1️⃣ The AI Application Layer Moment
For years, the AI story has been told in terms of scale — bigger models, larger datasets, more GPUs.
But in 2025, that center of gravity is shifting.
According to both the a16z AI Application Spending Report and Mercury’s Startup Economics Report 2025, the real momentum is moving upward in the stack, from training to using.
Money, hiring, and creative energy are flowing into AI-driven workflows — systems that understand business context and quietly augment daily work.
The new competition isn’t for tokens or teraflops but for relevance: which platform can make a company’s data, decisions, and language come alive the fastest.
We are entering what I call the Application Layer Era — where intelligence meets context, and context becomes capital.
And among all the domains competing for attention, one is emerging as the nervous system of them all: enterprise knowledge.
2️⃣ The Ten Domains of Applied AI
| # | Domain | Example Companies / Products |
|---|---|---|
| 1 | Foundational Models & Assistants | OpenAI · Anthropic · Perplexity AI · Manus |
| 2 | Enterprise Knowledge & Search | Glean · Notion AI · Rewind · Microsoft 365 Copilot |
| 3 | Meeting Notes & Comms Capture | Otter.ai · Fathom · Fixer |
| 4 | Creative & Content Generation | Canva · PhotoRoom · Descript · ElevenLabs |
| 5 | Sales & Go-to-Market Automation | Customer.io · Apollo · Clay |
| 6 | Customer Support | Crisp · Intercom AI · Ada |
| 7 | Legal Tech | Crosby · Harvey |
| 8 | HR & Recruiting | Ashby · Paradox |
| 9 | DevTools & Vibe Coding | Replit · Cursor · GitHub Copilot · Cognition Labs |
| 10 | IT Ops / Finance / Compliance | Ramp · Pulley · Vanta · Mercury |
Each category is maturing along its own trajectory, but all are gravitating toward a shared outcome: automation through understanding.
Roughly 60 % of AI startup spend is now sitting in this layer, and nearly 70 % of adoption is starting bottom-up — a single enthusiast, a team experiment, then org-wide scale.
Still, beneath every workflow, every creative suite, and every agent platform lies one core challenge: making enterprise knowledge findable, usable, and trustworthy.
3️⃣ From Assistants to App-Gen Platforms
When a16z describes Specialized App-Gen Platforms, they are capturing a movement that is redefining the purpose of AI in business.
If the first wave of AI was about training models and the second about prompting them, this third wave is about building on them — composable, domain-specific systems that let non-ML experts generate applications instead of answers.
Foundation models gave us cognition. Assistants gave us conversation. App-Gen Platforms are giving us construction — environments where finance teams are assembling reconciliation agents and HR leaders are creating onboarding copilots without touching a single pipeline file.
What’s emerging is a shift from prompt engineering to context engineering — a discipline that treats workflow state, user intent, and domain constraints as first-class citizens.
This is the quiet revolution that is powering the new Application Layer.
And the first arena where this approach is being tested most vividly is Enterprise Knowledge & Search — because nothing reveals context better than the way a company stores its memory.
4️⃣ Enterprise Knowledge & Search
a. Why Knowledge Search Matters
Every enterprise is both a generator and a graveyard of information.
The insights are already there — buried in email threads, Slack channels, and SharePoint archives.
AI’s highest ROI today is emerging from retrievability, not originality.
“Knowledge management is becoming cognition management.”
The question that every enterprise team is asking right now is no longer Can AI write? but Can AI remember responsibly?
b. Microsoft’s Copilot Architecture — The Closed Loop
Microsoft is embedding an agentic layer directly into its productivity and developer ecosystems. Microsoft’s play depends on its Graph + Tenant Isolation architecture:
- Microsoft Graph → Unified data model across Outlook, SharePoint, Teams, OneDrive, Dynamics, Azure AD.
- Every enterprise tenant’s data remains isolated under its own ID boundary.
- The Copilot orchestration layer uses Graph permissions to retrieve context, never training on customer data.
- Bring-Your-Own-Data (BYOD) connectors → Microsoft Search + Graph Connectors let companies plug in Salesforce, ServiceNow, or Confluence data.
- Retrieval works via vectorized embeddings + fine-grained ACLs.
- No cross-tenant learning; responses are contextual but ephemeral.
- Compliance & Governance → Integrated Purview, eDiscovery, DLP, and audit logging — a huge differentiator for regulated sectors.
- Model Layer – GPT-4 Turbo via Azure OpenAI, providing shared reasoning inside a governed sandbox.
- Orchestration Layer – an internal planner that is routing requests through Microsoft Graph to Outlook, SharePoint, Teams, and Dynamics.
- Application Layer – Copilots are appearing natively inside Word, Excel, Outlook, and VS Code.
Graph is serving as the connective tissue linking identity, permissions, and data lineage.
This architecture is turning Microsoft’s stack into a private knowledge loop — a self-contained ecosystem where data flows from query → reasoning → action → learning without leaving the tenant boundary.
Net effect: Microsoft sells trust and control as the moat — not just intelligence.
c. The Open Ecosystem Countermovement
Meanwhile, a counter-movement is unfolding around composability and portability.
Startups like Glean and Notion AI are building cross-SaaS knowledge graphs that follow the user instead of the vendor.
| Player | Core Idea | Differentiator |
|---|---|---|
| Glean | Unified permission-aware search | Deep cross-SaaS connectors (Slack, Google Drive, Jira, Salesforce) |
| Notion AI | Knowledge canvas + RAG summaries | Blends editing + reasoning in one workspace |
| Rewind | Personal memory engine | Local privacy; cross-app time-travel search |
| Cohere North | Custom retrieval agents | Multi-tenant LLM hosting with enterprise connectors |
| Hebbia | Research-grade semantic search | Optimized for financial & analyst workflows |
These systems are betting that the future of enterprise intelligence is context mobility — intelligence that travels where the user works today.
d. Data Sovereignty & Retrieval Strategy
Two architectural philosophies are defining the market:
| Model | Architecture | Trade-Off |
|---|---|---|
| Closed Loop (Microsoft Graph) | Unified internal ecosystem (Unified model → data graph → applications → security/compliance) | High trust / less extensible |
| Open Composable Loop (Cross-SaaS RAG) | Distributed connectors + vector DBs (Independent retrieval + open connectors + multi-model choice (Cohere, Anthropic, OpenAI, Mistral) | Flexible / harder governance |
Closed loops are thriving in regulated industries; open loops are accelerating in startups and creative teams.
Both approaches are converging as federated governance and encrypted retrieval become standard.
5️⃣ Where the Layers Collide — Convergence
We are now seeing boundaries between model, framework, and application starting to converge. Every layer of the AI stack is moving closer to the next:
- Models are learning to remember and use tools natively.
- Frameworks like LangGraph and Strands are becoming model-aware runtimes.
- Applications are embedding domain LLMs directly in their runtime.
In the near future, the typical enterprise knowledge system will look less like a search bar and more like a continuous reasoning agent with live connectors and domain-specific LLMs.
The stack that once looked like a pyramid is flattening into a lattice of feedback loops.
Foundation models are absorbing orchestration; frameworks are melting into SDKs; applications are hosting their own LLMs.
Development is starting to look less like API calls and more like composing context graphs that models are learning from in real time.
The stack is not collapsing in chaos — it is self-organizing.
This convergence isn’t happening in isolation.
Across the consulting and enterprise software world, partnerships and acquisitions are revealing where the real strategic battles are being fought — not over who trains the biggest model, but over who owns the workflow rails that enterprise AI will run on.
The CB Insights graphic below maps this new reality: global consultancies like Accenture, Deloitte, and KPMG are positioning themselves at the center of the AI stack through partnerships with Microsoft, AWS, Salesforce, and emerging agentic startups.
When frameworks vanish into models and models hide inside applications, users won’t ask what LLM a tool runs on — they’ll ask which domain truly understands them.
And as the layers finally converge, context will become the new competitive cloud.
My experiments with AI 
Leave a comment