Claude Cowork and Pega: beyond the “SaaS‑pocalypse” — agents, orchestration, and enterprise reality
Introduction: beyond the ‘end of SaaS’ narrative
In January and February 2026, the ‘end of SaaS’ narrative resurfaced with renewed intensity. This time it was triggered by something very tangible: Claude Cowork (Anthropic’s agentic desktop mode) moving from answering questions to executing real work across files, folders, data and documents.
When you see an agent do in minutes what used to take you hours, it is easy to conclude that an AI layer can sit on top of, and compress the value of, many traditional applications. The market reaction (and the label ‘SaaS‑pocalypse’) reflected that fear.
My view is more pragmatic: the concern is understandable, but the diagnosis is incomplete. Not all software exists to complete isolated tasks, and not all enterprise work is document-centric. The key distinction is between individual productivity and organization‑level process orchestration:
• Individual productivity: fast, local execution of unstructured work.
• Organizational orchestration: long‑running processes with persistent state, complex decisions, cross‑team accountability, governance, and auditability over time.
That is what this article is about: why Claude Cowork is highly effective at first, and why Pega remains difficult to replace at enterprise scale for the second.
1) Claude Cowork: Why it has attracted so much attention
Anthropic describes Cowork as ‘Claude Code for the rest of your work’: a desktop execution environment where Claude can operate on local files and tools with more agency than a chat interface. Claude Cowork does not just reason; it plans and acts. Anthropic also positions it as a research preview and explicitly notes that agent safety is still evolving.
In practice, the interaction model is simple and effective: you grant access to a specific context (folders/directories), describe the desired outcome (rather than a step‑by‑step recipe), and the AI agent returns the deliverable without you having to micromanage every click.
A key inflection point is extensibility. Claude Cowork introduced department‑oriented agentic plugins (sales, legal, finance, support…) designed to encode ‘how we do things here’: which data sources to use, which steps to follow, and which commands to expose. This approach is highly adaptable. If you have ever lived through the chaos of scripts, macros, and semi‑manual processes, you immediately see why this is compelling.
2) Where Cowork (and Claude Code) genuinely shines
Cowork perform particularly well when the work is unstructured and desktop‑centric. For example: cleaning up folders, extracting tables from PDFs, drafting reports, summarizing documents, synthesizing multiple sources, or producing a solid first draft.
If you try it for a couple of days, the value is obvious: it saves time without forcing you to redesign how you work. The mental model is closer to a ‘teammate’ than a chatbot (plan, execute, adjust) operating in a multi‑step loop.
Another important pillar is the Model Context Protocol (MCP), an open standard created by Anthropic to connect AI agents and assistants to business tools and data repositories (MCP servers and clients) through a common protocol. In practice, Cowork + MCP aims to become a ‘work layer’ across multiple systems, accelerating tasks that previously required switching between many different applications.
3) The “SaaS‑pocalypse”: more transition than extinction
The market reaction to Claude Cowork had two main drivers. First, a rational one: if AI enables one person to complete more work, seat‑based pricing models face pressure, and some software categories need to re‑articulate their value.
Second, a narrative driver. 2025 was the year of ‘everything with AI’ in decks, funding rounds, and valuations. In a market that rewards clean stories, ‘AI agents replace SaaS’ is easier to sell than a nuanced reality: AI agents shift which layers of the stack become commodity and which parts become strategically important.
In my view, that is what is really happening. Commodity layers (nice UI and isolated tasks) lose value; the hard parts (governance, long‑running processes, compliance, integration, decisioning) become more valuable. Within that landscape, Pega does not compete with Claude Cowork on creating documents or spreadsheets; it competes (and wins) at ensuring an end‑to‑end case is resolved on time, with operational control and auditable outcomes.
4) The natural limits of desktop-centric agents
This is where, in my opinion, the debate often breaks down: an enterprise does not live in sessions, it lives in states.
Claude Cowork is optimized for session-based actions on local files and tools. Enterprise operations, however, are usually built around a ‘case’ that has persistent state, SLAs, work queues, separation of duties, comprehensive audit trails, and cross‑functional reporting.
Real processes often span days, weeks, or months moving across teams and systems. The key question is not ‘Can an agent perform the task?’ but ‘Who is accountable and where do control and auditability reside?’.
Large language models are probabilistic by design. They are exceptional at proposing and generating content, but on their own, they are less suitable for scenarios that demand strict compliance, determinism, and traceability. This is a key aspect we should never forget.
5) Why Pega becomes more valuable in an agentic enterprise
Pega was not built to automate isolated tasks. It was designed to orchestrate work end‑to‑end. That difference becomes more remarkable, not less, when you introduce AI agents.
Four Pega capabilities are particularly relevant in the agentic enterprise world:
- Pega Case management for long‑running work: cases that span days, weeks, or months across multiple departments, with rigorous auditability, dynamic SLAs, and complex assignments.
- Rules and decisioning: The ability to change policies and logic without rebuilding the entire process.
- Governance and security: when an agent generates or executes, someone must answer to audit. In Pega, the trail lives in the case, not in an isolated chat transcript or log line.
- Enterprise integration: the real value is not producing an artifact but operating against core systems with resilience and control.
In addition, Pega GenAI Blueprint, which operates as a purpose-built design agent, combining generative AI with deterministic workflow and decisioning capabilities in a collaborative environment. It incorporates principles such as data isolation (regional residency and tenant‑separated storage), avoiding training foundational models on client data, and governance throughout the lifecycle using multiple specialized design agents. Enterprise‑grade, in practice, is defined by concrete architectural and governance guarantees. Could an AI agent achieve all of this?
6) The hybrid model that works
A practical way to think about coexistence is as a division of responsibilities:
• Claude Cowork as the executor of unstructured work: synthesis, drafting, reconciliation, research and document-centric deliverables.
• Pega as the orchestration layer: instantiating and governing end‑to‑end processes through cases, applying business rules, managing SLAs, assigning work, recording every step for audit, integrating with core systems and using AI with guardrails in a predictable and auditable approach.
A repeatable enterprise pattern looks like this:
(1) An event arrives through any channel (email, portal, API), (2) Pega creates or updates the Case Type and applies eligibility and routing rules, (3) When unstructured work is needed outside the workflow, Pega invokes an AI agent (e.g., Claude Cowork) through a controlled connector or API, (4) The agent returns results and supporting evidences, and (5) Pega determines and executes the next steps with end-to-end traceability within the case.
This is not a lone agent operating without constraints. It is an agent operating inside guardrails: AI drives local productivity; Pega provides state, management, governance, security, and accountability. Pega’s continued progress in agentic capabilities further strengthens the case for this hybrid model.
Closing perspective
Claude Cowork demonstrates that AI agents can now perform tangible work. But an enterprise is not a collection of tasks; it is a network of responsibilities. Pega functions as the orchestration layer that allows organizations to scale agentic capabilities without compromising governance or accountability. How are you thinking about the governance layer in your agentic deployments? Are you seeing a growing divide between individual execution and enterprise-level orchestration in your own organizations?