CRM & ERP integration
Put a review step between the OpenAI answer and the Salesforce action.
Nexus can select the Salesforce fields approved for a task, send that bounded context to OpenAI, validate the response and pause for a named reviewer before writing back. The execution record connects the input, prompt version, model output, human decision and resulting CRM change.
The problem
Why OpenAI + Salesforce needs a governed process
Sending a CRM record to a model is easy; deciding which fields should leave the environment and what the result may change is harder. An invented fact, inappropriate customer detail or unreviewed recommendation can become an operational mistake if model text is written directly into the record.
Responsibility
How the systems divide responsibility
External model-processing service
OpenAI
Processing of the submitted request under the customer's configured service and data-handling terms; no authority over CRM records.
Customer and commercial record
Salesforce
Accounts, contacts, opportunities, cases and the approved CRM fields or actions updated by the process.
Capabilities
What Nexus orchestrates
Pre-built action groups used in this pattern. Anything else runs through an API, protocol or database step, or Rules + Code, in the same workflow.
OpenAI
25 ACTIONS
- Chat Completions8
- Files4
- Uploads4
- Projects5
Salesforce
73 ACTIONS
- Cases12
- Accounts10
- Contacts10
- Opportunities10
Architecture
One governed run across both systems
How work starts
Salesforce (SOQL polling), webhook or schedule
A user requests a summary, a case is ready for triage, or an opportunity reaches a review stage that requires a draft assessment.
When a pre-built action does not cover it
Workflow patterns
What teams run across OpenAI + Salesforce
01
Case summary for agent review
Select the permitted case history, request a structured summary and present it to the support agent before saving the approved text.
02
Opportunity review brief
Assemble selected opportunity facts and ask the model to distinguish recorded information from missing context, then route the brief to the account owner.
03
Draft customer response
Generate a draft from approved case and reference context. A named person reviews it before the configured CRM action; generation alone does not send a message.
04
Controlled record classification
Validate the model's proposed category against the allowed values and route ambiguous or consequential classifications to the record owner.
Direction and ownership
Which way each change moves
Salesforce provides only the permitted context for a defined model task. OpenAI returns a draft, classification or recommendation. Nexus validates and routes that result under the agreed approval policy before an explicit Salesforce write; the model does not gain independent CRM access from this handoff.
OpenAI Salesforce
Validated and, where required, human-approved summaries, classifications or draft content for designated fields.
Salesforce OpenAI
Task-specific CRM fields and instructions selected under the customer's data-sharing rules.
Control
Approvals and exception handling
Human approval
Customer-facing content, consequential classifications and the designated CRM changes wait for a named reviewer. The reviewer can accept, edit or reject the result. Only the approved values reach the configured write step, with the model output retained separately from the final decision.
Exceptions
- Disallowed fields or missing data-sharing approval stop the model call before transmission.
- Invalid output, unsupported claims or an unrecognized category returns to review instead of being written directly to CRM.
- Model errors and uncertain Salesforce writes retain their separate outcomes; a failed write is not resolved by blindly generating another answer.
Security, audit and deployment
Every run leaves a reviewable record
- The permitted source fields selected for the model request, with sensitive values protected under the agreed logging policy.
- Prompt or instruction version, model identifier and recorded response.
- Reviewer identity, edits, decision and timestamp where required.
- Salesforce update request, response and final destination values under the configured audit permissions.
Credentials for OpenAI and Salesforce are held in the vault and used only by the steps that need them. Role-based permissions decide who can edit, approve and run the workflow. Nexus runs on premises, in private cloud or air-gapped, so the execution record stays inside your deployment boundary. Governance and audit
Talk to a solutions engineer
See this run against your OpenAI and Salesforce
Bring the process you run today. We map which system owns what, where approval belongs and the steps each side needs, against your own OpenAI and Salesforce, not a generic environment.
- The OpenAI + Salesforce workflow built and run end to end.
- Where a person is asked to approve, and what they see.
- The execution record an auditor would read afterwards.
- What it takes to run this inside your own deployment boundary.
Participating connectors
The connectors behind this pattern
Related
Continue from here
- Human in the loop: where approval belongs
Where a person belongs in an automated run, and where they do not.
- Governed AI automation
AI steps inside a process that records what it did.
- Enterprise integration
Connectors, interfaces and code in one platform.
- Integration explorer
Add more systems to this combination.
FAQ
OpenAI + Salesforce questions
- Does self-hosting Nexus keep OpenAI requests inside our network?
- No. This integration sends the selected data to the external OpenAI service. It requires approved connectivity and data-sharing rules and cannot operate as a fully air-gapped OpenAI workflow.
- Can we use a self-hosted model instead?
- Nexus also supports locally hosted AI models. That is a different model deployment choice; it should not be described as an OpenAI service integration.
- Will the model update records or message customers autonomously?
- Only explicitly configured workflow actions can do that. This design places validation and the required human approval before the CRM write or communication step.
- Is every model answer logged without restriction?
- Evidence collection follows the agreed access, redaction and retention policy. Auditability does not require exposing raw customer data to everyone who can inspect a run.
Integration review
Discuss this integration
Bring your OpenAI and Salesforce process. We will map ownership, approval and the steps each system needs.