Salesforce, Coupa and Asana Are Buying the AI Execution Layer

On June 15, Salesforce agreed to buy Fin, the customer service company formerly known as Intercom, for about 3.6 billion dollars. Fin's headline product is an AI agent that resolves support requests on its own across live chat, email, WhatsApp, SMS, phone and Slack. Salesforce was not paying for a chatbot. It was paying for software that closes tickets without a person in the loop, and it said Fin already resolves, on average, 76 percent of support volume end to end (Salesforce, TechCrunch).

That deal is the loudest note in a five week run where the largest names in business software all reached for the same thing. Not models. Not demos. The layer where AI agents actually finish work.

Four deals, one pattern

Take them in order. On May 12, Coupa, the spend management platform, acquired Rossum, a company whose software reads the invoices and shipping paperwork that move money and goods. Rossum is trained on tens of millions of documents and goes past older optical character recognition. Coupa CEO Leagh Turner put the math plainly: the company says it has delivered more than 300 billion dollars in customer savings over twenty years, and with Rossum it wants to help customers save the next 300 billion in five (Coupa).

On May 28, Asana bought StackAI for 75 million dollars. StackAI is a no code builder for AI agents that run inside the tools a company already uses, such as Salesforce, Slack and Google Workspace. Asana had been selling itself as the operating system for human agent teams, but its agents could plan and track work without doing it. CEO Dan Rogers said StackAI lets those agents go further, "agentifying the most complex business processes end-to-end" (TechCrunch).

On June 1, Salesforce signed to buy Contentful, the platform behind the content of more than 4,800 brands. Salesforce wants that content to feed Agentforce, its agent product, so agents can assemble and deliver personalized pages and messages without manual publishing steps. The price was not disclosed, and press reports placed it near 1 billion dollars, below Contentful's 3 billion dollar valuation from 2021 (Salesforce). Two weeks later came Fin, for 3.6 billion.

Why the execution layer, and why now

For two years the AI story in business software was about suggestions. A model drafts an email, summarizes a thread, proposes a next step, and a human decides. That phase is ending. None of the deals above are about better suggestions. They are about agents that read the invoice, write to the system of record, resolve the ticket, and publish the page.

Asana's own words are the tell. A planning tool that cannot execute is a notebook. The value sits one step later, in the action, so the incumbents are buying the action: document processing, workflow execution, content assembly, ticket resolution. Whoever owns that surface owns the workflow, and the data that runs through it.

There is a defensive read too. Salesforce, Coupa and Asana all face the same risk, that a customer wires up an agent from OpenAI or Anthropic and routes work around the incumbent's product. Buying the execution layer keeps the work, and the contract, inside the suite.

What this means for your business

At PaTech Labs we build AI agents for real operations, including voice agents that book and confirm rather than just answer. So we read these deals as a buying guide, not just headlines. Three practical takeaways for anyone choosing agent tools right now:

• Ask whether the agent executes, or only suggests. The market is rewarding execution. In a vendor demo, the question is not "can it draft this," it is "can it finish this and write the result back into our systems." Fin's 76 percent end to end resolution figure is the kind of number to ask for, measured in your own context.
• Weigh integration depth against lock in. A standalone agent builder is convenient until it becomes one feature inside a larger suite you may not use. If you adopt one of these tools now, assume the roadmap will start serving the acquirer's platform. Read the data and export terms before you commit.
• Watch who owns your data surface. These purchases are bids to own your documents, content and conversations. The more of your workflow runs through one vendor's agents, the harder it is to leave. Keep your systems of record, and your data, portable.

The pattern is clear enough to plan around. The money in enterprise AI is moving from the model that talks to the layer that acts. If you are deploying agents this year, judge every tool by what it finishes, not by what it says.

Related reading: our earlier look at enterprise AI agents learning to work together.