AI in Salesforce Records for Financial Services

You can put AI in Salesforce records without buying an Agentforce add-on in five ways: use the AI already in your edition (including the free Salesforce Foundations bundle, which turns on Prompt Builder with a limited pool of Flex Credits), bring your own model contract into Salesforce, call Claude or OpenAI from Apex or Flow through Named Credentials, connect Claude to Salesforce's hosted MCP servers, or install an AppExchange app. For a regulated firm, pick the pattern that gives you masking, human review and a retained log with the least custom work. Agentforce is still the better buy for customer-facing agents and multi-step autonomous actions.

This guide is for Salesforce owners at wealth managers, banks, credit unions and insurers who need to explain the cost and the controls to a CFO and a CCO. Salesforce changes its packaging often, so licensing statements below reflect what Salesforce published in October 2026. Check your own contract.

What AI is already included in Salesforce without Agentforce?

Salesforce Foundations. Salesforce describes Foundations as included at no extra cost with Enterprise, Unlimited and the older Einstein 1 editions. The Agentforce pricing page lists it at $0 with Prompt Builder and Agentforce Builder, and the Sales Cloud pricing page notes that "limited Flex Credits and Data 360 credits" come with it. Prompt Builder is the native way to put a generated summary on a record page, and Salesforce's setup documentation says Data 360 (formerly Data Cloud) is required for generative AI functions such as the Einstein Trust Layer. Once the starter credits run out, usage is metered in Flex Credits at $500 per 100,000.

Edition-level AI. In 2026 Salesforce reorganized Sales Cloud into Starter, Pro, Core ($195 per user per month), Advanced ($395) and Max ($550). Salesforce's FAQ says Core includes opportunity scoring and conversation intelligence, and the comparison table lists annual Flex Credit allotments for Core, Advanced and Max. Lead scoring and call summaries show as "available for purchase" on the lower tiers. Many financial firms are still on older Enterprise or Unlimited contracts, so ask your account executive what carried over.

Slack AI. If your Salesforce plan includes Slack Business+ (Core does), Slack's help center lists AI search, recaps and file summaries on Business+, plus conversation summaries on every paid plan. Salesforce channels and record previews work across plans. Those summaries live in Slack, not on the record.

Einstein Activity Capture. Before a broker-dealer or RIA relies on it, confirm with compliance where captured emails are stored, how long they're kept and whether they reach your archive. The "Sync Email as Salesforce Activity" option added in 2025 writes emails as native records, which changes the answer.

What can AI actually do on a financial services record?

These are the jobs we see most often in Financial Services Cloud (FSC) and Sales or Service Cloud orgs:

• Summarize a household or account record. Pull members, open opportunities, recent tasks and cases into a brief stored in a field.
• Draft a client follow-up from activity history. Draft an email from the last meeting note and open tasks for the advisor to edit and send.
• Classify inbound cases and complaints. Tag type and urgency, and flag language that may make it a written customer complaint for supervisor review.
• Extract data from uploaded documents. Propose field values from a trust document, statement or declaration page; a person confirms before anything is written.
• Flag missing KYC fields. Use validation rules or Flow for the structured check, since that's auditable. Use AI for the unstructured part, such as a note mentioning a new employer that isn't on the record.
• Meeting prep briefs. Household summary, service history and open items before a review.
• Next-best-action suggestions. A suggested follow-up (a beneficiary review after a life event, say) saved as a task for the advisor, never sent to the client.

Which non-Agentforce pattern should you use?

Prompt Builder through Foundations

An admin writes a template grounded in record fields, related lists and Flow outputs, and the output appears on the page or is written to a field by Flow. Because it runs through Salesforce's generative AI gateway, you get the Einstein Trust Layer, which Salesforce's documentation says includes grounding in CRM data, masking of sensitive data such as Social Security numbers, toxicity detection, an audit trail, and zero data retention agreements with third-party model providers. Salesforce's supported-models list says Anthropic and Amazon models run on Bedrock "entirely within the Salesforce Trust Boundary."

Bring your own model in AI Models

Salesforce's documentation now calls this "AI Models" (older material says Einstein Studio or Model Builder). It supports Amazon Bedrock, Azure OpenAI, OpenAI and Google Vertex AI as providers. Claude is reached through Bedrock. It suits a bank with an approved Bedrock or Azure OpenAI tenancy.

Direct callouts from Apex or Flow

This is the most flexible pattern and needs no Salesforce AI licence. Store the endpoint in a Named Credential and the key in an External Credential, grant access through a permission set, and call the model from an invocable Apex method or from Flow's HTTP Callout action (which generates an External Service). Bedrock works too, since External Credentials support AWS Signature Version 4.

The trade-off is that you own every control the Trust Layer would have given you. Anthropic's API documentation describes zero data retention arrangements enabled per organization through its sales team, and lists features outside ZDR such as batch processing and the Files API. OpenAI's documentation says API data isn't used for training by default, that abuse monitoring logs are kept for up to 30 days, and that approved customers can be excluded from those logs. AWS says model providers on Bedrock don't have access to customer prompts and completions. Get your terms in writing.

Hosted MCP servers and Claude

Salesforce now runs hosted MCP servers, including SObject Reads (read and query only), SObject Mutations, SObject All, Data 360 and Tableau Next. They're disabled by default. Salesforce's reference says field-level security, object permissions and sharing rules apply to every tool call, with per-user authentication. An admin creates an External Client App, and each user connects Claude with their own Salesforce login.

It fits meeting prep and questions across many records. It doesn't write summaries into fields on its own, and the conversation lives in Claude, so check your Claude plan's retention settings and whether transcripts need archiving. Salesforce and Anthropic announced Claudeforce on August 26, 2026 (a "Salesforce in Claude" plugin with 37 prebuilt sales skills); pricing hasn't been announced.

AppExchange apps

Packaged apps add record-page AI with little build. Our own RecordIQ accelerator is one example: a Lightning Web Component that sends the record context you configure to Claude, OpenAI or other models using your own API key, logs each call to an object in your org, and doesn't require Agentforce, Einstein or Data 360 licences. Run any vendor through third-party risk review.

Example: summarize an FSC household with Claude, with human review

A minimal screen flow an advisor launches from the household record, using an Apex invocable action and no Salesforce AI licence.

1. Create the credentials. Create an External Credential with a custom header x-api-key whose value is the formula {!$Credential.Anthropic_Key.ApiKey}, and a Named Credential called Anthropic_API pointing to https://api.anthropic.com with "Allow formulas in HTTP header" turned on. Grant the principal only to an "AI Household Summary" permission set.
2. Add two fields and a log object. Add AI_Summary__c (long text) to Account, and create an AI_Interaction_Log__c object with fields for record ID, user, model, prompt, raw output, approved text, reviewer and request ID.
3. Deploy the Apex action below. It queries WITH USER_MODE so the advisor's field-level security and sharing apply, strips obvious identifiers, and returns the draft without saving it.
4. Build the screen flow. Call the action, show the draft in an editable long text component with an "I have reviewed this summary" checkbox, then update AI_Summary__c and create the log record only after the advisor approves.
5. Add the quick action to the household page layout and assign the permission set to a pilot group.
6. Test in a sandbox with missing data, conflicting notes and sensitive text, and file the results with your use-case approval.

Here is an illustrative version of the Apex action. Have a developer review and test it in a sandbox before use.

public with sharing class HouseholdSummaryAction {
    public class Request {
        @InvocableVariable(required=true) public Id householdId;
        @InvocableVariable(required=true) public String modelId;
    }
    public class Result {
        @InvocableVariable public String draftSummary;
        @InvocableVariable public String promptSent;
        @InvocableVariable public String requestId;
        @InvocableVariable public String errorMessage;
    }

    @InvocableMethod(label='Draft household summary with Claude' callout=true)
    public static List<Result> run(List<Request> requests) {
        List<Result> results = new List<Result>();
        for (Request req : requests) {
            results.add(summarize(req));
        }
        return results;
    }

    private static Result summarize(Request req) {
        Result res = new Result();
        res.requestId = System.Request.getCurrent().getRequestId();
        Account hh = [SELECT Name, Description FROM Account
                      WHERE Id = :req.householdId WITH USER_MODE LIMIT 1];
        String context = 'Household: ' + hh.Name + '\nNotes: ' + hh.Description + '\nRecent activity:\n';
        for (Task t : [SELECT Subject, ActivityDate, Description FROM Task
                       WHERE WhatId = :req.householdId WITH USER_MODE
                       ORDER BY ActivityDate DESC LIMIT 20]) {
            context += String.valueOf(t.ActivityDate) + ' ' + t.Subject + ': ' + t.Description + '\n';
        }
        context = context.replaceAll('\\b\\d{3}-\\d{2}-\\d{4}\\b', '[SSN removed]')
                         .replaceAll('\\b\\d{8,17}\\b', '[number removed]');
        res.promptSent = context;

        Map<String, Object> body = new Map<String, Object>{
            'model' => req.modelId,
            'max_tokens' => 800,
            'system' => 'Write an internal household summary for a financial advisor. '
                + 'Use only the data provided, say when information is missing, '
                + 'and do not make investment recommendations.',
            'messages' => new List<Object>{
                new Map<String, Object>{ 'role' => 'user', 'content' => context }
            }
        };
        HttpRequest httpReq = new HttpRequest();
        httpReq.setEndpoint('callout:Anthropic_API/v1/messages');
        httpReq.setMethod('POST');
        httpReq.setHeader('content-type', 'application/json');
        httpReq.setHeader('anthropic-version', '2023-06-01');
        httpReq.setTimeout(60000);
        httpReq.setBody(JSON.serialize(body));
        try {
            HttpResponse resp = new Http().send(httpReq);
            if (resp.getStatusCode() == 200) {
                Map<String, Object> parsed = (Map<String, Object>) JSON.deserializeUntyped(resp.getBody());
                List<Object> blocks = (List<Object>) parsed.get('content');
                res.draftSummary = (String) ((Map<String, Object>) blocks[0]).get('text');
            } else {
                res.errorMessage = 'Model API returned status ' + resp.getStatusCode();
            }
        } catch (CalloutException e) {
            res.errorMessage = e.getMessage();
        }
        return res;
    }
}

Keep the model ID in custom metadata and pass it in from the flow, so a model change goes through your normal change control rather than a code deploy. Add FSC financial accounts using the object name your org uses (managed-package and standard-object versions differ). The request ID lets you join this log to Salesforce event logs, the correlation approach Salesforce's developer blog recommends for cross-system tracing.

Which compliance controls do financial firms need?

Send the minimum. Free-text notes are where account numbers and health details hide. Prompt Builder gets Trust Layer masking; direct callouts need your own.

Respect field-level security and sharing. Run Apex with sharing and query WITH USER_MODE. Hosted MCP servers enforce the user's permissions on every call, which will expose any over-broad permission sets.

Human review before anything reaches a client. FINRA's Regulatory Notice 24-09 (June 2024) says its rules are technology neutral, and that the communications standards in Rule 2210 apply whether a person or a tool wrote the content. Write drafts to a draft field and keep the send step with the advisor.

Log prompts and outputs. FINRA Rule 4511 requires records to be kept in a format that complies with SEC Rule 17a-4, for at least six years where no other period applies. Whether a prompt or a raw draft is a required record is your compliance team's call, but the sent communication is, and keeping the prompt, raw output and approved version lets you show what the advisor changed. A log object in Salesforce isn't an archive, so plan the export to your 17a-4 storage.

Govern models even though the bank guidance changed. On April 17, 2026, the Federal Reserve, OCC and FDIC replaced SR 11-7 and OCC Bulletin 2011-12 with revised model risk guidance (SR 26-2). It says generative and agentic AI are "not within the scope of this guidance" and that a bank's own risk management and governance practices should set the controls. Keep a use-case inventory with a risk rating, documented testing, monitoring and a named owner for each.

Change control. Treat a prompt edit like a code change: logged, tested in a sandbox and approved by compliance when it touches client communications.

When is Agentforce the better choice?

Buy Agentforce when the AI must act on its own across several steps (open a case, update the record, schedule a follow-up), when it faces clients through chat or voice, or when you want Salesforce to own the Trust Layer, model hosting and support rather than your developers. A custom Apex integration is code you maintain through three Salesforce releases a year. Many firms prove use cases and controls with the patterns above, then move high-volume ones to Agentforce.

Buyer checklist for financial services teams

• Which Salesforce contract are we on, and does it include Foundations, Data 360 and any Flex Credits?
• Which record-level use case comes first, and who signs off on the output?
• Which fields are allowed in a prompt, and who approved that list?
• What retention and training terms apply with our model provider, and are they in our contract?
• Where are prompts, outputs and approved versions stored, for how long, and do they reach our archive?
• How will we test accuracy before launch and monitor it afterwards?
• Who reviews permission sets before an AI tool can read across records?
• Who maintains custom Apex when Salesforce or the model provider changes an API?

FAQ

Can we use AI in Salesforce without buying Agentforce?

Yes. Enterprise and Unlimited customers get Salesforce Foundations at no extra cost, which includes Prompt Builder and a limited pool of Flex Credits. You can also call Claude or OpenAI from Apex or Flow with Named Credentials, connect Claude through Salesforce's hosted MCP servers, or install an AppExchange app, none of which need an Agentforce licence.

Does the Einstein Trust Layer protect direct API callouts?

No. The Trust Layer applies to requests that go through Salesforce's generative AI gateway, such as Prompt Builder. If Apex calls a model provider directly, your team builds the masking, logging and review steps and relies on the provider's contract terms for retention.

Is it safe to send client data to Claude from Salesforce?

It can be, with the right terms and design. Use Claude through Salesforce's Bedrock-hosted models, your own Bedrock account, or the Anthropic API under a zero data retention arrangement, send only the fields the task needs, and have compliance approve the use case before go-live.

Do AI-drafted client emails need to be archived?

Any email actually sent to a client is a business communication and falls under your existing retention and supervision rules, whoever drafted it. Many firms also keep the prompt and raw output to show what the advisor changed.

How much does it cost to add AI to Salesforce records?

It ranges from nothing beyond your edition (Foundations and included features) to Flex Credits at $500 per 100,000, to model API usage for direct callouts. The Agentforce add-ons start at $125 per user per month, or $150 for Industries clouds such as FSC.

Does SR 11-7 still apply to generative AI in Salesforce?

SR 11-7 was replaced on April 17, 2026 by SR 26-2, which explicitly excludes generative and agentic AI from its scope. The agencies said firms should apply their own risk management and governance practices to those tools, so an AI governance policy and use-case inventory are still expected.

Get help choosing the right pattern

Vantage Point is a Salesforce partner, a Claude (Anthropic) Partner Network partner and an OpenAI Partner Network partner, working with financial services firms since 2018 across 700+ engagements. We configure the controls and your compliance team approves them.

• See our Claude services and Agentforce services.
• Compare scoped AI packages or start with an AI readiness assessment.
• Read how Claude inside the Salesforce trust boundary changes vendor review, or our 2026 guide to Agentforce and Einstein.
• Connect Claude to Salesforce with our free Claude MCP accelerator, or try RecordIQ on your record pages.

Not sure where to start? Book a free CRM assessment and we'll map your first AI use case to the pattern, licence and controls that fit.