Short answer
AI connects to Salesforce through Agentforce, Salesforce’s own agents billed by Flex Credits or per user, through the REST and GraphQL APIs, or through Salesforce Hosted MCP Servers, generally available since the Summer ‘26 release. For New Zealand sales teams a good first project is an opportunity brief, read-only and built under the rep’s own permissions, before each customer meeting.
- Agentforce Flex Credits cost US$500 per 100,000; a standard action uses 20, about US$0.10.
- Hosted MCP servers let Claude, ChatGPT and custom agents reach Salesforce with standard OAuth.
- Custom MCP servers respect your org’s sharing and security model.
- SOAP login() for API versions 31.0 to 64.0 retires in Summer ‘27.
What Salesforce’s built-in AI does today
Salesforce’s AI is Agentforce, sold three ways: Flex Credits (consumption), Conversations (per conversation, price on request) and per-user licences (Agentforce pricing, checked 19 October 2026). Flex Credits cost US$500 per 100,000; a standard Agentforce action uses 20 credits and a voice action 30. The Sales, Service and Field Service add-ons cost US$125 per user per month for unmetered employee agent use, and the US$5 Agentforce User Licence still needs Flex Credits. Sales Cloud itself is now sold as Starter Suite, Pro Suite, Core, Advanced and Max (Sales pricing). There is no NZD option; NZ customers pay in USD or AUD.
Integration routes and when to use each
| Route | What it gives you | Use it when |
|---|---|---|
| Agentforce | Salesforce-built agents and actions inside the org | Work that stays in Salesforce and fits its actions |
| REST and GraphQL APIs | Records, SOQL and history under your own code | Briefs that combine Salesforce with other systems |
| Change Data Capture or record-triggered flows | Events when opportunities or activities change | Preparing a brief when a meeting is booked |
| Hosted Standard MCP Servers | SObject reads and writes, SOQL, search, Data 360, Tableau | Staff using Claude, ChatGPT or Cursor against the org |
| Custom hosted MCP servers | Only the Apex actions, flows or named queries you expose | Narrow, governed tools for agents |
| n8n, Make or Zapier | Salesforce triggers plus AI steps | Light automations outside the core sales process |
Salesforce’s Summer ‘26 developer guide says custom MCP servers “respect the full sharing and security model” and can be built from Apex actions, flows, the Named Query API, Prompt Builder prompts and Agentforce agents. A named query that returns exactly the brief’s fields is usually better than giving an agent general SOQL. Our guide to MCP for business systems explains why narrow tools are safer.
Automation 1: the opportunity brief
- Trigger
- A meeting is booked against an opportunity, or the rep clicks “Prepare brief”.
- Inputs
- One read-only named query returning the fields in the table below, run as the rep.
- Model step
- A one-page brief: where the deal stands, what changed since the last meeting, open risks and questions to ask. Each statement cites the record it came from.
- Checks in code
- Distinguish a field the rep cannot see from a field that is empty; never let the model treat hidden as blank.
- Human review
- The rep reads it. Nothing goes to the customer.
- Write-back
- Optional completed Task noting that a brief was generated and by which prompt version.
- Measures
- Rep usefulness rating, preparation time, briefs with errors reported.
| Brief section | Salesforce source | Note |
|---|---|---|
| Deal status | Opportunity.StageName, Amount, CloseDate, NextStep | Multi-currency orgs: show CurrencyIsoCode |
| What changed | OpportunityHistory | Stage, amount and close date moves since the last meeting |
| Who is involved | OpportunityContactRole, Contact.Title | Roles as recorded, never guessed |
| Recent contact | Task and Event, last 90 days | Show the date of the last logged activity |
| Service issues | Open Cases on the Account | Flags an unhappy customer before a renewal |
This mirrors our research briefing preparation workflow, and our lesson on using AI for meeting preparation covers what reps actually need in a brief.
Automation 2: after-call updates, approved by the rep
- Capture. The rep pastes call notes or a transcript into a screen flow.
- Propose. The model suggests a new next step, close date, stage and follow-up task, each with the sentence from the notes that supports it.
- Approve. The rep accepts, edits or rejects each change. Stage moves past your commit stage, or amount changes over a threshold, also need a manager.
- Write. Approved changes are saved under the rep’s identity, so field history records who changed what.
Permissions, scopes and approval design
Run briefs under the rep’s own OAuth identity so they contain only what the rep can see. Background jobs should use a dedicated integration user with a permission set limited to the objects and fields in your query, never View All Data or Modify All Data. The Summer ‘26 guide also warns that SOAP login() in API versions 31.0 to 64.0 retires in Summer ‘27: integrations that log in with a username and password must move to OAuth before then. Our agent permissions and approvals guide covers approval tiers.
Costs and limits
Enterprise Edition orgs get 100,000 API calls per 24 hours plus 1,000 per Salesforce licence, and production orgs allow 25 concurrent long-running requests (API limits). Professional Edition needs API access enabled.
Illustrative example: 50 briefs a week through Agentforce, at three actions each, is 60 Flex Credits or US$0.30 a brief, about US$65 a month. A custom brief reading 15,000 tokens and writing 1,200 on Claude Sonnet 5.5 (US$2 and US$10 per million tokens) costs about US$0.04, roughly US$9 a month for the same volume, plus build and support. All figures are list prices on 19 October 2026.
Where it goes wrong
- Hidden fields read as empty. Field-level security can make a model say “no amount set”. Pass visibility explicitly.
- Missing activity. If emails are not logged to Salesforce, the brief will miss them. Show the last logged activity date so reps know.
- Shared API allocation. Marketing, ERP and AI integrations draw on the same daily allocation. Monitor it.
- Edition uncertainty. Salesforce’s API table still uses legacy edition names, and the edition needed for hosted MCP was not confirmed officially when we checked.
- Manual recovery. Enable field history tracking on the opportunity fields AI can propose changes to, and keep a log of approved changes so a bad batch can be reversed by hand.
Smaller team? Compare AI and HubSpot. Salesforce also owns Slack; see AI and Slack. We design and build these as part of our AI integration services.
Sources
- Agentforce pricing, Salesforce, accessed 19 October 2026
- Sales Cloud pricing, Salesforce, accessed 19 October 2026
- The Salesforce Developer’s Guide to the Summer '26 Release, Salesforce Developers Blog, June 2026
- API request limits and allocations, Salesforce Developers, accessed 19 October 2026
- Claude API pricing, Anthropic, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.