From conversation to execution
A model can decide the next step.
Your system decides what is allowed.
An agent uses a model to interpret a goal, select tools, inspect their results and continue or escalate. A fixed workflow follows a predefined sequence. A generated summary alone is not an autonomous business process.
Use the simplest design that solves the task. A predictable approval or calculation often belongs in Flow or Apex; an agent is useful when the next step depends on the information it discovers.
Anthropic: agent and workflow patterns ↗
Your AI experience, inside Salesforce
Ask in one place.
Act within your business rules.
Give employees or product customers an AI workspace in Salesforce. It can bring together an initial briefing, conversational questions and approved actions, using Claude, OpenAI or another supported model behind the experience.
01 / THE EXPERIENCEStart with the work, not a blank chat.
Surface assigned tasks, exceptions and decisions from authorised data. Let users ask follow-up questions or request a next step from a Lightning page or your product interface.
Examples: What changed? Why is this blocked? Which items need my approval?
02 / THE INTELLIGENCEChoose the model behind the experience.
Use supported Salesforce model services or a controlled provider integration to connect Claude, OpenAI or other models. Define which context leaves the org, what is retained and which tasks each model handles.
Buyer benefit: the user stays in Salesforce; model selection becomes an architecture choice.
Salesforce: model integration capabilities ↗03 / THE CONTROLKeep the authority in the product.
Validate every proposed operation through application services, Flow or Apex. Keep credentials server-side and expose only the operations the user may perform.
MCP when useful: use it for compatible tool connections. You do not need to expose your entire product through an MCP server to embed AI.
Integration and tool design ↗ Salesforce can remain the place users work and business rules are enforced. Model processing and connected services may still run outside the org; the data boundary must be designed and verified.
A unified business platform
Bring more of the business together.
Reduce the hand-offs between systems.
Use shared records, custom objects and workflows to connect processes across departments on Salesforce. AI can work across that context instead of treating each task as a separate conversation.
Revenue & customers
Account research, proposals, opportunity follow-through and customer requests.
Finance operations
Invoice queries, collection follow-ups and exception explanations, with controlled access to financial systems.
Procurement & suppliers
Request triage, supplier evidence, purchase approvals and contract follow-up.
Delivery & assets
Project dependencies, order exceptions, maintenance requests and asset history.
People & internal operations
Employee requests, onboarding, policy retrieval and administrative coordination.
Your industry product
Workflows built around your domain objects, customer permissions and operating rules.
These patterns can be adapted across industries. Salesforce-native processes can share a common data model; ERP, banking, specialist applications and internet sources still require supported connections. Consolidation should reduce unnecessary integration, not pretend it disappears.
For Salesforce product companies
Add AI to your product.
Keep control of the customer experience.
Respect each customer’s boundaries.
Resolve the current org and user, apply package permissions, and retrieve only data relevant to the request. Keep customer configuration and entitlements explicit.
Build a stable action layer.
Wrap domain operations in validated interfaces. An AI request uses the same rules as the product, including approval, duplication and state-transition checks.
Ship and support it as a product.
Version prompts and action contracts, test upgrades against customer configurations, and monitor adoption, failed actions and usage costs.
Connected research / Account intelligence
“Research Acme.
Prepare me for the conversation.”
Combine the account record with authorised external research, then return a briefing inside Salesforce. A useful answer includes its sources, retrieval date and uncertainties.
- Resolve the company. Match the domain, legal entity and geography so similarly named businesses are not confused.
- Retrieve evidence. Use approved search or data providers for the company website, filings, announcements and market information.
- Separate facts from inference. Market share needs a defined market and period; valuation may be unavailable for a private company. Report unknowns instead of inventing figures.
- Identify relevant roles. Suggest stakeholder roles and sourced professional contacts where permitted, with an explanation of relevance.
- Propose a next step. Save an approved briefing or create a follow-up task. Apply configured approval before outreach or material record changes.
External text is untrusted evidence, never an instruction to the agent. Provider permissions, licensing, freshness and data handling are part of the implementation scope.
Reference architecture / One bounded transaction
One case. A defined set of actions.
A result you can inspect.
One implementation pattern: an Acme Case is blocked by missing evidence. The same action-contract approach can be adapted to orders, assets, projects or your product’s custom objects.
01 / CONTEXTRetrieve only the evidence needed.
Resolve Case.AccountId and ContactId, check the current Status, and retrieve authorised agreement metadata. Handle missing or ambiguous matches explicitly. Treat document text as evidence, not instructions.
02 / ACTION CONTRACTMake the operation narrow.
An example linkVerifiedAgreement action accepts a Case ID, document ID and request key. Server-side checks verify the relationship, access and current case state before writing.
03 / TRANSACTIONEnforce the sequence in code.
Keep mandatory checks and writes within a defined Flow or Apex operation. Reject conflicting changes; prevent duplicate updates on retries. A model’s proposed action is not authorisation.
04 / EXCEPTIONStop when the decision changes.
A revised commercial term becomes an approval task. No matching agreement becomes a work item for the case owner. A timeout returns pending or failed status, never an invented completion.
Acceptance evidenceCorrect record changedUnauthorised write rejectedRetry causes no duplicateException reaches its owner
Controls designed into the implementation
Reduce exposure.
Detect failure. Recover deliberately.
No LLM implementation is risk-free or guaranteed error-free. The architecture must limit what a failure can affect and make incorrect behaviour observable.
Identity and Salesforce access
Define the execution identity for each channel. Enforce record sharing, object permissions and field access in custom actions. with sharing alone does not enforce CRUD or field-level security; choose appropriate user-mode operations or explicit checks.
Salesforce: secure Apex ↗
Grounding and data handling
Retrieve relevant authorised records and approved knowledge, with a clear response when evidence is unavailable. Check the actual data path and provider terms. Salesforce documents that Trust Layer capabilities vary; its guidance specifically notes that LLM data masking is disabled for agents. Custom external integrations need their own data controls.
Salesforce: Trust Layer scope and limitations ↗
Actions, approvals and recovery
Use narrowly scoped tools, validated inputs, bounded retries and request identifiers. Put financial thresholds and mandatory approvals in enforceable logic. Define compensation for cross-system partial failures; not every external action can be rolled back. Include a way to disable writes without losing the investigation trail.
Anthropic: containing agent access ↗
Evaluation before release and after changes
Run sandbox scenarios for ordinary requests, missing records, denied access, malicious document instructions and failed integrations. Use Agentforce Testing Center, DX or Testing API where applicable, alongside action and integration tests. Salesforce notes that agent tests can modify data and consume credits.
Salesforce: agent testing options ↗
Faster delivery, measured properly
Automate repeatable engineering.
Keep release decisions evidence-based.
Our delivery approach applies AI to requirements drafts, Apex and LWC scaffolding, test scenarios and documentation. These outputs enter the same review, permission checks and deployment gates as other changes.
Track togetherLead timeReworkEscaped defectsCost per successful task
Begin with a read-only pilot, then approved writes, then bounded automation when agreed evaluation criteria are met. Expand permissions only as the evidence supports it.
MCP development for your business
No MCP connection?
We can build it around your workflows.
Connect Salesforce, your business systems and an LLM through purpose-built tools. We map your processes and expose the context and permitted actions an assistant needs to reduce repetitive work.
MCP is the connection layer. The model interprets the request; your business logic validates and executes the action. We can configure Salesforce-hosted tools or build a custom server where your systems require one.
Research and enrich a lead.
Combine CRM context with approved search and company-data sources. Return a dated briefing, suggest relevant stakeholder roles and prepare record updates for review.
Answer questions across systems.
Retrieve an order status, invoice or support history through scoped tools, and bring the answer back into the Salesforce experience.
Automate the follow-through.
Turn a validated request into a task, an approval or an update. Apply identity checks, duplicate protection, error handling and an activity trail.
Company pages and permitted search providers can inform contact research. LinkedIn access requires an approved route; the design does not assume unrestricted scraping or access to private profiles.
Discuss your MCP workflow ↗Salesforce: custom MCP servers ↗Platform direction
Salesforce. Claudeforce. AIforce.
AIforce brings Salesforce capabilities into different AI experiences. Claudeforce connects Salesforce with Claude; Salesforce also continues to describe Agentforce capabilities, including an experience inside Lightning. We select the path around where your users work and the control your product needs.
Explore Salesforce’s current AI experiences ↗What a scoped engagement produces
An architecture you can review.
A workflow you can test.
Agree the system and data-flow design, action contracts, permission model, test scenarios, acceptance measures and operational handover. Start with one workflow before committing to broader automation.
Discuss the workflow and its constraints ↗