAI with your own API key, and its listing describes automatic personal-information masking with saved masking history. It wins for OpenAI-only Sales Cloud orgs that want masking and published band pricing ($900 to $16,000 per company per month); its listing names OpenAI as the only supported provider.

Sources: Ortoo Orchestrator on AppExchange; Ortoo Orchestrator; Ortoo Orchestrator features; Peeklogic AI Orchestrator on AppExchange; Peeklogic AI; Oliva.AI on AppExchange.

Last reviewed: 19 October 2026. Vendor facts were checked against public documentation on that date and are re-audited at least quarterly. Beta programmes and list prices change often, so confirm current terms with each vendor before you buy. The trademark and affiliation disclaimer is at the foot of this page.