Generic selectors
Exact matches only
Search in title
Search in content
Post Type Selectors
About the Author
About the Author
About the Author

Software Ate the World. AI Isn’t Devouring It. It’s Sharing the Table. The High-Stakes Fight for Enterprise Power and Value.

Being Human for a Better Tomorrow in the Age of AI

  • February 13, 2026
  • 66th edition
Gemini Generated Image dkg2jzdkg2jzdkg2 scaled


Marc Andreessen’s famous 2011 observation that software will eat the world has played out. In Feb 2025, in the 35th Mindvista, I hypothesised that AI will now eat software. There was early evidence from legacy software transforming with AI like Autodesk, HubSpot, Figma and also a new breed of AI native software like Replit, Runway, Jasper integrating generative capabilities and workflows into software expanding the value. It also felt metaphorically as some kind of a karmic payback to software as a fallout of technology progress.

One year on, a lot has changed. AI capabilities and efficiency have increased. Enterprises have started deploying AI in production and a few leaders are rolling out AI enterprise wide across processes and employees.

A new class of Agentic AI goes beyond intelligence on demand to autonomous planning and execution of action. Claude CoWork, OpenClawd, Qwen have created the tools for Agentic AI deployment and a new class of AI native apps like Clay in marketing, Serval in IT service management and Devin in software engineering are delivering outcomes.

With the advent of these tools, the volume of the debate has increased enormously. “SaaS is dead.” “AI kills software.” “SaaSpocalypse.” These claims fill financial media and technology commentary in early 2026, driven partly by real valuation pressure on enterprise software companies riggered by debate around Claude Cowork launch and partly by genuine anxiety about what agentic AI means for the structure of enterprise technology. The debate, in most places, is binary and reflects a deep divide. As Jamin Ball puts it, the software bears feel that software is mature, lacks the focus to adapt, and will be vibe coded away — while software bulls say AI vibe coded software does not work and cannot be trusted by business.

Is it time to revisit the hypothesis that AI will kill software? Will software turn bearish, tamed by AI bulls?

In this edition we revisit the hypothesis and take a fresh examination along the following lines:

1. Exploring AI native applications with agentic or agent like capabilities — where are they genuinely breaking out?


2. Examining legacy and SaaS companies — what is actually changing inside enterprise software and what can be learnt from early successes and struggles?


3. A buy side view — what does the application stack look like for AI native businesses? Are they using legacy systems or vibe coded applications? How are enterprises like JPMorgan and Goldman Sachs with capital and conviction moving on enterprise wide AI? What modifications are they making in their application stack?

When we synthesise these different views, what emerges is a more nuanced picture driven by a new emergent enterprise application architecture pattern.

AI will not eat the software lunch — but it has pulled up a chair to the same table. And quietly, almost unnoticed, AI Services — a third guest — has joined, and may turn out to have the largest appetite of all.

Enterprise AI Architecture: The Three-Layer Model Reshaping Enterprise Software

The enterprise technology architecture is broadly restructuring into three distinct layers: Systems of Record, Systems of Execution, and Systems of Context,Control and Observability. They play the role of the Ledger, the Agent, and the Fabric that holds the two together. AI native vendors and software firms must choose their play in one of these categories — and value is migrating between them in ways that reward clarity of purpose and punish confusion of role.

Systems of Record: Why Enterprise Software Still Owns Business Truth

This is the bottom layer and it is not going anywhere. The Systems of Record are where organisational truth lives — financial accounts, employee records, customer contracts, compliance trails. Workday holds HR truth. NetSuite and Oracle hold financial truth. Salesforce holds customer relationship truth. These Systems persist not because of switching costs alone but because they are bound to legal accountability. You cannot satisfy a SOX audit with an agent’s working notes. You cannot IPO on a vibe coded database. The ledger must exist, be structured, and be defensible.

What is changing is the interaction model. AI agents increasingly bypass the user interface, the menu clicking, the form filling, and the manual entry but they must write back to the record to make their work real and auditable. The Systems of Record is holding the truth that is intact and nonnegotiable.

Systems of Execution: How Agentic AI Is Transforming Enterprise Software

This is where the genuinely new story is being written. AI native applications in this layer are not only copilots or chat assistants bolted onto existing software. They aim to be functional actors that own a bounded unit of work end to end — assembling context, planning, acting, and closing the loop, while leaving truth keeping to the systems of Record above. This is illustrated with few examples.

Clay operates as an AI driven go to market engineer: it gathers intelligence, enriches prospect data, sequences outreach, and writes results back into Salesforce. It does not replace Salesforce — it makes a human sales rep’s entire research and outreach workflow autonomous. Clay is reportedly valued at $1.3B on an estimated revenue $50-$70M.

Sierra operates as an AI-driven conversational agent for the enterprise: it integrates directly with internal systems (ERPs, CRMs, and order management), reasons through complex customer intent, and executes real-world transactions like processing returns or modifying subscriptions. It does not replace the customer service platform (like Zendesk or Salesforce Service Cloud)—it transforms it from a passive ticketing queue into an autonomous resolution engine that handles end-to-end customer needs without human intervention. Sierra is reportedly valued at $4.5B on a $100M+ estimated revenue .

Serval operates as an AI-driven IT operations engineer: it autonomously triages, diagnoses, and resolves complex technical tickets by navigating an enterprise’s internal infrastructure. It does not replace the service desk infrastructure—it acts as the autonomous execution layer that logs into systems, resets configurations, and clears bottlenecks. By the time a human lead looks at ServiceNow, the ticket is already marked “resolved” with a full audit trail of the actions taken.Serval is a unicorn valued at estimated $1 billion.

Devin, from Cognition, operates as an autonomous AI software engineer: it handles the entire development lifecycle from initial specification to tested, deployed code. By navigating terminal environments, browsers, and editors, it functions as an independent teammate that operates upstream of GitHub governance, eventually writing its output back into version control for human review. (Devin became the first AI to pass the threshold of “long-horizon” reasoning required to solve real-world engineering tasks autonomously.). Devin has also became a unicorn with a $10 billion valuation.

These breakouts share a structural pattern. They succeed when scope is bounded, outcomes are measurable. They do not try to own the world model or the ledger. They execute within defined boundaries and pass truth upward. The systems of Execution is growing fast — but it grows fastest when it respects the other two layers by not doing too much and compete with systems of context and control or replace systems of record.

AI native systems of execution have not only gone from chats to outcomes. They have elevated their delivery and pricing models with outcome based pricing capturing more value.

Enterprise AI Fabric: Systems of Context, Control & Observability

This is the most strategically important layer in the current moment.

AI agents are semantically capable but causally blind. They understand language, generate plans, and execute discrete actions. What they cannot do, as research from Skyfall AI’s World of Workflows benchmark shows, is model the hidden dynamics of enterprise systems under conditions of partial observability — where workflow dependencies, approval chains, and system states are not fully visible to the model. When an agent updates a record in Salesforce, it cannot automatically know that this triggers a downstream approval workflow in ServiceNow, which feeds a compliance report in Oracle, which has a 48-hour review cycle that a human manager must sign. The agent sees the surface. It cannot see the plumbing. In constrained enterprise tasks, current frontier models succeed at only 2 to 14 percent of attempts when operating without contextual grounding. (Source: Skyfall AI / World of Workflows, 2025.) Silent failures, actions that appear completed but are causally broken are the specific risk.

Beyond causal blindness, agents also struggle with human realities — ambiguous definitions, exception handling, and explainability demands. These are not edge cases; they are everyday enterprise conditions.

The Systems of Context, Control, and Observability solve this problem in three complementary ways.

  • On the context side, it supplies the semantic world model—data relationships, business logic, system interdependencies—that agents need to reason and act without triggering downstream chaos.
  • On the control side, it enforces the guardrails, orchestration rules, and governance structures that enterprises demand before entrusting automated systems with consequential work.
  • And on the observability side, it provides continuous visibility into what the agents are actually doing—tracking performance, detecting drift, surfacing silent failures, and explaining decisions—so humans (and regulators) can trust the system without flying blind.

 

Three companies illustrate the context, control and observability layer from different angles.

Databricks is an example on context . Its Unity Catalog and Lakebase architecture create a governed semantic layer across fragmented data sources — the unified intelligence fabric that allows agents to reason about enterprise data without crashing into what practitioners call the “data swamp.” Databricks reached a $5.4 billion annualised revenue run rate in early 2026, growing at 65 percent year on year. (Source: Databricks, February 2026.) Its neutrality — it sits across all systems of records rather than competing with them — is the strategic choice that makes it trusted.

ServiceNow is an example on the control. Its Configuration Management Database maps enterprise systems. dependencies, and permitted workflows — the causal world model that agents currently lack. Bill McDermott’s positioning of ServiceNow as the “control tower for business reinvention. NowAssist and the agent platform insert AI into existing workflows without bypassing the control structure. ServiceNow reported $600 million in AI Annual Contract Value in its most recent results. (Source: ServiceNow earnings, 2025.)

Datadog provides an example of observability. You cannot control what you cannot see. As enterprise AI stacks grow in complexity — multiple agents, multiple data sources, multiple SORs interacting — knowing when something is failing, drifting, or behaving unexpectedly is nonnegotiable. Datadog’s LLM Observability tools track model performance and drift in production. Its Bits AI SRE Agent autonomously investigates alerts and surfaces root causes before engineers are paged. Datadog posted $3.43 billion in fullyear 2025 revenue, 28 percent growth, and is guiding to $4.06–4.10 billion in 2026. (Source: Datadog Q4 2025 earnings.) The growth reflects enterprises discovering that deploying AI without observability is flying blind.

Context, Control, Observability the enterprises that build this fabric well are the ones whose AI investments compound.

 

Enterprise AI Transformation: Two Architecture Patterns Leading Enterprises Are Using

What does this architecture look like in practice? The evidence points to two distinct implementation patterns — one incremental, one more radical — and both are instructive.

Goldman Sachs and JPMorgan represent the incremental pattern, pursued at scale. JPMorgan has deployed over 600 AI use cases across lending, compliance, fraud detection, and customer service, with a reported $2 billion annual AI investment. (Source: JPMorgan Chase, 2025.) They retain the Systems of Record — the core banking ledgers, the compliance infrastructure — and build a proprietary systems of Context and Control on top. The agents run on top of the fabric; the fabric runs on top of the ledger.

Goldman Sachs committed over $1 billion to AI transformation and embedded Anthropic engineers directly within its technology teams a build with us model that treats the intelligence fabric as a co-created strategic asset rather than a vendor purchase. Early Claude based agent deployments in reconciliation, KYC onboarding, and due diligence workflows are reporting 20 to 30 percent productivity gains in targeted processes.

Anthropic and OpenAI are new age but large enterprises with annualised revenue of $14B, and $20B respectively and growing at 10x and 3x respectively. Yet despite their size and growth they represent a more radical pattern compared to traditional enterprises.

Based on limited public information it seems both companies use standard Systems of Record internally Workday for HR, Salesforce or HubSpot for go to market, NetSuite for finance. (Sources: Company job postings, 2025.) Neither replaced these with AI agents. What they built instead was an aggressive custom intelligence layer on top, so employees rarely interact with legacy UIs directly.

The pattern is consistent: treat the ledger as utility infrastructure, build rest of the layers as a competitive asset a conservative for systems record, radical on the Fabric and Execution.

This definite but minimalist use of systems of records and custom build the Fabric and Execution may point to the evolution of new archetype of enterprise applications and technology capability.

Enterprise AI Business Models: Choosing the Right Layer to Win

When you have clarity on what is the play and is backed by agency execution align the results are evident and where they don’t, the costs are high.

Clay, Serval,Devin, Sierra and such systems of execution succeed because they made a precise architectural choice: own the execution loop completely, respect the systems of record as the truth layer, and stay out of the fabric business. Their growth is the reward for that clarity. Databricks and ServiceNow, Data Dog are winning because they identified the Context and Control and Observability layer (fFabric) as the strategic prize before most competitors. Their revenue growth is the reward for building depth in the right place at the right time.

Microsoft and Salesforce are the cautionary evidence. Microsoft Copilot reached Fortune 500 penetration on licence sales but daily active usage remains well below 10 percent of licensed employees. (Source: as reported in Multiple enterprise surveys, 2025.) Salesforce recognised the problem, killed Einstein Copilot, and pivoted to Agent Force a move in the right direction, but Agent Force remains bound to the quality of the underlying CRM data, which in most enterprises is not clean and complete enough to support reliable agent execution. Both companies are attempting to span all three layers simultaneously and paying the coordination tax. The lesson is not that they will fail both have the resources and customer relationships to course correct but that confusion of architectural role is causing concern.

AI Services Enterprise: The Missing Layer in Enterprise AI Transformation

Realising the three layer architecture needs not just software or AI vendors but also AI Services the human capability required to bridge the gap between what the architecture enables and what enterprises can actually build and govern.

Building  Systems of Context, Control and Observability is not a product purchase. It requires understanding the business logic embedded in legacy systems accumulated over decades, remediating years of data quality debt, designing governance frameworks that satisfy compliance and security requirements, and managing the human and organisational change that accompanies any significant shift in how work gets done.

The Goldman Anthropic embedded engineering model is a services model. JPMorgan’s internal platform build is a services investment running into the billions. The “build with us” pattern that leading enterprises are adopting — whether with vendors, systems integrators, or through aggressive internal hiring — is the services story playing out in real time.

While the demand of AI services is real, the business model impact is serious. The deflation of effort and service pricing and change from time and material or fixed cost to focus on outcomes create new challenges for sustaining the growth and margins. Service providers have the benefit of customer access but their traditional strengths of large employee base, capital and legacy application knowhow and pricing models need to reconfigured for IP enabled , small, nimble and talented team delivering outcomes.

Future of Enterprise Software: Who Wins in the Age of AI?

For AI-native systems of execution builders : The next competitive frontier is depth, not breadth. Generalist agents are stalling precisely because they cannot navigate enterprise complexity. The breakouts of the past year like Clay, Sierra, Serval and Devin, succeeded by going deep into a single functional domain and owning the complete execution loop, including the write-back to the systems of Record. The builders who do this cleanly, making the complexity invisible to the user and the outcome undeniable to the buyer, become systems of execution.They own outcome and get the valuations proportional to growth.

For Systems of Context, Control and Observability: Fabric builders hold the keys to making AI reliable, consistent, and better, turning probabilistic tools into dependable enterprise assets. Lead the fight against the reliability slowdown: deliver seamless continuity across model upgrades (no breaking changes or forced resets), reduce hallucinations and reasoning drift through richer semantic/causal grounding, enforce proactive observability to surface silent failures early, and minimize verification debt and cognitive load so humans spend less time doubting and fixing.

Measure success as output quality per unit of input effort for higher reliability and consistency for lower effort skepticism and cost.

For systems of record vendors: The market’s reassessment of enterprise software valuations is not primarily about AI replacing your systems. It is about the engagement layer, the user interface, the “where work happens” function of your platform being eroded. The survival strategy — remaining the trusted ledger — is secure. The growth strategy requires something harder: opening your data models as semantic foundations for the intelligence layer, becoming the bedrock that agents and fabric platforms rely on rather than competing with them for the execution layer. Workday’s agentic workflow investments, SAP’s Business AI embedding, Oracle’s Fusion AI expansion — these are the right directional moves. The question is pace and depth of commitment that can compete with fast movers.

For IT services firms: The “AI staff augmentation” model  providing humans who use AI tools is a transitional business. The structural opportunity is owning build for systems of execution or the systems of the Context and Control and deliver as a managed outcome. This means underwriting the reliability of the intelligence layer, not just implementing it. The commercial model shifts from selling implementation hours to selling governed intelligence infrastructure. The firms that make this transition, moving from integrator to intelligence architect will capture the margin that currently belongs to no one.

 

The Future of Enterprise AI Architecture

The karmic cycle described in Edition 35 did not complete as anticipated. Software was not devoured by AI but the cycle is re-layering application architecture, creating new strategic plays that impact future of business and employees.

Lunch is on the table. AI natives, software, and services are all sitting. Now it is for business and technology leaders to decide the menu, the offering, and the terms.

Love to hear your reactions, thoughts and comments.

Best wishes,

Sidebar: References

MindVista Series

  1. MindVista Edition 35: Software Ate the World. Now AI Is Devouring Software (February 2025)
  2. MindVista Edition 56: 2026 First Movers—Enterprise AI to Enterprise-Wide AI (Mid-2025)
  3. MindVista Edition 64: AriaVista—My AI Ally for Exploration: Everyone’s Using AI, But Few Can Rely on It


Blogs & Research

  1. Rohan Paul: SaaSpocalypse Hits World Capital (2026)
  2. Skyfall AI: World of Workflows Benchmark – Bridging AI Safety Gap in Enterprises via World Models (2025)


X Posts & Announcements

  1. Jamin Ball: Travel Analogy Thread on Systems of Record and New AI Layers (December 19, 2025)
  2. Jamin Ball: Follow-up Thread on Enterprise AI (2026)
  3. Jared Sleeper: ServiceNow Teardown and AI ACV Analysis (February 9, 2026)
  4. Zoomer: Bill McDermott Email on ServiceNow as the Enterprise Control Tower (February 9, 2026)
  5. Ali Ghodsi: Databricks Results Commentary (February 9, 2026)


Other Sources

  1. Goldman Sachs / Anthropic Partnership (Reuters, 2025)
  2. Datadog Q4 2025 Earnings (Datadog Investor Relations, February 2026)
  3. JPMorgan AI Investment and Deployment (JPMorgan Chase Disclosures, 2025)
  4. Anthropic ($14B ARR, 10× Growth) & OpenAI ($20B ARR, 3× Growth) (Company estimates, 2026)

II. Questioning / Asking

Good conversations flow from well‑sequenced questions—topical, simple, coherent, cohesive.
LLM Conversation Example 1
Q: What are empirical judgments?

A: Empirical judgments are based on observation, experience, or experimentation.
Q: What are moral judgments?

A: Moral judgments are based on ethical principles and values.
Share:

Leave a Reply

Your email address will not be published. Required fields are marked *