The January 2026 SaaS market dislocation wiped over a trillion dollars from software market caps. But the stock moves are just the symptom. What’s actually happening underneath is structural — it’s about the specific mechanisms by which AI agents are attacking SaaS business models.
The best framework for understanding this is Pakodas’ eight disruption theories. They map out how agentic AI is eroding SaaS business model assumptions across multiple attack surfaces at the same time. We’ll run through all eight, then go deeper on the mechanisms that matter most: AI unbundling, the dumb pipe risk, the cost collapse, the deterministic/probabilistic divide, and where value is accumulating in the three-layer agentic stack. For the broader picture of the SaaS reckoning, start there.
What exactly are AI agents — and why do they threaten SaaS differently from previous AI tools?
An AI agent is an autonomous multi-step action system. It perceives its environment, plans a sequence of actions, and executes them end-to-end — without needing human instruction at each step. That’s what separates agents from copilots. Copilots wait for a prompt, return a result, and wait again.
Previous AI tools worked with the SaaS interface. Agents bypass it entirely. They pull data via APIs, execute workflows, and take actions without ever loading a dashboard.
Anthropic’s Claude Cowork launched on January 12, 2026 and added eleven plugins covering legal, finance, marketing, sales, product management, and data analysis — work that previously required a stack of specialised SaaS products. When it launched, LegalZoom‘s stock dropped 20% in a single session. Not because Cowork is better than LegalZoom. Because document generation is now a commodity.
The consequences for SaaS are structural. Per-seat pricing is built on the assumption that human users need logins and dashboards. One AI agent can do the work of many human users — and agents don’t need seats. As Satya Nadella put it: “Business applications are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents.”
What are Pakodas’ eight theories of AI disruption — and how do they map to the SaaS business model?
Pakodas’ eight disruption theories aren’t a menu where only one turns out to be right. They are all happening at the same time, at different speeds across different market segments.
Theory 1 — The Superagent Eats the Interface. AI agents become the primary way humans interact with software — sitting above all apps, talking to their APIs. SaaS tools become back-end services; value creation moves to the agent layer.
Theory 2 — The Great Unbundling. A sales team paying $30,000 per year for Gong can now replicate the core feature with Claude Code in a weekend. Bundle pricing power erodes when users can acquire individual features at near-zero cost.
Theory 3 — The Uncertainty Tax. Even SaaS companies with strong fundamentals are getting repriced downward because markets can’t confidently model what their business looks like in five years. ServiceNow beat Wall Street expectations for the ninth consecutive quarter, raised guidance, and the stock dropped 11%.
Theory 4 — The $0 MVP. Traditional SaaS MVPs cost $500,000 to $1 million to build. AI coding tools collapse that to near-zero. The cost moat protecting SaaS from substitution is gone.
Theory 5 — Compound Engineering. A single developer using AI coding agents can maintain and ship five software products simultaneously. AI-native companies operate at roughly half the headcount of traditional SaaS at equivalent output. The evidence in AI-native growth rates is well documented.
Theory 6 — The Invisible App. Agents don’t need good design; they need good APIs. The SaaS product becomes invisible — a back-end service users never directly interact with.
Theory 7 — The Probabilistic Divide. Probabilistic SaaS — content creation, marketing automation, task management — faces the most acute displacement risk. Deterministic SaaS — payroll, ERP, healthcare records — retains more defensibility.
Theory 8 — The End of the Seat. AI agents do the work that previously required multiple human users. Salesforce is already experimenting with Agentic Enterprise Licence Agreements — flat-fee structures for companies deploying agents at scale.
Theories 2, 4, and 7 are the mechanically richest. The next sections go deeper.
How does AI unbundling work — and why can’t SaaS vendors just add AI to defend against it?
AI unbundling is the process by which AI agents let companies extract individual features from bundled SaaS products and replace them with purpose-built alternatives at near-zero cost.
The SaaS bundle worked because building custom software was expensive. Vendors packaged multiple features together and users paid for the bundle even if they only used part of it. That friction is now gone. A sales team paying $30,000 a year for Gong can replicate the core call analysis feature with Claude Code in a weekend.
Jasper AI is the canonical example. It peaked at roughly $90M ARR in 2023. Then ChatGPT commoditised its core content generation feature, revenue dropped, and both co-founders stepped down. A Retool survey of 817 builders found 35% had already replaced at least one SaaS tool with a custom build, and 78% expect to build more in 2026.
Adding AI features to an existing bundle doesn’t restore the bundle’s value proposition. The real competitive pressure is bundle pricing against à la carte alternatives at zero marginal cost.
What is the “dumb pipe” risk — and which SaaS vendors face it most acutely?
“Business applications are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents.” — Satya Nadella.
Here’s the dumb pipe scenario: an AI agent accesses a SaaS vendor’s data via API and executes all business logic externally. Interface value, workflow value, and intelligence value migrate to the agent layer. The vendor narrows to a data store with no differentiation.
The structural distinction that matters: systems of record hold authoritative, governed data. Systems of engagement are the workflow and UI layers humans use to interact with that data. Systems of engagement are the most vulnerable; systems of record retain defensibility as long as they protect the data moat.
The most exposed vendors: HubSpot (down 51% year-over-year), Monday.com (down 44%), Atlassian at the workflow layer (down 26.9%). Their differentiation is precisely what agents are replacing.
More defensible: Workday (HCM compliance data), Epic (healthcare records), SAP/Oracle ERP. Bain names Procore’s project cost accounting and Medidata’s clinical-trial randomisation as “core strongholds.” The evaluative question is simple: does this vendor’s primary value sit in the interface layer, or in proprietary governed data? How these mechanisms map to the Bain four-scenario framework gives you the full vendor evaluation methodology.
Why has the cost of building software alternatives collapsed?
Traditional SaaS MVPs cost $500,000 to $1 million to build. AI coding tools have collapsed that to near-zero. The mechanism is vibe coding: AI-assisted development using Cursor, Claude Code, and GitHub Copilot that lets a single developer replicate core SaaS functionality in days. The skill threshold has dropped from senior engineering teams to competent solo developers.
Compound engineering extends this further. A single developer with AI coding agents can maintain and ship five products simultaneously.
Cursor crossed $1B ARR in less than 24 months with approximately 300 employees — $3.3M ARR per employee, three to five times more efficient than the best public SaaS companies. Salesforce runs at roughly $800K ARR per employee. The full picture of what these growth trajectories mean for incumbents is worth a look.
The only remaining barriers to entering a SaaS category are proprietary data, compliance requirements that embed the vendor into regulated workflows, and network effects requiring multi-sided marketplace participation. Every SaaS renewal now needs a build assessment.
What is the difference between probabilistic and deterministic SaaS — and which category is more at risk?
The probabilistic vs. deterministic distinction is the single most actionable risk-stratification tool you have for evaluating a vendor portfolio.
A deterministic system produces the same outputs given the same inputs, every time — payroll, invoicing, compliance, clinical trials. Error tolerance is near-zero. Regulatory requirements embed the vendor deeply into your operations.
A probabilistic system is one where “good enough” is acceptable: content generation, marketing automation, project management, customer support routing. AI agents handle these workflows at near-zero marginal cost.
Defensible examples: Workday, SAP, Oracle, Epic, Procore, Medidata — vendors holding governed, regulated data that can’t be casually replicated.
Vulnerable examples: HubSpot, Monday.com, Zendesk, Jasper AI, Atlassian’s workflow layer. ServiceNow reported that AI agents now resolve 90% of IT and 89% of customer support requests autonomously inside its own operations — which is what probabilistic workflow automation looks like at scale.
One important nuance: deterministic is a spectrum. Vendors like Salesforce and Workday have probabilistic workflow layers built on top of their deterministic core data. The workflow UI can be unbundled even if the underlying system of record is defensible. Probabilistic tools get heightened build-vs-buy scrutiny at every renewal. The Bain four-scenario framework maps this directly to portfolio decisions.
Where does value accumulate in the three-layer agentic stack?
Bain’s three-layer agentic stack maps out where economic value is moving as AI agents mature.
Layer 1 — Systems of Record. The data repositories: Workday, Salesforce, SAP, Epic. Their edge is unique data structures, long activity histories, and built-in regulatory logic. They remain valuable as long as they hold proprietary, governed data — and lose that value if they become pure CRUD stores.
Layer 2 — Agent Operating Systems. The middleware that orchestrates actual work: Microsoft’s Azure AI Foundry, Google’s Vertex AI Agent Builder, Amazon Bedrock Agents. These systems plan tasks, remember context, and invoke tools. This is where substantial platform value is currently accumulating — Amazon, Google, and Microsoft are forecast to spend close to $500 billion on AI infrastructure in 2026 alone.
Layer 3 — Outcome Interfaces. How humans consume agent outputs: Teams, Slack, custom apps. Layer 3 is increasingly commoditised — building a custom outcome interface has near-zero cost with vibe coding.
The emerging battleground is the semantic layer between these three levels. Anthropic’s Model Context Protocol (MCP) and Google’s Agent2Agent (A2A) standardise how agents package tool calls, security tokens, and results as they move among layers — and both show the kind of network-effect dynamics where the first standard to achieve broad adoption takes most of the market.
Vendors concentrated in Layer 3 face the most exposure. Vendors with a defensible role in Layer 2 or irreplaceable data in Layer 1 are better positioned. For what this means for your software stack, the implications are practical.
Are AI agents replacing SaaS or augmenting it — and does the distinction matter for evaluating your stack?
Both are happening. Which one applies to a specific tool depends on the probabilistic/deterministic framework from the previous section.
The replacement scenario is real. Klarna consolidated 1,200 SaaS applications into an in-house AI stack. AI handled the workload of approximately 700 customer service agents and revenue per employee grew from $300K to $1.3M. But customer satisfaction declined and the company reversed course, rehiring staff. Full stack replacement carries operational risk.
The augmentation/starvation scenario is the more common near-term mechanism. Menlo Ventures documents enterprise AI spending growing from $1.7 billion in 2023 to $37 billion in 2025 — discretionary IT budget that would have funded SaaS expansion is now flowing to AI tools instead.
And augmentation still erodes per-seat revenue. If AI agents handle 40% of a team’s workflow, that team renews at a reduced seat count. Headcount-driven SaaS growth stalls even without product replacement.
For probabilistic tools, think in replacement terms — build-vs-buy assessment, migration planning. For deterministic tools, think in starvation and renegotiation terms — seat contraction, usage-based pricing conversations, renewal discipline.
The mechanism-level picture is where the analysis starts. For what this means for your software stack, that is the full context.
FAQ
What is the difference between an AI agent and a copilot?
Copilots require human prompting at each step. AI agents are autonomous: they perceive their environment, plan multi-step actions, and execute tasks end-to-end without per-step human guidance. Agents interact with systems directly through APIs — they don’t surface a UI.
What are Pakodas’ eight theories of SaaS disruption?
The eight theories are: (1) The Superagent Eats the Interface, (2) The Great Unbundling, (3) The Uncertainty Tax, (4) The $0 MVP, (5) Compound Engineering, (6) The Invisible App, (7) The Probabilistic Divide, and (8) The End of the Seat. They are additive — multiple theories can attack a single vendor simultaneously.
Which SaaS tools are most at risk from AI agents right now?
Probabilistic SaaS tools — those where output accuracy doesn’t need to be exact — face the most immediate risk. HubSpot, Monday.com, Zendesk, Jasper AI, and Atlassian’s workflow layer are the commonly cited examples. Deterministic tools — payroll, ERP, healthcare records, compliance — retain more defensibility.
What does “dumb pipe” mean in the context of SaaS?
“Dumb pipe” (Satya Nadella’s term) describes the risk scenario where a SaaS vendor becomes a passive data store — accessed by AI agents via API, with all business logic executed in the agent layer. The vendor loses interface differentiation and pricing power, reduced to a CRUD database. Vendors most at risk are those whose primary value is in the workflow and UI layer rather than in proprietary governed data.
Why did Klarna replace Salesforce CRM with an in-house AI stack?
Klarna consolidated 1,200 SaaS applications as part of an AI-first rationalisation. AI handled the workload of approximately 700 employees. Customer satisfaction declined and Klarna reversed course, rehiring staff. Full stack replacement is technically feasible but operationally risky at scale.
What is vibe coding and why does it matter for SaaS buyers?
Vibe coding is AI-assisted software development using Cursor, Claude Code, and GitHub Copilot. A developer can now replicate core SaaS functionality in days at near-zero marginal cost. This changes the build-vs-buy calculus: tools that were cost-prohibitive to build internally are now feasible alternatives, which gives you real renegotiation leverage at renewal.
Is Cursor really a threat to traditional SaaS companies?
Cursor crossed $1B ARR in 24 months with approximately 300 employees — $3.3M ARR per employee, three to five times more efficient than the best public SaaS companies. That growth rate validates Compound Engineering in practice: AI-native companies can reach significant commercial scale with structural cost advantages incumbents can’t easily replicate.
What is MCP (Model Context Protocol) and why does it matter?
MCP is Anthropic’s standard for how AI agents communicate across systems — packaging tool calls, security tokens, and results as they move between applications. Google’s Agent2Agent (A2A) is its counterpart. Both show strong network-effect dynamics — winner takes most — and the outcome will influence which agent OS platforms integrate most easily with existing SaaS stacks.
What is the three-layer agentic stack?
Bain’s three-layer agentic stack: Layer 1 (Systems of Record — Workday, Salesforce, SAP), Layer 2 (Agent Operating Systems — Azure AI Foundry, Vertex AI Agent Builder, Amazon Bedrock Agents), and Layer 3 (Outcome Interfaces — Teams, Slack, custom apps). Value is accumulating most rapidly at Layer 2. Vendors concentrated in Layer 3 face the greatest displacement risk.
Are per-seat SaaS contracts still appropriate in an agentic world?
Per-seat pricing is increasingly misaligned with AI agent economics. Agents don’t need logins or user licences. Gartner predicts over 30% of enterprise SaaS solutions will incorporate outcome-based pricing components. Salesforce’s Agentforce already offers Flex Credits at $0.10 per standard action. Ask your vendors about usage-based and outcome-based alternatives at your next renewal.
What is Pakodas’ “Uncertainty Tax” — and why does it matter?
The Uncertainty Tax (Theory 3): SaaS companies with strong fundamentals get repriced downward because markets can’t model their five-year trajectory under AI disruption. Depressed valuations affect vendor stability — financially pressured vendors may cut R&D, accelerate price increases, or become acquisition targets.
Does the probabilistic/deterministic distinction mean ERP and payroll vendors are completely safe?
No — it means they are more defensible, not invulnerable. Deterministic SaaS vendors still face the starvation scenario — budget contraction, seat reduction. Vendors like Workday and SAP have probabilistic workflow layers built on top of their deterministic core data. Those workflow layers remain vulnerable even if the underlying data store is defensible.
