The release of “Claudeforce” by Anthropic and Salesforce has rekindled debate over who will win in software in the AI era.

The core question is whether large software companies that control existing systems of record holding key corporate information will continue to dominate, or whether vertical AI startups specialised in specific tasks will seize new opportunities. Against this backdrop, Seema Amble (시마 앰블), a partner at Andreessen Horowitz (a16z), drew attention by addressing related perspectives in a recent post on social media platform X (Twitter).

She first introduced the argument that incumbent system companies could gain an advantage. The logic is that if customers connect Salesforce data directly to AI such as Claude or Codex, there is no longer a need to use a new AI-only app.

Claudeforce supports handling CRM tasks within Claude without having to open the Salesforce app directly. Incumbent system companies such as DocuSign, Atlassian and Klaviyo are also moving beyond providing information toward carrying out tasks on users’ behalf.

The issue is that as general-purpose agents such as Claude handle work across multiple apps, the interface people see on screen and the system that holds the actual data are slowly splitting apart.

Amble says this situation will create opportunities for vertical AI startups. “For a vertical startup to win, it has to do one specific task better than an incumbent enterprise agent or a general-purpose agent,” she said. “By focusing on a single task, it can go deeper, build clear data assets, learn the details of the work through them, and secure a learning loop that improves over time.”

Amble divided agents into four stages. The first is a search type that finds and summarises information. The second is a processing type that updates records according to set rules. The third is a policy type that judges matters with no single right answer based on company policies. The fourth is an autonomous type that decides for itself what an organisation should do based on strategy and risk.

“A year ago, most incumbent enterprise products were stuck at the search stage,” Amble said. “Now, a significant number are handling real work, and some, like DocuSign Iris, have started to make judgments in a narrow range. But the scope of those judgments is still limited to data stored in their own systems.”

Amble stressed that real work is much larger than simple records. “A single contract does not contain an entire legal workflow, and one ticket does not show the whole process of resolving a customer problem,” she said. “Even if a general-purpose agent can connect to multiple apps, that is not everything. Different apps record the same customer information differently, causing confusion, and you may not be able to access the other company’s system in the first place.”

Amble laid out four conditions under which vertical AI businesses can work. They are whether an expert can quickly judge if there is something wrong with the output, whether the task requires judgment, whether the work is repeated often enough to provide many learning opportunities, and whether it can start with a small task and expand to the full scope of the job.

“Many legal, tax and accounting tasks fit these conditions. Investigating the cause of defects on manufacturing sites is the same,” she said. “Some work is sufficient with incumbent systems and general-purpose agents alone, but in areas where work is repeated often, judgment is required and the learning loop is solid, vertical AI startups still have significant room to win.”