Software as a service reinvented software delivery by moving to the cloud; now generative and agentic Artificial Intelligence are triggering a new discontinuity. These systems are already drafting code in Cursor, resolving support tickets in ServiceNow, preparing journal entries in Workday Financial Management, and writing ad copy in Adobe’s Experience Cloud. Model costs are falling fast while accuracy rises, with OpenAI’s latest reasoning model dropping 80 percent in two months. Bain forecasts that within three years many routine, rules-based digital tasks could shift from human plus app to agent plus API, forcing SaaS leaders to reassess where value is created and captured.
The firm recommends mapping each workflow on two axes: the potential for Artificial Intelligence to automate user tasks and the potential for Artificial Intelligence to penetrate the workflow itself. Six signals of user-task automation include task structure and repetition, error risk, context dependence, data availability and structure, process variability and exceptions, and reliance on human interfaces. Penetration potential is indicated by external observability, industry standardization, proprietary data depth, switching and network frictions, regulatory or certification barriers, and agent protocol maturity. High automation can expand the market and top-line opportunity, while high penetration makes it easier for third-party agents to siphon usage and margin.
Plotting against these axes yields four scenarios. In “Artificial Intelligence enhances SaaS” strongholds (low automation, low penetration), incumbents should use Artificial Intelligence to boost productivity and price time savings, as in Procore’s project cost accounting or Medidata’s clinical-trial randomization. Where “spending compresses” (low automation, high penetration), such as HubSpot list building or Monday.com task boards, incumbents must launch their own agents, deepen integrations to raise switching costs, and restrict critical endpoints. In “Artificial Intelligence outshines SaaS” gold mines (high automation, low penetration), like Cursor’s code editor and Guidewire’s claims adjudication, leaders should build end-to-end agents and shift from seats to outcomes. In “Artificial Intelligence cannibalizes SaaS” battlegrounds (high automation, high penetration), including Intercom tier 1 support, Tipalti invoice processing, and ADP time-entry approvals, winners will scale agent orchestration and decide whether to be the neutral agent platform or the unique data supplier; only a few giants such as Salesforce can do both.
Bain says a new three-layer stack is forming: systems of record at the base, agent operating systems in the middle (early versions include Microsoft Azure AI Foundry, Google Vertex AI Agent Builder, and Amazon Bedrock Agents), and outcome interfaces on top via tools like Teams, Slack, or custom apps. Interoperability is the bottleneck. Protocols such as Anthropic’s Model Context Protocol and Google’s Agent2Agent standardize syntax for tool calls and handoffs but stop short of a shared semantic vocabulary and policy layer. The first widely adopted semantic standard will likely trigger network effects and reshape value flows, creating a powerful gatekeeper role. Incumbents can lead by moving quickly, even if that requires selective open sourcing or new monetization models.
The playbook for SaaS leaders: make Artificial Intelligence central to the roadmap and deliver do-it-for-me workflows; turn proprietary usage, domain, and transaction data into the defensible moat; invest according to the four scenarios; close the semantic gap by standardizing internal objects, open sourcing select schemas where you lead (as ServiceTitan and Guidewire do), encoding constraints like approvals and compliance in your data model, and rallying partners and customers to your standard; rethink pricing toward outcomes, with precedents from Intercom and Salesforce; and build company-wide Artificial Intelligence fluency. The conclusion is clear: disruption is mandatory, but obsolescence is optional.
