AI tool decision matrix: what to use when in Salesforce

AI in a Salesforce programme follows one baseline rule: the agent works in a sandbox or scratch org, gets data read-only through Hosted MCP, and releases go only through the pipeline with review. Which tool your team or partner uses inside that rule is usually decided by customer contracts, data residency and the existing toolchain, and less often by model quality. For agents in production, ERP access and cost per action come on top. The matrix below turns this into twelve scenarios that actually come up at manufacturers, from workshop prototype to order status from SAP.

Which question decides the tool choice?

The first question is whether you are talking about building or running. That sounds obvious, but in workshops both topics often end up on the same slide.

• Building: a coding agent such as Agentforce Vibes, Claude Code, Codex or Copilot writes Apex, LWC, flows and metadata. Billing is by tokens or per user. The output is code that has to pass review and the pipeline.
• Running: an Agentforce agent answers dealer enquiries, pre-checks warranty claims or reads order status from the ERP. Billing is per action in Flex Credits. The output is an answer to a customer or dealer.

Each world has its own criteria, its own budget and its own tests. The terms behind them are explained in chapter 1, Salesforce’s own tools in chapter 2 and the external coding agents in chapter 4.

The decision matrix for manufacturers

Status as of 27 October 2026, based on the linked pages. Prices are US list prices.

Building: coding agents, data and pipeline

#	Scenario	Tool or approach	Why	Cost and governance
1	Clickable prototype for a workshop, such as a dealer portal in LWC	Vibe coding with Agentforce Vibes in a Developer Edition org or Claude Code against a scratch org	Throwaway code, speed matters, no customer data; Martin Fowler places vibe coding in software for a small circle (opens in a new tab)	Developer Edition: 110 requests or 1.5M tokens (opens in a new tab), one month, no refresh; discard the code afterwards
2	Project features, Flex Credits already in the Salesforce contract	Agentforce Vibes 2.0 (opens in a new tab) with rules files and Plan Mode	Runs under the Salesforce contract, billed in Flex Credits	Flat Fee $125 per user per month (opens in a new tab), premium model capped at 2,500 requests or 36M tokens (opens in a new tab) per 30 days
3	Large refactorings, features across many metadata types, architecture reviews	Claude Code with the salesforce-development plugin (opens in a new tab)	About 40 Salesforce skills, Apex language server, deploy validation hooks, an architecture review agent	Anthropic cites around $13 per developer per active day (opens in a new tab), $150 to $250 per month; settle the data contract
4	Customer runs GitHub Enterprise with Copilot	Copilot, or Claude and Codex through Agent HQ (opens in a new tab), plus the open Salesforce skills library	Admins control permitted agents and models, audit logs exist	Enterprise $39 per user including $39 of AI Credits (opens in a new tab); Claude and Codex in Agent HQ launched as public preview in February 2026
5	Org sits in the Hyperforce EU Operating Zone, or EU processing is mandatory	Claude Code through Amazon Bedrock in an EU region, or OpenAI API projects with EU residency	Vibes is not available in the EU Operating Zone (opens in a new tab); the direct Claude API offers only “us” and “global” (opens in a new tab)	Data processing agreement, works council, no production data in prompts
6	Agent needs data, such as checking the installed base before a migration	Hosted MCP with the sobject-reads server against a sandbox	Named user, PKCE, field security and sharing apply (opens in a new tab), read-only	Docs say only for customers with Flex Credits (opens in a new tab); the meter for agentic access is still “TBA”
7	Code quality, tests, deploy validation	sf CLI with Code Analyzer and Apex tests as a hook and in the pipeline	Deterministic and free of token cost; Salesforce itself recommends the CLI over MCP (opens in a new tab) when no LLM is needed	Free; the only reliable check that something is done
8	Existing setup with the local DX MCP Server @salesforce/mcp	Pin version 0.30.15, move to skills, plugin and Hosted MCP	End of life on 2 November 2026 (opens in a new tab)	Restrict access to fixed sandbox aliases, never to all orgs

Running: agents in service, sales and ERP

#	Scenario	Tool or approach	Why	Cost and governance
9	Agent that pre-checks warranty claims from dealers	Agentforce Builder with Agent Script, tests before go-live	Salesforce’s rule: what can be expressed as code belongs in logic (opens in a new tab)	20 Flex Credits per action in production, 16 in a sandbox (opens in a new tab)
10	Agent returns order status or availability from SAP S/4	MuleSoft API, synced to API Catalog (opens in a new tab), used as an agent action	Documented pattern, named credentials are created automatically	MuleSoft licence; integration user with minimal permissions
11	Old ERP or MES without an API	MuleSoft IDP with RPA, or the Systems Integration Agent	IDP routes documents under 80 percent confidence to a human (opens in a new tab); the Systems Integration Agent has been GA since Dreamforce 2026 (opens in a new tab)	RPA and the Systems Integration Agent work through the user interface; plan monitoring
12	Sales wants to use Salesforce directly in Claude	Salesforce in Claude from the AIforce layer	Available to all customers in beta (opens in a new tab)	“Pricing and packaging are subject to change”; pilot with a small group

Six questions in the right order

The matrix is for looking things up. In a conversation with the IT lead and the programme manager I prefer to walk through these six questions in order, because early answers remove later options.

1. 01

   Prototype or production?

   Vibe coding in a Developer Edition or scratch org is fine for a prototype. Production needs agentic engineering: a human reads the code, and tests and the pipeline decide on the merge. Fowler keeps the two terms apart for exactly this reason.

2. 02

   Which data and residency rules apply?

   If the org sits in the EU Operating Zone, Agentforce Vibes is out. If data protection requires EU processing, Anthropic's direct Claude API is out, while Claude through Bedrock in an EU region remains. For Codex I found no statement from OpenAI that the EU residency for API projects (opens in a new tab) applies there too; I would get that in writing.

3. 03

   Which contracts and tools already exist?

   If Flex Credits are already in the Salesforce contract, Vibes is the natural choice. If IT runs GitHub Enterprise, Copilot is. With a free choice I use Claude Code as the reference, because Salesforce built its first official plugin for it. The full comparison is in Agentforce Vibes, Claude Code or Copilot.

4. 04

   What is the job?

   Code goes to the coding agent, metadata knowledge to the API context server, data to Hosted MCP, business rules of a runtime agent to Agent Script, releases to the pipeline.

5. 05

   Which cost model fits?

   Heavy, steady use favours flat per-user fees; light or fluctuating use favours pay-per-use. The numbers are in chapter 7.

6. 06

   Which checks are deterministic?

   Code Analyzer, Apex tests, a validation deploy and human review come before every merge, whichever agent wrote the code. What those checks look like is covered in chapter 8.

For agents in production a seventh question comes before all others: is the process worth it at all? I go through volume, data quality, ERP access, cost model and edition in Agentforce for manufacturers: start now or wait?.

Example: three initiatives at a machine builder

Assume a machine builder with Sales Cloud and Service Cloud in an org that sits in the Hyperforce EU Operating Zone. Headquarters runs SAP S/4, while an acquired plant still uses an older ERP without an interface. IT runs GitHub Enterprise, and an implementation partner supplies six developers. All circumstances and numbers are assumptions. Three initiatives are on the table.

A prototype for the dealer portal. Sales wants to show in a workshop how dealers could order spare parts. That is row 1. A separate Developer Edition org holds no customer data, so the partner can build there with Vibes or Claude Code, as long as data protection signs off on the test data. After the workshop the code is deleted, not reused.

The partner’s coding agents. Because the org sits in the EU Operating Zone, Vibes is out for project work (row 5). That leaves Copilot Enterprise on the existing GitHub licence (row 4) or Claude Code through Bedrock in an EU region (rows 3 and 5). With Bedrock, AWS bills the usage, so its price list belongs in the comparison, not Anthropic’s. How I set up an approved-tools list, a cost estimate and a comparison on real backlog tasks is shown in the scenario in chapter 4. The works council and the data processing agreement come before the decision, as described in AI coding tools, works council and GDPR. Whatever the tool, agents work only in sandboxes, and the pipeline checks with Code Analyzer and Apex tests (row 7).

An agent for order status. Dealers ask about delivery dates every day. A MuleSoft API to SAP S/4 exists, which is row 10. The agent calls it as an action, and the decision about which data a dealer may see sits in Agent Script. The details are in Agentforce and SAP order status. For the acquired plant without an API (row 11) I would not roll out the agent at first. RPA breaks when screens change, and I would test whether the Systems Integration Agent really copes with that on a copy of the system first. The options are in legacy ERP without an API.

Where does the matrix not help?

Three questions no table answers.

Question	Why the matrix does not answer it
Model quality for Apex	There is no public benchmark for Apex, flows and metadata, as described in chapter 4. Only a comparison on your own backlog can decide, measured by review findings and rework.
Ownership	Who maintains the rules file in the repository, who approves Hosted MCP access and who pays for Flex Credits is a contract question between you and your partner. I covered that in chapter 5.
Whether Salesforce is the right place	Decisions SAP already makes do not belong in an agent in Salesforce. For routine work without language, a flow or an ERP job is often cheaper than any agent.

Sorting out these decisions for your org and your partners is part of my work in Salesforce architecture. Where agents carry weight in manufacturing today is the subject of chapter 10; all chapters are listed on the masterclass overview.