AI Agents 2024-2026: The Next Layer After SaaS

Between September and November 2024, the AI industry pivoted in public. The vocabulary shifted from “AI assistants” to “AI agents”. The product launches shifted from chat windows to workflow runners. The pricing models shifted from per-seat subscriptions to per-action consumption. And the structural question for every SaaS founder in 2026 became: where does the value capture move, when the user no longer types into your application but tells an agent to use it?

This piece walks through the 2024 watershed launches, what each one actually shipped, and the three structural shifts they imply for the SaaS playbook that the 2003-2010 cohort wrote down (see our SaaS pioneers’ playbook 2003-2010 for the precursor).

Why “agentic AI” became the 2024 watershed

The word “agent” had been floating in generative AI discourse since the 2022 ChatGPT launch. AutoGPT and BabyAGI in March 2023 showed the public what a self-prompting loop looked like, but both were research toys that broke on real workflows. LangChain in October 2022 and CrewAI in 2023 gave developers libraries for chaining tool calls, but production deployments were rare.

What changed in 2024 was the combination of three things that had to land at the same time. Models with reliable tool-use became commodity. OpenAI’s function calling shipped in June 2023, Anthropic’s tool use shipped in May 2024, and by mid-2024 every major frontier model could be trusted to call APIs without garbling JSON. Context windows expanded from 8k tokens in early 2023 to 200k in mid-2024 (Claude 3) and 1M in late 2024 (Gemini 1.5), which made multi-step workflows tractable. And the cost per token dropped roughly 10x in eighteen months, which made multi-call agents economically viable for routine tasks.

The result was that the same agent loop that crashed on the AutoGPT demo in 2023 could, by 2024, complete a real ticket triage, a real lead enrichment, or a real expense report. The watershed was not a research breakthrough; it was a cost-and-reliability threshold being crossed.

Artificial intelligence historians will probably mark the September-to-November 2024 window as the equivalent moment for agents that 2007-2008 was for cloud computing: the year the technology graduated from credible demos to enterprise procurement.

Salesforce Agentforce launch September 2024 + agentic CRM thesis

Salesforce Agentforce was announced at Dreamforce 2024 in mid-September. The pitch from Marc Benioff was explicit: agents are not a feature on top of Salesforce, they are the next layer that everything else ships through. The product included a no-code agent builder (Agent Builder), a library of pre-built agents for sales, service, marketing and commerce roles, and a marketplace called Agentforce Exchange that mirrored AppExchange but for autonomous workflows.

The technical bet was sharp. Agentforce agents are built on a model called Atlas (Salesforce’s reasoning layer), grounded in the customer’s Data Cloud (the unified data layer Salesforce had been building since 2020), and given access to the Salesforce metadata as their tool catalogue. An agent does not call “the API”; it reads the Salesforce object model, the flow definitions, the field-level security, and decides which records to touch.

Pricing landed at $2 per conversation for the Service Agent tier at launch, dropped to per-action billing in early 2025, and by Q1 2026 most Agentforce SKUs are consumption-priced (action-counts or token-counts) rather than per-seat. The shift away from per-seat is the most consequential pricing change in enterprise SaaS since 2007.

Three months after launch, Salesforce reported on the Q3 FY25 earnings call that more than 1,000 Agentforce deals had been signed, with reference customers including Wiley, OpenTable, ADP and Saks Fifth Avenue. By Q3 FY26 (late 2025), the figure was reported as more than 8,000 customers with Agentforce in production. The thesis that “every Salesforce customer eventually becomes an Agentforce customer” was clearly not just marketing.

For context on the underlying platform evolution that made Agentforce possible, see our Salesforce 2003-2026 retrospective.

Microsoft Copilot Agents + the M365 enterprise agent layer

Microsoft’s move came in two waves. The first was Copilot Studio (originally announced as Power Virtual Agents 2.0 in late 2023), which let enterprise admins build custom agents on the Microsoft Graph (the unified API across Outlook, Teams, SharePoint, OneDrive, Dynamics). The second was the Microsoft 365 Copilot Agents announcement at Ignite in November 2024, which positioned every Microsoft 365 customer as a potential agent author.

The Microsoft thesis differs from Salesforce in one important way. Where Salesforce treats the agent as the new front end on top of the CRM, Microsoft treats the agent as a workflow runner across the M365 stack. The agent reads your Outlook calendar, your Teams chats, your SharePoint documents, your Dynamics CRM records, and acts on all of them. The unifying layer is the Microsoft Graph, not any single application.

Pricing on the M365 side started at $30 per user per month for Copilot, with custom agents billed on top via a metered consumption model (per-message billing in Copilot Studio, with rates depending on the model and the message type). By 2026, Microsoft had introduced “agent-as-a-license” SKUs that let enterprises commit to a per-agent monthly fee rather than per-user, which is a clean signal of where the pricing is heading.

The strategic implication for SaaS founders is that Microsoft is using its existing enterprise distribution (more than 400 million paid Microsoft 365 seats globally) to push agents into the same workflows that third-party SaaS sells into. A vertical SaaS company in the Microsoft customer base now competes not against the Microsoft application, but against a Microsoft-hosted custom agent built by the customer’s own admin.

Anthropic Claude Computer Use October 2024: agents that drive the desktop

The most architecturally interesting launch of the 2024 window was probably the smallest in dollar terms. Anthropic released Claude Computer Use on October 22, 2024 as a public beta feature in the Claude 3.5 Sonnet API. The capability was simple to describe and hard to do well: given screenshots and a virtual keyboard and mouse, the model decides where to click, what to type, and what to do next, to complete a task on a real desktop.

The demo showed Claude opening a browser, navigating to a website, filling out a form, downloading a file, opening a spreadsheet, copying data across, and emailing the result. Each of those steps was a separate decision: where on the screen is the link, has the page loaded, did the click work, what is the next field. The model held state across all of them.

Computer Use is not, in late 2024 or even in 2026, a stable production primitive. The error rate on multi-step tasks is high enough that no enterprise has deployed it autonomously on real workflows. But it is the cleanest existence proof of the architecture every other vendor is building toward: an agent that does not need a custom API per application, because it can drive the application’s normal UI the way a human would.

The downstream implications are significant. If desktop-driving agents become production-reliable, the SaaS vendor that exposed a clean API to be called gets the same agent traffic as the SaaS vendor that never exposed one. The lock-in of “we have integrations to X, Y, Z” evaporates. The differentiator moves to the data and the workflow, not the integration surface.

OpenAI shipped a similar capability called Operator in January 2025, and Google’s Vertex AI followed with Project Astra integrations across 2025. By Q1 2026 every major model vendor offers a computer-using agent in either GA or public beta. The architectural bet has moved from contested to consensus.

AWS Bedrock Agents + Google Vertex AI Agents: platform-level agent infrastructure

The hyperscalers took the platform-infrastructure role rather than the application role. AWS Bedrock Agents had been in preview since July 2023, went GA in November 2023, and through 2024 and 2025 picked up multi-agent collaboration (March 2024), agent memory (May 2024), and prompt caching for cost reduction (August 2024). By 2026 Bedrock Agents are the default substrate for AWS-hosted enterprise agent workloads, with thousands of customers running production agents on the service.

Google’s Vertex AI Agent Builder shipped in April 2024 at Google Cloud Next, and combined the Gemini model family with the same retrieval-augmented-generation and tool-use plumbing AWS had. Vertex AI added the “Agentspace” surface in late 2024, which is the Google equivalent of a corporate agent marketplace.

What both hyperscaler offerings do, deliberately, is stay out of the application-layer competition. AWS does not ship a CRM agent; it ships the runtime that lets you build one. Google does not ship an HR agent; it ships the components. The bet is that the agent layer is too horizontal to be locked into one application vendor, and that the infrastructure provider that wins the most agent-runtime workloads inherits the same kind of share AWS won in cloud compute.

The pricing model on Bedrock and Vertex is per-token-consumed, with surcharges for agent-specific features (memory storage, tool-call orchestration, evaluation harnesses). The economic shape is much closer to traditional cloud billing than to per-seat SaaS, which is consistent with the broader shift from seat-based to consumption-based monetization that 2026 is making consensus.

What this changes for SaaS founders 2026 (UX shift, monetization shift, platform-tax tension)

The three structural shifts a SaaS founder in 2026 has to plan around are:

UX shift. The user is increasingly not in your application. They are in an agent surface (Agentforce, Microsoft Copilot, a custom Bedrock agent, a Slack channel where an agent is a participant) and your application is being called as a tool. The implication is that the SaaS vendor needs to expose an agent-friendly surface (a clean API, well-documented tool schemas, good error messages, idempotent operations) in addition to the human-friendly UI. The vendors who treat API quality as a back-burner concern in 2026 will lose agent traffic to the vendors who treat it as a feature.

Monetization shift. Per-seat pricing assumed one user per seat per month. Per-action or per-workflow pricing assumes a billable event per outcome. The 2026 transition is not clean: most SaaS vendors are running hybrid models (per-seat base plus per-action overage) because cutting over to consumption-only loses revenue predictability. But the direction of travel is consensus. The 2003-2010 cohort that wrote the per-seat playbook (see SaaS pioneers’ playbook) is now being asked to rewrite it.

Platform-tax tension. This is the 2026 version of the Lock-In is Inevitable argument written from inside Salesforce in 2008. When an agent platform sits above your SaaS application and orchestrates calls across multiple SaaS vendors, the platform captures the cross-application value, and the underlying applications become commodities. The agent platform is the new lock-in. Salesforce, Microsoft, Google, AWS, Anthropic and OpenAI are all trying to own that layer. The independent SaaS vendor who underwrites this tax in 2026 is the equivalent of the AppExchange ISV who signed up for the 15 percent Salesforce platform fee in 2007: the platform owner takes a margin on every workflow, and the application owner takes whatever is left.

There is no easy exit from the tension. Building your own agent platform means competing with three trillion-dollar incumbents and Anthropic and OpenAI. Refusing to integrate means the agent traffic routes around you to a competitor that did. Most SaaS founders in 2026 are picking the second-worst option: integrate with two or three agent platforms, expose a clean tool API, and try to push enough product differentiation into the data and the workflow that the platform tax is bearable.

The 2026 topic page for SaaS on this archive collects the longer pieces on what survived from the 2003-2010 playbook, what got reinvented under new branding, and where the agent layer fits into the multi-decade picture. The short version is that agents are not a new SaaS competitor; they are the new SaaS distribution channel, and the founders who treat them as such will outlast the ones who treat them as either hype or threat.

FAQ

What is an AI agent vs a chatbot?

A chatbot answers a prompt with a reply. An AI agent is given a goal and a toolbox (APIs, file system, browser, scripts), and decides itself which tools to call, in what order, and when the goal is reached. The agent maintains state across multiple steps, recovers from errors, and only returns control when it has a result. ChatGPT in 2022 was a chatbot. Anthropic Claude Computer Use in October 2024 was an early agent: the model could see screenshots, click, type, and run shell commands on its own. The distinction matters for monetization, because a chatbot is billed per message and an agent is billed per task completed.

When did agentic AI become mainstream?

September to November 2024 is the conventional watershed window. Salesforce launched Agentforce at Dreamforce in mid-September 2024, Anthropic released Claude Computer Use on October 22, 2024, and Microsoft announced custom Copilot Agents in Microsoft 365 in November 2024 at Ignite. AWS Bedrock Agents had been in GA since November 2023 and picked up multi-agent collaboration and memory features through 2024. Google Vertex AI Agent Builder shipped in April 2024. By Q1 2026 every major enterprise platform (Salesforce, Microsoft, ServiceNow, Workday, SAP) had shipped at least one agentic surface, and the question had moved from “is this real” to “which platform do we standardise on”.

Will AI agents replace SaaS?

Not replace, layer on top of. SaaS applications still own the data, the workflow, and the compliance posture. Agents call the application as one of their tools. The real shift is in the UI: if a sales rep tells an agent “update the deal stage and notify the AE”, the agent calls the CRM API instead of opening the CRM screen. The CRM is still there, but the user-visible surface area shrinks, and the SaaS vendor competes for traffic from the agent platform instead of from the human user directly. Vendors who refuse to integrate with agent platforms will see their UI traffic flat or declining while the agent-mediated traffic at competing vendors grows. The pattern looks similar to the 2010 mobile-versus-desktop transition: the underlying applications did not disappear, but the surface they were consumed through changed.

How do AI agents change SaaS monetization?

Three shifts are already visible in 2026. Per-seat pricing erodes because the consuming unit is now the agent, not the user, and an agent can run thousands of seats’ worth of work on a single subscription. Per-action or per-task pricing rises (a workflow completed by an agent is a billable unit, similar to how Bedrock and Vertex bill per token). And the platform hosting the agent (Agentforce, Microsoft Copilot, a hyperscaler’s agent runtime) can extract a margin on every cross-application action, which is the new platform-tax debate. Most SaaS vendors in 2026 are running hybrid pricing (per-seat base plus per-action overage) as a transition, but the long-run consensus is consumption-based monetization aligned with the agent layer, not seat-based monetization aligned with the human user.