Salesforce Opens the Door to Claude. Who Keeps the Software Budget?

Marc Benioff spent Tuesday explaining how customers could use his company's software from somebody else's AI assistant. At Dreamforce, Salesforce introduced AIforce, which makes its data, workflows and business rules accessible through outside interfaces, including the Salesforce in Claude beta ^1,17.

The stakes are a business that reported $11.35 billion in quarterly revenue on August 26. Salesforce (CRM) wants an employee starting work in Claude to remain a paying Salesforce customer underneath ².

That is a sensible defense. It also exposes the central problem facing enterprise software: customers can keep the database while changing which company they depend on to get work done. Partnerships with model providers preserve access to better AI. They do not automatically preserve the incumbent's share of the software budget.

The partners also want the work

An AI assistant that drafts a customer email needs information. An assistant that changes a contract needs authority: the correct account, approved terms, permissions and a record of what happened. Salesforce's pitch is that those controls remain in its platform even when the request begins elsewhere ^1,3.

But OpenAI and Anthropic are pursuing that operational layer themselves. In February, OpenAI paired its engineers with Accenture, Capgemini, McKinsey and BCG to help companies move agents into core business processes. The competition reaches implementation and customer relationships, beyond access to a model ⁴.

Anthropic moved further into finance in May with ten agents for work including credit memos and financial-statement audits. Those products can be adapted to a firm's policies. They demonstrate an ambition to sell completed professional work, although an announcement does not establish reliable performance at scale ⁵.

The strategic inference is uncomfortable: today's model supplier can become tomorrow's application vendor without terminating the partnership.

Three businesses, three defenses

Salesforce's defense is its customer data and the business rules attached to it. Its August agreement with Anthropic combines product integration with reciprocal adoption, including Anthropic choosing Salesforce as its preferred CRM. That creates a distribution opportunity, but the announcement does not disclose a revenue-sharing formula that establishes who captures the incremental spending ⁶.

ServiceNow (NOW) has a more explicit answer to the charging question. It announced expanded relationships with both OpenAI and Anthropic in January. Its Action Fabric strategy allows outside agents to execute work through ServiceNow. Calls through its agent connector consume Assist units, although usage already included in a customer's contract does not immediately produce additional revenue ^7,8,9,16.

The implication is that ServiceNow can earn money even when an employee starts in another vendor's assistant. Its defense depends on the approval, fulfillment or other paid action continuing to run through ServiceNow. If customers move that execution elsewhere, connecting to more assistants provides less protection.

Palantir (PLTR) sells a different kind of dependency. Its Artificial Intelligence Platform, AIP, connects models to an operational representation of a business: data, logic, actions and security. Its documentation supports multiple model families, including OpenAI and Anthropic. The intended asset is the machinery that turns a model's answer into a controlled business action ¹⁰.

That can be expensive to replace because the customer must recreate how decisions connect to operations. Yet this advantage is conditional. OpenAI's deployment engineers and consulting partners are competing to build those connections too. Palantir has to make deployment and ongoing operations sufficiently valuable that customers keep buying the platform ^4,10.

The earnings reject a single software story

The latest reported quarters show three growing businesses at very different speeds: Salesforce's total revenue rose 11%, ServiceNow's 24% and Palantir's 93% from a year earlier. These company-reported figures are corroborated by independent earnings coverage ^{2,11,12,13,14,15}.

The chart defeats a blanket claim that enterprise software is already collapsing. It cannot identify which partnership creates lasting pricing power. Palantir's figure includes government revenue; Salesforce's includes an acquired business. Different starting sizes and customer mixes also matter.

The AI-specific numbers require more care. Salesforce expanded its Agentforce annual recurring revenue definition this quarter to include additional AI offerings, Slackbot and Headless 360. Comparing that measure with ServiceNow's AI annual contract value would mix different definitions and mistake sales metrics for recognized revenue ^11,12.

Keeping the database is not the same as keeping the customer's budget.

More automation can weaken pricing power

The strongest case for incumbents is practical. Their customers have already configured permissions, approvals and operating processes. Better models make those investments more useful. An employee who previously needed a specialist to navigate the software might now initiate valuable work through a familiar assistant.

This could expand the market: more people requesting more actions, with the existing platform collecting usage revenue. The expensive work of cleaning data and establishing governance still has to happen, a limitation highlighted by analysts examining Salesforce's launch ³.

The problem is how the gains are divided. If automation reduces the number of employees needing paid seats, new usage revenue must compensate. If the AI provider owns the daily customer interaction, it can influence which applications get invoked. An incumbent may process more work while losing negotiating power at renewal. These are economic risks, not outcomes established by this quarter's results.

Costs create another constraint. ServiceNow said its 2026 subscription gross-margin outlook reflected increased hyperscaler usage and faster AI adoption. More AI activity can increase both sales and the bill for delivering them ¹².

The useful measure is therefore gross profit from reliably completed workflows after model and delivery costs. Agent counts and demonstrations cannot establish that profitability.

Watch the contract after the demo

The likely dividing line is between software that remains necessary to execute a business process and software whose contribution an assistant can reproduce. Salesforce, ServiceNow and Palantir each have a credible claim to the former. Their partnerships strengthen distribution while testing how much of the customer relationship they can surrender without surrendering economics.

The next results and renewals should answer three questions. Does outside-agent activity generate incremental paid usage? Does that revenue offset any pressure on human-user subscriptions? Do customers and vendors retain enough savings after model costs, integration and exception handling?

For Benioff, opening Salesforce to Claude is a bet that customers will pay for the business controls behind the conversation. The next useful evidence is whether those customers expand their contracts once the work moves into Claude.