"Why should you ever log into Salesforce again?"
When Salesforce introduced Headless 360 this spring, co-founder Parker Harris posed that question. Salesforce has rebuilt the platform around its answer: AI agents can work directly in Salesforce, without a screen between them and the system. The question is reasonable.
My next question would be this: if no one opens Salesforce, who catches a problem in a record?
At one of our clients, Salesforce contained about 113,000 account records for a market of about 8,500 companies. We compared the accounts with Dun & Bradstreet company data and couldn't connect 70% to any company we could identify. A website appeared on only a quarter of the records. More than 13,000 duplicated companies already represented in the system, and 428 accounts belonged to a single corporate family.
Connect an AI agent now, and Salesforce can give it every one of those records. It can quickly and confidently read, reason and act on them. That is both the appeal of headless and the difficulty it presents.
What Salesforce announced
Across six months, Salesforce announced the changes in three stages.
April 15, 2026: Headless 360. Salesforce used its TDX developer conference to introduce Headless 360 as "the capabilities your agents need most, exposed as an API, MCP tool, or CLI command." Put simply, an AI agent can now use code to perform what you previously did through Salesforce's screens. More than 60 new MCP tools and more than 30 preconfigured coding skills accompanied the launch. MCP means Model Context Protocol, the standard that connects AI agents to outside systems. Governance was part of the release too: Agent Fabric, Salesforce's control plane; observability and session tracing that reveal agents' actions and reasoning; and a testing center for evaluating agents before and after launch.
September 2026: AIforce. Dreamforce brought a new name and a broader concept. Salesforce describes AIforce as a "live, composable interface layer that brings the rest of the platform to any surface." Its foundation is now named the Headless Toolkit. There, "every capability is accessible via API, MCP, or CLI," and people and agents "work from the same data, logic, and governance." Another addition, MCP Risk Scores, checks an outside MCP server at registration for threats including prompt injection, before authorizing a connection.
November 2026: registration and metering. The next stage appears in a Salesforce help article dated September 17. November is Salesforce's target for new security controls. Registration will be required for each AI agent using its APIs or MCP servers, giving that agent a separate identity and, if you choose, permissions narrower than a human user's. A successful call will count as a "Headless Platform Interaction" that consumes Flex Credits. Salesforce hasn't published the rate and promises 30 days' notice before metering begins. For existing customers with agents using APIs, registration and the new billing model arrive at renewal. Anyone using a Salesforce MCP server must register agents within three months of Salesforce's notice.
The shift extends beyond Salesforce. HubSpot's remote MCP server became generally available on April 13, 2026, allowing writes to contacts, companies, deals and activities. Claude, ChatGPT and Copilot can therefore read and update that CRM directly. Across the industry, CRMs are becoming infrastructure for agent calls rather than applications that people need to open.
What "headless" actually means
Software combines two parts. One holds the data and enforces its rules; the other presents that information to people. Screens, dashboards and forms make up the visible part, or "head."
Making a platform headless leaves its data, business logic and permissions in place while making the visible interface optional. An agent can request records, change a field, create a task or execute a workflow directly. None of those actions has to pass through a screen.
Salesforce APIs aren't new. The change is whom the platform is being built for: agents are becoming its primary users, and its capabilities are being designed for them to call.
What people think headless will let them do
Four capabilities account for much of the enthusiasm. Each is real.
Ask instead of look. A leader can put the question to an assistant instead of navigating a dashboard: "Which deals will close this quarter? Why is pipeline down? Which accounts should we focus on?" The AI retrieves the answer from Salesforce.
Use any AI, not just Salesforce's. MCP support lets you connect Claude, ChatGPT, Copilot or your own agent. That frees your CRM from depending on a single vendor's AI.
Let agents do the work. Without a person typing, agents can change records, record activity, route leads, draft and send outreach, and advance deals through stages.
Build faster. One customer Salesforce cites put production agents into use in 12 days. Developers can work in their own coding tools instead of building through Salesforce's interface.
There are substantive governance controls as well. Existing permissions and sharing rules still constrain agents; registered identities will identify them. Salesforce flags risky MCP servers and lets you trace agents' actions. Those protections deserve to be used.
What headless doesn't change
None of those capabilities repairs the underlying record.
Data quality appears in neither the Headless 360 announcement nor Salesforce's Dreamforce summary for IT leaders. Those announcements have a different purpose, so that omission isn't a criticism. The problem is that many buyers will assume the data has been handled when it hasn't.
When headless removes the screen, it also removes the place where people quietly corrected records. A blank field prompts a sales manager to question the rep. A repeatedly deferred close date tells a leader to treat it as a placeholder rather than a forecast. A recent Inc. article about AI and CRMs captured the issue: "the gaps that humans compensated for without thinking become the exact places agents fail."
The agent has no way to know which fields deserve trust. It uses the available entries and infers what's missing. Those inferences can produce answers that are confidently wrong.
Gartner makes a similar distinction. Through 2026, it expects organizations to abandon 60% of AI projects that aren't supported by AI-ready data. It also reports that 63% either lack suitable data management practices for AI or don't know whether they have them. Salesforce's 2025 survey of 4,050 sales professionals supports that concern. Respondents using AI agents put manual errors and duplicate data first among their data problems; 46% reported harm to sales from data-quality issues. Salesforce sells both the platform and the agents. This is a candid vendor survey, and should be read as one.
For practical purposes, "AI-ready data" means records that won't mislead a reader who brings no context. Most CRMs fall well short of that standard.
Four ways bad data goes headless
My book, Effort Is Not Evidence, describes four related problems in revenue data. Headless introduces none of them; it takes away the person who previously spotted them.
1. The agent can only see the market that's in your CRM
An agent choosing priority accounts can choose only from the accounts it sees. At another of our clients, an IT-services firm, the offer fit 18,775 companies. Salesforce contained about a quarter of them. Using that CRM alone, an agent would confidently recommend accounts from that quarter without mentioning the other three-quarters of the market. They would be invisible to it.
2. The agent reads seller activity as buyer progress
CRM entries largely describe sellers' actions: booking meetings, creating opportunities and moving stages. At a third client of ours, a B2B fintech, 267 of 397 demos became "qualified" opportunities, about two in three. Yet 80% stopped at the first stage, and just 5% closed. A forecast based on those records would describe what sellers entered rather than what buyers did. Holding a meeting establishes engagement, but doesn't establish a deal's stage, amount, probability or close date. An agent can distinguish those things only when the record distinguishes them.
3. The agent inherits every place your systems disagree
Your CRM shares the picture with other systems. Outreach software, call recordings, meeting logs and finance each maintain a version of events. At the global services client from the opening example, a new outbound program reported sixteen opportunities. We uncovered three meanings of "sixteens": an outbound meeting-log status, contacts we had matched, and a written account of program results. None represented sixteen CRM opportunities. An agent reading across those systems could have returned the version it encountered first.
4. AI scales all three, and now it writes back
These agents also write: they change fields, record activity, send messages and move stages. LeanData sells lead-routing and revenue orchestration software. In May 2026, it surveyed 157 of its own customers' go-to-market practitioners. The results concern those customers rather than companies generally. Even with that limit, they are telling: 93% had at least one deployed AI agent, nearly one in three couldn't count the agents acting on their records, 30% had discovered actions lacking an audit trail, and 27% had experienced overlapping prospect contact from multiple tools or agents.
Our own work exposed this fourth problem too. An early version of our AI research engine couldn't look anything up. Over more than 1,000 runs, it claimed evidence of budget authority for every researched contact. Its cited evidence was simply the job title we had supplied. Unless the surrounding system accepts "I don't know" as an answer, an AI unable to find one will invent it.
The meter changes the math
Another consequence of headless has received little discussion so far.
When metering begins, Salesforce agent calls will consume Flex Credits. We still don't have a published rate, but we do know the charging basis: interactions rather than seats. To answer a question about 8,500 companies, an agent reading 113,000 account records also reads 104,500 records that don't matter. Eventually those reads will incur charges. Duplicates will carry a bill as well as create confusion.
I'd wait for Salesforce's rate before assigning a cost to that. Poor data already wastes effort; a metered model will put some of that waste on an invoice.
Before you plug an AI into Salesforce
The implication is to sequence the work, not postpone it. This is the order I'd follow.
1. Match your CRM to the market first
Begin with a company database, not Salesforce. Build a clean market list of companies that fit your offer, counting each at the parent level, and match your Salesforce accounts against it. Any AI project should begin with the match rate. It reveals how much of the agent's input corresponds to real companies and how much of your market lies beyond its view.
- Who owns it: revenue operations.
- What good looks like: you know the proportion of accounts that match real companies in your market and the proportion of that market absent from Salesforce.
2. Inventory every agent and tool that touches the CRM
Turn the coming registration requirement into an inventory. Record every agent, integration and tool with read or write access to Salesforce, identifying its owner and the changes it can make. Apply the narrower permission sets registration enables. For most agents, write access isn't necessary at the start.
- Who owns it: the CRM administrator, with security.
- What good looks like: every agent appears on a complete list with its purpose, owner and minimum workable permissions.
3. Define the numbers before agents answer them
Without a written definition, the response to "what's our pipeline?" will follow whatever definition the records suggest. Specify the qualifying stages, evidence and companies. Separate the evidence levels: a held meeting or reply establishes engagement; an accepted mutual plan or initiated security review shows buyer progress; a contract signed and recorded in finance is a verified outcome. Those levels aren't interchangeable.
- Who owns it: revenue operations, with sales and finance leadership.
- What good looks like: each number a leader will ask about has one documented definition.
4. Make unknown mean unknown
An empty field invites a guess. Require meaningful entries in important fields, block the duplicates you can, and label missing information as unknown. Finance won't invoice a nonexistent customer. Records supporting your CRM forecast should face a comparable standard.
- Who owns it: the CRM administrator.
- What good looks like: every field an agent depends on contains a real value or an explicit unknown.
5. Read first, write later
First let agents answer checkable questions with read-only access. Put their proposed changes in a queue for a person to approve before a write occurs. Broaden write permissions only as those proposals repeatedly prove correct.
- Who owns it: the CRM administrator, with the business owner of each agent.
- What good looks like: a person approves record changes until the agent's track record supports independent writes.
6. Log every write, and give exceptions an owner
Salesforce's tracing covers its agents. Your records need equivalent coverage for every writing tool, regardless of vendor. Maintain a single list of unexplained edits, duplicates and differences across systems. Assign each exception an owner, a likely cause and a date. Remedy the source of the problem, whether a field, rule, integration or incentive, as well as the record.
- Who owns it: revenue operations.
- What good looks like: key-field edits can be traced to their human or automated source, and the unresolved exception list gets smaller over time.
7. Make the AI show its records
Ask an agent to identify the records supporting its answer before trusting it. Without them, it's guessing. Before using it for questions you can't answer yourself, test it on questions whose answers you already know.
- Who owns it: whoever owns the agent.
- What good looks like: the records supporting an answer accompany every response to a leader.
The common trap
The usual trap is blaming AI for an incorrect answer and changing the prompt, model or vendor. After six months, the answers remain wrong because the inputs remain what your team entered.
Waiting for flawless data is the other trap. Flawless data won't arrive, and it isn't required. Know its quality, weaknesses and the person responsible for repairs. Make sure agents can respond "I don't know" when the evidence in a record is insufficient.
What this paper can't tell you
Growth Orbit's anonymized client figures describe our work at three companies, not a market survey. Salesforce's details reflect its announcements and documentation as of October 12, 2026. Neither registration nor metering has begun, and no metering rate is published; the cost implications discussed here are reasoning rather than measurements. LeanData's and Salesforce's surveys describe their respective respondents. AI agents can and will operate in Salesforce. Their performance will depend on the records they consume.
Where to start
Your first agent needn't wait for answers to every question here. Begin by establishing how much of the market your CRM includes and how much of its contents belongs to your market.
A Coverage-Gap Analysis establishes that. Working read-only, it compares the CRM with companies that fit your offer and exposes the gap before an agent reasons from those records. For a wider view, the GTM Assessment examines market, data and sales motion together.
Whichever route you choose, establish what the AI will be reading before it begins.
Frequently asked questions
What is Salesforce Headless 360?
Salesforce announced Headless 360 in April 2026 to make the platform directly usable by AI agents. Its capabilities became available through APIs, MCP tools and command-line calls. The September 2026 announcement renamed the concept AIforce, powered by the Headless Toolkit.
Will Headless 360 or AIforce fix my CRM data?
No. Headless changes access to the records rather than their contents. Duplicate accounts, unmatched companies, blank fields and seller-entered stages remain. Unless the records explicitly expose those gaps, an agent will treat the entries as facts.
What should I do before connecting AI agents to Salesforce?
Begin by matching CRM accounts to your market and listing every agent and tool. Document the definitions of numbers agents will report, explicitly identify unknown information, and give agents read-only access first. Trace every write, assign exception owners, and require the underlying records with each answer.
The shorter companion is Salesforce MCP Won't Clean Your CRM. Here's What Will.