AIforce, originally Salesforce Headless 360, exposes the Salesforce platform to AI agents through APIs, MCP tools and command-line calls. Agents can work without a screen, but the records they use don't get repaired along the way. In nearly every Salesforce org Growth Orbit audits, we encounter the same gap: agent access is ready well before data quality is.
At one of our clients, Salesforce contained about 113,000 account records for a market of about 8,500 companies. Comparing them with Dun & Bradstreet company data left 70% unconnected to any company we could identify. An AI agent connected through headless can now read every one of those records.
What is Salesforce Headless 360?
Salesforce rebuilt its platform as Headless 360 for direct use by AI agents. The April 15, 2026 announcement described it as "the capabilities your agents need most, exposed as an API, MCP tool, or CLI command," and introduced more than 60 new MCP tools and 30 coding skills. At September's Dreamforce, the name became AIforce. It uses a Headless Toolkit in which "every capability is accessible via API, MCP, or CLI."
The intent was clear in co-founder Parker Harris's question: "Why should you ever log into Salesforce again?"
What does "headless" mean for a CRM?
A CRM's "head" is its visible interface: dashboards, forms and record pages. Making it headless retains the data, rules and permissions while removing the need to use those screens. Agents can retrieve records, edit fields and execute workflows directly, without a person opening the application.
HubSpot is taking a similar path. Its remote MCP server became generally available on April 13, 2026, allowing CRM records to be written as well as read.
What changes in November?
Registration and call metering are planned for agents using Salesforce's APIs or MCP servers. A Salesforce help article from September 17 describes separate identities for agents and the option of permissions narrower than those granted to people. Successful calls will count as a "Headless Platform Interaction" consuming Flex Credits. The rate remains unpublished. Salesforce promises 30 days' notice before metering begins.
What do companies expect headless Salesforce to do?
Companies generally expect four capabilities. Each is real:
- Ask instead of look: a leader can question an AI assistant about pipeline rather than navigate a dashboard.
- Use any AI: Salesforce MCP can connect Claude, ChatGPT, Copilot or an agent you build.
- Let agents do the work: agents can edit records, capture activity, route leads and send outreach.
- Build faster: one customer Salesforce cites put production agents into use in 12 days.
Substantive controls accompany that access. Your permissions constrain agents, registration flags risky MCP servers, and tracing lets you see what an agent did.
Why won't Headless 360 fix bad CRM data?
Headless 360 removes a screen without repairing the data behind it. That screen was where people quietly supplied corrections: a blank field could prompt a manager to question the rep. An agent instead uses the entries and infers what's missing. An October 2026 Inc. article about AI and CRMs described the consequence: "the gaps that humans compensated for without thinking become the exact places agents fail."
The studies point in the same direction:
- Through 2026, Gartner expects organizations to abandon 60% of AI projects not supported by AI-ready data.
- Salesforce's State of Sales survey covered 4,050 sales professionals. Respondents using AI agents ranked manual errors and duplicates as their leading data problems, and 46% reported harm to sales from poor data quality. Read it as a candid vendor survey.
- LeanData's May 2026 LeanData survey covered 157 go-to-market practitioners at its own customers. Among them, 30% had discovered agent actions without an audit trail.
What does bad data look like to an AI agent?
An agent can turn poor records into a confident answer. When choosing target accounts, it selects from the accounts it sees. At another of our clients, an IT-services firm, Salesforce held only about a quarter of the 18,775 companies that fit its offer. Forecasting agents likewise use the stages sellers entered. At a third client of ours, a fintech, 267 of 397 demos became "qualified," but only 5% closed. Give an agent access to disagreeing systems, and its answer follows whichever it reads first.
| What the agent reads | What it will conclude | What's actually true |
|---|---|---|
| Your CRM's account list | This is the set of best targets | Much of the market may be absent |
| Stages entered by sellers | The pipeline is substantiated | Holding a meeting doesn't establish buyer progress |
| Unmatched accounts and duplicate records | The account count is 13 times the real number | Most entries don't match any company you can identify |
| Conflicting records across systems | The first figure encountered | The systems have not been reconciled |
For an agent to distinguish good records from bad ones, the records must expose that distinction.
Does metering make bad data more expensive?
Probably, although the size of the increase can't yet be stated. Metered calls would mean an agent reading 113,000 accounts to answer a question about 8,500 companies also consumes interactions on irrelevant records, eventually incurring charges. Without a published rate, this supports cleaning the records before metering begins. It doesn't support a cost forecast.
What should you fix before plugging an AI into Salesforce?
This is the sequence Growth Orbit recommends:
- Match your CRM to your real market. Establish the match rate before starting any agent.
- Inventory every agent and tool. Build the list through Salesforce's registration and restrict each agent to the minimum permissions it needs.
- Define the numbers agents will answer. Distinguish engagement from buyer progress and verified outcomes.
- Make unknown mean unknown. Require meaningful entries in the fields agents use and prevent the duplicates you can.
- Read first, write later. Have people approve proposed edits until the agent's history supports independent writes.
- Log every write and give exceptions an owner. Resolve the underlying cause alongside the record.
- Make the AI show its records. Without a source, it shouldn't supply an answer.
Frequently asked questions
Is Salesforce AIforce the same as Headless 360?
Yes. Salesforce's help documentation identifies AIforce as the concept introduced at TDX 2026 under the Headless 360 name. The Headless Toolkit supplies its functionality.
Does Salesforce MCP let any AI connect to my CRM?
Yes, subject to limits. MCP-compatible agents can connect using the permissions you grant. Under Salesforce's plans, registration will be mandatory for each agent, and metering will apply after billing starts.
What is AI-ready data in Salesforce?
It means records that won't mislead a reader with no context. Accounts must correspond to real companies. Opportunities need a single owner and a source. Important fields must contain real information or an explicit unknown, and every edit needs a traceable origin.
Establish what your AI will read before connecting it. The full white paper, Salesforce Just Went Headless. Your Data Didn't., explains the broader picture. Or begin by examining your coverage gap.
Growth Orbit was founded by Steve Schilling. Is Your CRM Data Ready for AI? examines how a weak CRM affects AI.