GPT will impress anyone in a sandbox. Putting it in front of customers or employees, where it has to be reliable, affordable, and safe on inputs nobody anticipated, is a software engineering problem that happens to involve a model. The teams that stall keep reaching for a better prompt when the thing blocking them is evaluation, observability, data quality, or a plan for when the model gets something wrong. That gap is where our OpenAI work lives.
We build production systems on the GPT family: direct API integration, Azure OpenAI for regulated and enterprise deployments, the Responses and Realtime APIs, and fine-tuning in the rare cases it is actually justified. Around the model we build the parts that decide success, evaluation so you know whether output is right, observability so you can see what is happening in production, and cost management so the bill does not quietly outrun the value.
The predictable failures: a model grounded in nothing, so it invents answers; no evaluation, so quality is a matter of opinion; runaway token costs no one is watching; and a pilot that works on the happy path and falls apart on the inputs real users send. We design for grounding, measurement, and cost from the start, because those are the parts that separate a feature that ships from a demo that does not.
We build production AI for a living, with the engineering discipline that turns a capable model into a dependable system. As a Salesforce consulting, implementation and integration partner, we integrate GPT where it does real work inside your stack, and we are honest about when a different model, or no model, is the better answer.
The engagements we run most often on OpenAI / GPT, from first implementation through optimization.
Production applications built on current GPT and reasoning models: research, drafting, summarization, classification, and orchestration.
Enterprise-grade Azure OpenAI deployments with VPC, data residency, content filtering, and audit logging for regulated workflows.
Agentic applications on the Responses API with built-in tools, MCP and persistent conversations, realtime voice with the Realtime API, and migration of legacy Assistants API builds, which OpenAI retired in August 2026.
OpenAI embeddings for semantic search, classification, recommendation, and RAG retrieval pipelines.
Custom fine-tunes for narrow workflows where prompt engineering hits its limit and accuracy or cost demands a tuned model.
Token tracking, request observability, evaluation frameworks, and cost optimization at production scale.
Engagements are measured by movement on the numbers that matter. These are the directions of travel we commit to.
Predictable phases. Clear deliverables. No surprises.
One to two working sessions to map your current state, business goals, and gaps. We come out with a written scope and recommendation.
Documented architecture, realistic timeline, and transparent commercial proposal. No surprises and no hidden scope.
Configuration, development, integrations, data migration, and QA, with weekly demos and on-the-fly adjustments.
Training, change management, hypercare, and ongoing optimization. We do not disappear at go-live.
Consultant-level analysis from the consultants delivering the work.
How Salesforce data and actions reach ChatGPT, Gemini Enterprise and Claude today, what is in beta, how permissions carry over, and how to decide which assistant each team should use with your CRM.
Read the full post →OpenAI’s Responses API lets an application send a prompt plus a list of functions the model may request. Examples include looking up an opportunity or logging a call. The model returns a structured request, and your code decides whether to run it against Salesforce. Nothing touches the org unless your integration executes that call under an identity you selected.
ChatGPT Business and Enterprise workspaces can add custom apps that point at an MCP server, including Salesforce’s hosted MCP servers, which authenticate each person through OAuth. Salesforce has also announced Agentforce experiences inside ChatGPT. Workspace admins control which apps appear, so enabling a Salesforce connection should be a governed decision. Confirm current availability for your plan and org.
Agentforce includes OpenAI models among its Salesforce-managed model choices, and Salesforce’s October 2025 announcement highlighted GPT-5 for building agents and prompts. Requests pass through the Einstein Trust Layer for masking, grounding and audit logging. Selecting the model is a configuration inside Salesforce, so the agent’s subagents, actions and permissions stay where your admins already manage them.
Salesforce’s bring-your-own-LLM feature, configured in AI Models (formerly Einstein Studio), can connect a model from your own OpenAI or Azure OpenAI account. Prompt Builder and Agentforce then use it while inference runs under your contract with the model provider. This suits companies with negotiated OpenAI terms or tuned models. Check which model versions and Trust Layer features Salesforce currently supports.
OpenAI embedding models convert text such as knowledge articles, case resolutions and product documentation into vectors for semantic search. A custom application retrieves the closest passages, then passes them to a GPT model along with the Salesforce record in question. Keep the index synchronized with record edits and deletions, or answers will cite articles your team already retired.
Strict JSON schema output lets GPT return fields like case reason, urgency and product line in a fixed shape your code can check against Salesforce picklists. OpenAI’s platform also offers evaluation tooling for scoring responses against graded datasets. Together they make it practical to measure classification accuracy before any generated value lands in a production field.
List candidate tasks and decide where each belongs: ChatGPT with a Salesforce app for employee questions, the API for custom software or Agentforce for agents inside Salesforce. Write down the objects and fields each task touches. That single map prevents three teams from wiring GPT to the same org in three incompatible ways.
Before production data flows, have security and legal review OpenAI’s current business terms, retention defaults, data residency options and whether zero data retention applies to your endpoints. If Azure OpenAI is the route, review Microsoft’s terms instead. Record the decisions, so later projects inherit an approved configuration rather than reopening the same debate.
Use per-user OAuth for ChatGPT connections so Salesforce sharing and field-level security decide what each person sees. For API integrations, create a dedicated integration user whose permission set covers only required objects and actions. Avoid shared admin credentials entirely, and set token and session policies on the external client app from day one.
Run the first release read-only and assemble a graded test set from real records and questions. Score correctness, refusal behavior and cost per task. Then add one write action at a time, such as logging a call, with a person confirming each change and validation logic checking values before the save happens.
Track token usage, latency, error rates and evaluation scores on a regular schedule, and alert when any of them moves sharply. When OpenAI releases a new model version, rerun your test set before switching. Review ChatGPT app usage alongside Salesforce login history, and remove access that nobody has used in months.
OpenAI sells ChatGPT Business and ChatGPT Enterprise as per-seat workspaces, while the API platform is billed on usage by model and token type. GPT models used through Agentforce are contracted with Salesforce and draw on Salesforce’s own usage entitlements, so ask your account executive how that consumption is metered. Bring-your-own-LLM connections bill inference to your OpenAI or Azure account instead. Salesforce hosted MCP servers depend on your edition, so confirm eligibility. OpenAI states it does not train on business or API data by default, and it offers data residency and zero data retention to qualifying customers. Review the current terms at openai.com before connecting Salesforce data.
When every team calls OpenAI with the same key, nobody can attribute cost, rotate credentials safely or cut off one misbehaving integration. Create separate projects and keys per application, store them in a secrets manager and set usage limits. On the Salesforce side, pair each application with its own integration user so audit trails stay readable.
GPT can return a plausible picklist value that does not exist, a date in the wrong format or a confident guess where the source was silent. Writing that straight into Salesforce corrupts reports quietly. Validate every value in code, reject anything unrecognized and route uncertain results to a review queue instead of saving them.
Newer GPT versions often change tone, length and edge-case behavior, even when overall quality improves. Pointing production at an alias that updates silently can break a parser or shift classifications overnight. Pin explicit model versions, rerun your graded test set against each new release and switch only when the scores hold.
Employees who paste account notes or customer emails into personal ChatGPT accounts move Salesforce data outside your business terms and admin controls. Offer an approved workspace with a governed Salesforce connection, publish a clear usage policy and make the sanctioned route easier than the workaround. Enforcement works better when the approved tool is genuinely useful.
If the agent serves customers or employees inside Salesforce channels, choosing a GPT model within Agentforce keeps subagents, actions, masking and audit in one managed place. Direct integration fits when the experience lives elsewhere, such as ChatGPT, a portal or a data pipeline. Many companies do both, keeping each use case on the surface its users already work in. The deciding question is where the conversation happens and who maintains it.
OpenAI states that, by default, it does not train on inputs or outputs from ChatGPT Business, ChatGPT Enterprise or its API platform. API data is normally retained for a limited period for abuse monitoring, and qualifying customers can request zero data retention for eligible endpoints. Consumer ChatGPT accounts follow different settings. Terms change, so confirm the current wording at openai.com with your legal team before connecting Salesforce.
It can, when connected through an MCP server that exposes write tools, such as Salesforce’s hosted MCP servers or a custom build. Whether it should depends on your controls. Start with a read-focused configuration, then enable specific write tools once testing shows reliable behavior. Require users to confirm each change, and keep Salesforce validation rules and permissions as the final gate on every update.
Azure OpenAI, now offered through Microsoft Foundry, runs OpenAI models inside your Azure subscription under Microsoft’s terms, with Azure networking, identity and regional deployment. Companies with Azure commitments or strict network requirements often prefer it. Salesforce’s bring-your-own-LLM supports Azure OpenAI as a provider, so Agentforce can use those deployments. Model availability can lag OpenAI’s direct API, so check the versions you need first.
Measure cost per completed task, not per request, and set budgets and alerts on each API project. Send short, focused context: the fields a task needs rather than whole records with every related list. Use smaller models for classification and larger ones only where reasoning demands it. Cache repeated prompts where the platform supports it, and compare spending with the value each use case delivers.
Gemini models on Google Vertex AI for enterprise AI deployments and Google Workspace integration.
Copilot Studio, Microsoft 365 Copilot, and Azure OpenAI for Microsoft-first organizations.
Enterprise API platform, for connecting Salesforce, ERP, data warehouses, and the rest of your stack.
Modern iPaaS for SaaS-heavy stacks, automations, recipes, and AI-augmented workflows.
Cloud data platform, for unified analytics, AI grounding, and data sharing across the enterprise.
Executive analytics, embedded dashboards, and CRM Analytics built on Tableau.
Tech Talk
A monthly brief for the people who own Salesforce, AI and revenue technology
What changed in Salesforce and AI this month, and what to do about it.
One email a month. Written by the consultants who deliver the work, not by a marketing team, for the leaders who make the technology decisions.
Consultant analysis, not vendor recaps. One click to leave.
