AI Absorbing the Work Stack: Layered Automation Before AGI

Research

Metaflow AI

Last Updated on

Jan 16, 2026

Build Your 1st AI Agent

At least 10X Lower Cost

Fastest way to automate Growth

Free Forever

Build Your 1st AI Agent

At least 10X Lower Cost

Fastest way to automate Growth

Free Forever

Over the next 2–3 years, artificial intelligence will permeate multiple layers of the work stack – especially in software development and marketing – not as a singular AGI overlord, but as myriad task-focused AIs woven into workflows. Instead of a sudden leap to general-purpose superintelligence, we’re seeing a gradual “sponge effect,” with AI soaking up manual workflows into software and agentic environments. This deep-dive examines how emerging AI workbenches, embedded SaaS agents, and domain-specific tools are transforming work, and what that means for companies and professionals.

AI-Infused Workbenches for Developers

In the software realm, AI is moving from a handy autocomplete to an integrated development partner. Early tools like GitHub Copilot (launched 2022) began as in-editor code autocomplete, reportedly boosting productivity up to 55% . But today’s AI workbenches go much further, enabling developers to go “from idea to code to software all in natural language.”

GitHub Copilot Workspace is emblematic of this evolution. Announced in April 2024, it’s described as a “Copilot-native developer environment” where developers can brainstorm, plan, build, test, and run code using natural language across an entire project . Rather than just suggesting one line at a time, Copilot Workspace uses multiple AI agents to handle tasks from GitHub issue triage to writing code, with the human in full control . Developers can start with a plain-language task (e.g. a feature request or bug report), get a step-by-step plan drafted by the AI (based on the repo’s code and context), edit or approve that plan, and then generate and test code – all within one environment . Importantly, everything is editable and under developer oversight, aligning with the vision that AI should “deliver – not replace – developer creativity” . This workbench lowers the barrier to entry for new coders while empowering experienced devs to work at a higher level of abstraction (more like “systems thinkers” than syntax tinkerers) .

Another example is Cursor, an AI-first IDE built on VS Code. Cursor augments the familiar coding experience with AI-driven workflow tools: context-aware chat and code completion that understand your entire repo, persistent project memory (“Rules”), and the ability to spin up long-running agents that plan and execute multi-step coding tasks . In the latest version (Cursor 2.0), the interface itself has become agent-centric – developers can deploy a “small team of agents” in parallel, each working on a different task (one refactoring, one writing tests, etc.), with their conversations and code diffs presented like reviewable pull requests . This reflects a broader trend: treating AI not as a single magic box, but as multiple specialized collaborators. Cursor even introduced its own optimized coding model (“Composer”) to accelerate these workflows . By combining VS Code’s ecosystem with planning UIs, browser controls, and background agents, such AI workbenches turn coding into a higher-level collaborative process – where much of the boilerplate and rote work is handled by AI under human guidance .

Anthropic’s Claude Workbench similarly provides an environment for prompt engineering, coding, and agent execution. Claude 2, released in 2023, brought a massive 100K token context window – roughly 75,000 words – meaning it can ingest hundreds of pages of code or documentation in one go . In practice, Claude can digest an entire codebase or book in under a minute and answer detailed questions or make edits that require synthesizing information across that huge context . This effectively gives developers a near-eidetic memory assistant for large projects. Claude’s Workbench (available through the Claude Console) allows developers to test prompts and chain tasks, taking advantage of Claude’s long memory and advanced reasoning. For example, a user can drop a full API spec or multiple files into Claude and have it write new code or documentation consistent with all that context – something that previously required human developers to painstakingly read and remember numerous documents.

Historically, these advances build on a trajectory from simple autocomplete (e.g. IntelliSense) to pair-programming bots (Copilot) and now to AI orchestrators in the IDE. Just a few years ago, “AI coding assistant” meant quick suggestions; now it means a multi-agent system embedded in your editor, capable of planning entire features and managing context beyond what any human could hold in mind. We see AI steadily absorbing the lower-level labor of coding – setting up project boilerplate, writing routine functions, generating tests, running linters – allowing humans to focus more on design and complex logic. The creative control remains with developers, but the execution stack underneath is increasingly handled by AI agents.

Embedded AI Agents in Software and Marketing Platforms

Beyond developer tools, AI is being embedded as agents within the software that knowledge workers already use, especially in marketing, sales, and customer service. Large SaaS platforms are integrating conversational and autonomous agents directly into their user experience – essentially giving every user a personalized AI assistant that understands their business data and workflows.

Salesforce’s Einstein Copilot is a prime example. Announced at Dreamforce 2023, Einstein Copilot is a generative AI assistant built into every Salesforce application (Sales Cloud, Service Cloud, Marketing, etc.), working alongside users in their flow of work . A salesperson, for instance, can ask in natural language for an account research summary, and Einstein Copilot will pull from Salesforce Data Cloud (which aggregates customer data, emails, call transcripts, Slack messages, and more) to give a contextual answer . Critically, it’s grounded in secure proprietary data – by design, Copilot’s answers are “relevant and trustworthy” because they use the company’s own CRM data and content as context . This sidesteps one of the biggest barriers to AI adoption in enterprises (hallucinations and data privacy) by anchoring the AI to first-party data.

Salesforce also introduced Einstein Copilot Studio, a toolkit for organizations to customize their AI agents with specific prompts, skills, and even bring-your-own models . This allows, for example, tailoring an agent to automate a particular business process – like an insurance claim triage or a marketing content generator – all within the Salesforce ecosystem. Notably, Einstein Copilot can not only answer queries but proactively suggest actions: e.g. after a sales call, it might recommend an action plan or draft a follow-up email . In customer service, it can auto-summarize cases and even take actions like creating knowledge base articles from resolved issues . All of this happens inside the familiar Salesforce interface (often as a side-panel conversational UI), rather than requiring users to hop to a separate AI app . By deeply embedding AI, Salesforce is essentially turning its CRM into an intelligent co-worker that augments each role – whether it’s writing code for developers on the platform, generating marketing segments, or assisting support agents .

Similarly, HubSpot has gone all-in on embedded AI for marketing and sales. In early 2023, HubSpot launched ChatSpot, a chat-based natural language UI for its CRM . This allowed users to query and command their HubSpot data with simple prompts – for example, “Show me the new contacts added last week” or “Draft a blog post outline on topic X using our top-performing keywords”. ChatSpot was essentially an AI concierge that could fetch analytics, generate content, or update records by understanding the user’s intent in plain English . This innovation meant even non-technical users could tap into the wealth of data (marketing stats, sales pipeline, website metrics) that HubSpot contains, without fiddling with reports or dashboards – “company data becomes useful using natural prompts” as one observer noted . It was a recognition that a huge amount of value in CRMs is locked behind complex UIs, and AI can unlock it by acting as an intermediary between humans and data.

Fast-forward to 2025, and HubSpot’s AI has evolved into Breeze, a suite of AI features and “Breeze Agents” that act as specialized team members. Breeze Agents are described as “AI-powered specialists that extend your team’s marketing, sales, and service capabilities” . In practice, these are domain-specific agents for tasks like content creation, prospecting, customer support, social media scheduling, and more. For example, a marketing team might use a Content Breeze Agent to generate and A/B test copy, while a sales team uses a Prospecting Agent to research leads and draft outreach emails. These agents work within HubSpot, automating workflows “from planning to social media, sales prospecting, customer service, and more” . The platform provides a Breeze marketplace where users can choose which agents they need and then configure them in-app . By capturing entire workflows (not just single tasks), HubSpot aims to make these AI agents true virtual employees. They are integrated with the HubSpot CRM and tools, so they can take actions (e.g. update a contact record, schedule a campaign) rather than simply providing advice. HubSpot’s positioning here is clear: increase productivity “without expanding your staff” by letting AI handle the grunt work of growth and customer management .

The common theme with Salesforce, HubSpot, and others (e.g. Microsoft’s Copilots for Office, Adobe’s Firefly in Creative Cloud) is that AI is being built into the existing UX and dataflows of major software platforms. This embedded approach has several advantages:

Contextual Awareness: Because these AIs sit on top of first-party data stores, they can use a company’s own data as context for generation and decisions. For example, Einstein Copilot can draft an email in the company’s tone and with accurate customer info pulled from CRM records . This is far more powerful than a standalone AI chatbot that lacks access to proprietary data. It also grounds the AI in truth: Salesforce’s Trust Layer ensures Copilot’s answers are grounded in the customer’s secure data to avoid hallucinations .
Flow of Work Integration: Users don’t have to leave their primary tools to leverage AI. The AI assistant is right where work happens, whether it’s a chat sidebar in a CRM or a slash command in a design tool. Slack’s upcoming Slack AI works this way too – summarizing channel discussions or drafting updates directly in Slack threads. This minimizes friction and increases adoption: if asking an AI is as easy as clicking a familiar button or typing a query where you already store information, people will use it more often.
Role-Specific Agentic Behavior: These embedded AIs are not one-size-fits-all ChatGPTs; they are being tuned into job roles and domains. A marketing-focused AI knows about SEO, social media metrics, and branding guidelines. A sales-focused AI knows about lead qualification, CRM objects, and so on. By narrowing the scope, vendors make the AI more trustworthy and useful for specific tasks. The agent can even take autonomous steps within its domain (like logging a call, or creating a draft campaign) rather than just advising. In essence, every professional might soon have a specialized AI colleague integrated into their software – a Sales Co-Pilot, a Marketing Planner, a Customer Support triager, etc.

From a historical viewpoint, this trend parallels earlier technology waves in enterprise software. Just as CRMs and ERPs gradually automated clerical tasks (data entry, basic analytics) and marketing automation took over manual email campaign steps, now LLM-based agents are set to absorb the next tier of work – generating content, interpreting data, and orchestrating multi-step processes on behalf of users. It’s a step-change in that the AI can handle unstructured inputs (like a conversation or a vague request) and produce structured outputs or actions. We’re essentially adding a new “AI layer” atop the SaaS stack that wasn’t possible before advanced language models.

Domain-Specific Frameworks and Tools: From Design to Customer Support

Not all AI integration is happening through big platforms. There’s a flourishing ecosystem of domain-specific AI tools and frameworks that target specific professions and workflows – from design and content creation to coding, customer support, and beyond. These are often specialized startups or features within products that use AI to supercharge a particular domain.

Take Figma for example, a leading design tool. In mid-2024 at their Config conference, Figma announced a slate of generative AI features integrated into its design platform. One headline feature allows designers to generate UI designs from a text prompt – essentially “prompt to prototype.” In a demo, Figma’s CPO showed how typing a description of a restaurant app instantly produced a reasonable mockup with menus, tabs, and even appropriate icons in place . In another example, a prompt for a recipe page yielded a layout complete with an AI-generated image of a cookie . While these first drafts are generic, they give designers a starting point that can be immediately tweaked. Figma AI can also automatically connect design frames into a clickable prototype and suggest text content, saving hours that would be spent on rote linking and writing lorem ipsum . Other features include AI-powered visual search (find design assets by describing them) and auto-generation of copywriting in designs . The philosophy, as stated by Figma’s team, is to “lower the floor and raise the ceiling” for design work – meaning AI helps beginners overcome blank-page syndrome, and helps experts iterate more quickly. Notably, Figma’s AI features run on third-party models and were not trained on private user files , highlighting a common concern in domain-specific AI: respecting proprietary or creative data. By limiting training to public data and fine-tuning on UI-specific concepts, Figma aims to avoid IP issues while still offering design-savvy AI assistance .

OpenAI’s Responses API is another important development in domain tooling – specifically for developers building AI into applications. Launched in 2025 as a successor to the older chat-completions API, the Responses API introduced a more stateful, agent-oriented model for AI apps . Instead of the client managing conversation history and tool usage manually, the API allows developers to maintain a conversation state on OpenAI’s side (via a conversation ID) and utilize built-in “tools” and function calling more seamlessly . One motivation for this, as observers noted, is that OpenAI’s more advanced “reasoning” models (like the anticipated GPT-4/5 with chain-of-thought) keep internal reasoning traces that are not exposed to users . The Responses API’s statefulness lets OpenAI hide that reasoning in the backend (to protect proprietary methods) while still allowing multi-step agent reasoning to carry over between calls . In practical terms, the Responses API makes it easier to build robust agents – you can call multiple tools in one session without resending full context, and the model’s hidden scratchpad of reasoning persists, making it more capable at complex multi-step tasks . It has also been implied that certain agentic features (like more efficient function calling, or tool use with less prompt engineering) are only fully unlocked via this new API . For developers, this represents an evolution of the toolkit: rather than orchestrating everything client-side with stateless prompts, you can lean on OpenAI’s infrastructure to manage memory and tool orchestration – a step towards “agent SDKs” for building AI-driven software. Indeed, OpenAI has been heavily promoting this API as “faster, more flexible, and easier” for AI apps (though some note that it mainly compensates for closed-off reasoning traces ). Regardless, it shows the direction of travel: stateful AI services that maintain context and handle tool integrations for you, effectively letting developers plug in high-level instructions and get complex sequences of actions in return.

Anthropic’s Claude is also pushing domain-specific capabilities like long-term memory and tool use. We mentioned Claude’s 100K context – hugely valuable for domains like legal (summarizing contracts), finance (analyzing reports), or software (understanding entire codebases). On top of that, Anthropic introduced features in late 2025 to make Claude a more effective agent with tools. They added what they call advanced tool use: Claude can now “discover, learn, and execute tools dynamically” rather than only using a fixed, pre-defined set . For example, an IDE assistant built with Claude could have access to dozens of devops tools (Git, package managers, test frameworks) and only load the ones needed at runtime . Anthropic achieved this via an open standard called Model Context Protocol (MCP), which defines how external tools and data sources are described to AI assistants . With MCP, developers can set up “MCP servers” for various services (Google Drive, Slack, GitHub, databases, etc.), and AI agents like Claude can query those servers for what actions are available . The advanced tool use update included a Tool Search feature, meaning Claude doesn’t need every possible tool loaded in its prompt (which could be 100K+ tokens of API definitions); instead it can search for relevant tools by name and load only those, preserving 95% of its context window . It also introduced programmatic tool use, allowing the AI to execute tool calls via code (Python) for complex logic and looping . These capabilities move AI from being a static question-answer system to being an adaptive problem-solver that can interface with software and data on the fly. For instance, a Claude-based agent could on its own decide it needs to pull data from a database and plot a chart to answer a question – and then actually do it by invoking the appropriate tools, all within one session. This is highly relevant to specialized workflows: think of an AI financial analyst that connects to Bloomberg and Excel, or an AI growth marketer that hooks into Google Analytics, Facebook Ads, and a CMS. Thanks to open protocols like MCP (spearheaded by Anthropic in 2024 as an industry standard ), we’re seeing a convergence of AI and APIs – where AI agents become a new kind of user interface for software, discovering and using APIs as needed rather than through rigid integrations.

In customer service and knowledge management, domain-specific AI is appearing as well. We have tools like Intercom’s Fin, a support chatbot that was fine-tuned on support transcripts, or Gong’s AI for sales coaching. Even design tools have niche AI plugins (e.g. accessibility checkers that suggest fixes, or copywriting assistants specialized for UX writing). The pattern is clear: whatever the domain, there’s an effort to bolt on an AI brain that knows the domain’s language, best practices, and typical tasks. Over the last 12–18 months especially, nearly every software category – from code editors to graphic design, from CRM to project management – has seen AI integrations as a key selling point.

While about 80% of this push is contemporary (post-GPT-3/4 era), it’s worth noting a bit of history: domain-specific AI assistants aren’t entirely new. Vertical AI like “AI legal researcher” or “AI medical scribe” have been talked about for years, but earlier efforts relied on brittle rules or narrow models. What’s changed is the emergence of general models (LLMs, vision models) that can be quickly adapted to niches via prompt engineering or fine-tuning. This has dramatically reduced development time – e.g., Figma didn’t need to invent image generation from scratch; they plugged into existing models and focused on the UX for designers . Similarly, many SaaS vendors are using OpenAI, Anthropic, or Cohere under the hood for text generation, adding their domain data for grounding. This means the pace of innovation is fast and the ubiquity of AI features has exploded since 2022. The next couple of years will likely bring even more specialized AI “co-pilots” – for architects, for educators, for HR managers, you name it – all layering into the software tools those professionals already use.

New Protocols and Architectures Enabling Scalable Integration

For AI to truly soak into the work stack at scale, there’s a need for standards and robust system architectures. Companies and open-source communities have recognized that bespoke one-off integrations won’t cut it; we need the AI equivalent of a “USB standard” for connecting to data and tools. This is where initiatives like the Model Context Protocol (MCP) come in, alongside new agent frameworks and orchestration platforms.

MCP, introduced by Anthropic in late 2024, is essentially a universal translator between AI assistants and external systems . It defines a standard way to expose a data source or service (e.g. a SaaS app, a database, an internal knowledge base) as a server that an AI model can query. Instead of every AI vendor writing custom integration code for Gmail, Slack, Jira, etc., those systems can have MCP connectors that any compliant AI can use. Anthropic open-sourced MCP and even provided a library of pre-built servers for common apps like Google Drive, GitHub, Slack, and Postgres . Early adopters like Block (Square) jumped on this to connect AI agents with their internal systems, and dev tool companies (Zed, Replit, Sourcegraph, etc.) are building MCP support to let their AI features retrieve relevant context from code repositories and dev data . The vision is that down the line, “AI systems will maintain context as they move between different tools and datasets, replacing today’s fragmented integrations with a more sustainable architecture.” In other words, just as modern web apps rely on standard protocols (HTTP, REST/GraphQL, etc.) to talk to each other, AI agents will use MCP-like protocols to seamlessly pull in whatever information they need when they need it.

Another piece of the puzzle is agent orchestration and execution frameworks. Running an AI agent in production involves more than just the LLM; you need to handle concurrency, state persistence, error recovery, scaling to many instances, etc. Projects like Metaflow (by Netflix/Outerbounds) have emerged to serve as the “baseplate” for agentic systems . Metaflow is a workflow engine originally for data science pipelines, but it has added features (in 2025) like conditional and recursive steps to better support the loop-like nature of AI agents that plan, act, observe, and repeat . An Outerbounds article noted that while agent frameworks (LangChain, Microsoft’s AutoGen, Google’s Agent Toolkit, OpenAI’s Agent SDK, etc.) handle the logic of prompting and tool use, they still need a reliable execution environment . A robust agent system in production may involve fleets of agents running in parallel (for throughput or redundancy), which means concerns like auto-scaling, distributed state management, and fault tolerance come in . Metaflow provides those via its cloud pipeline capabilities – it can snapshot the state of an agent (so if it crashes, it can resume), schedule many agents, route requests to specialized models (e.g. use a smaller, cheaper model for simple queries and a big one for hard queries) , and crucially, integrate with MCP tool endpoints securely . This means an enterprise that wants to deploy, say, an “autonomous research agent” that reads thousands of documents and generates reports can rely on an orchestrator like Metaflow to handle the heavy lifting behind the scenes (managing 100s of GB of data, long model context, etc.), rather than building that infra from scratch. The key point: scaling AI agents is not just an ML problem, but a software engineering problem, and new standards and platforms are solving that.

We’re also seeing the rise of no-code or low-code AI orchestration tools for specific domains. Earlier, we highlighted a startup also named Metaflow (not to be confused with Netflix’s) targeting marketing workflows. Its founder described it as allowing non-engineers to “compose AI workflows as naturally as arranging ideas on a whiteboard”, with nodes that “search the web, parse documents, or execute code” without writing code yourself . This signifies a trend toward democratizing AI agent creation. Domain experts – marketers, analysts, etc. – could design an agentic process (like: listen on social media > summarize sentiment > draft content > loop until approved) using visual tools. By abstracting the technical bits (APIs, model calls, memory handling), these platforms empower a much broader set of people to harness AI in their processes. This is analogous to how spreadsheet macros or RPA (robotic process automation) allowed non-programmers to automate tasks in the past, but now applied to intelligent AI-driven tasks.

All these developments (MCP, agent frameworks, orchestration engines) are creating a stack for AI integration. We might think of it as the “AI middleware” that will underpin many of the user-facing capabilities. Much like a web app relies on a web server, database, and protocols underneath, the future AI-suffused workplace relies on this new middleware to connect the AI “brains” with enterprise “body” (data, tools, and computing resources). The encouraging part is that it’s being built with openness in mind – open standards, open-source tools – to avoid a heavily siloed world. That said, big players will certainly try to set de-facto standards via their platforms (OpenAI with its plugins and API, Microsoft with the Azure AI ecosystem, etc.). The outcome will likely be a blend: some universal protocols (like MCP) plus some platform-specific ecosystems (like Salesforce’s closed ecosystem of certified partners feeding Einstein).

Defensibility and Moats: Data Gravity, Workflow Lock-In, and Ecosystem Control

As AI becomes pervasive in software and workflows, a natural question arises for businesses: where is the competitive moat? If everyone has access to more or less similar AI models (since models or their capabilities rapidly commoditize), the differentiation must come from integration and execution. Industry analysts often cite factors like data gravity, workflow capture, and distribution as key moats in this landscape .

Workflow lock-in is one powerful moat. This means owning the end-to-end workflow so tightly that your product becomes the system-of-record for getting a job done . For example, developers might start living entirely in an AI-powered IDE (like Cursor or Copilot Workspace). If that IDE not only helps write code but also manages the project, versions, tests, and deployment via AI, then switching away isn’t trivial – all the “evidence, artifacts, and state” of the dev process are captured in that tool . Similarly, if a marketing team uses HubSpot’s AI agents for content, campaigns, and CRM updates, HubSpot becomes the primary place work happens (beyond just a database). When “teams run their process inside your product, switching has a cost that UI clones cannot beat,” as one analysis noted . In other words, capturing the workflow and data exhaust makes an AI feature more than a feature – it makes it an integral part of operating the business. Products achieving this become platforms that are hard to dislodge, even if competitors clone the AI capabilities.

Data gravity complements this. AI’s quality often improves with access to unique data – e.g. fine-tuning on a company’s customer interactions or continuously learning from user feedback. If a platform captures proprietary data exhaust (like the corrections a user makes to AI outputs, or domain-specific terminology and knowledge), it can refine its AI to serve that user better and better . Competitors, lacking that data, cannot easily replicate the improved quality. For instance, Salesforce’s Einstein gets to train on millions of CRM records and support logs (within each customer’s silo) – a new entrant wouldn’t have that history. Data becomes gravitational: once your processes and historical data sit in one AI-enabled system, there’s a strong pull to stay, because moving would mean losing that accumulated learning. We also see explicit strategies like retrieval-augmented generation where the platform builds a knowledge repository (FAQs, docs, past decisions) that the AI taps into. Over time, this repository (and the feedback loops) constitute a moat; the AI agent becomes smarter and more tailored with use, and that advantage doesn’t transfer if you switch products. In summary, the more you use it, the better it gets – and competitors can’t access that flywheel .

UX surface area and ecosystem control is another defensibility angle. If your AI is embedded in many touchpoints of user interaction (the “surface area”), you become the go-to interface for work. Microsoft, for example, is infusing Copilot across Office apps, Windows itself, and even devices. This broad presence makes their AI sticky – it’s just everywhere the user might need assistance, reducing the need to look elsewhere. Ecosystem control refers to owning complementary pieces so others must integrate on your terms. Salesforce’s approach shows this: they not only have the CRM and the data platform (Data Cloud) but also Slack for communication and MuleSoft for integration – and Einstein spans them all . This means a competitor AI can’t easily wedge in without playing by Salesforce’s rules (e.g., Slack’s new AI features are naturally tuned to Salesforce’s context). Owning the distribution channels and having customer trust (especially in enterprises with security/compliance needs) is another moat . Incumbents with strong enterprise relationships (think Salesforce, Microsoft, SAP) can bundle their AI into existing contracts and guarantee certain compliance measures (like the Einstein Trust Layer that emphasizes security ). Customers might stick with a slightly less flashy incumbent AI if it’s the one that meets their IT policies and is deeply integrated with their stack, versus a startup offering a separate agent app.

Finally, cost and infrastructure moats should be noted. Running AI at scale can be expensive (token costs, inference latency issues). Companies that can optimize usage – by routing simpler tasks to cheaper models or using caching of results – will have an economic edge . For example, an AI platform might use a small local model for trivial requests and only call an expensive API for complex ones, keeping margins healthier . These optimizations compound at scale and can deter challengers who don’t have the same volume or infrastructure savvy. In essence, as AI becomes part of the “plumbing” of work, vendors that manage to do it efficiently and reliably (with good SRE practices, evaluations for quality, and cost control) will outlast those who just layer AI on top sloppily . It’s reminiscent of how early “API wrapper” startups had to either grow into full platforms or get outcompeted; offering a thin UI on top of GPT wasn’t enough, you needed three compounding loops: capturing workflow, capturing data, and improving economics .

In short, the moats in this pre-AGI era of AI are less about having the “best model” and more about having the best integration: owning the process (workflow lock-in), owning the data feedback loop (data gravity), ensuring ubiquity and trust (ecosystem & distribution), and managing efficiency (infrastructure and cost advantages). Companies that achieve these will build defensible positions, while those relying purely on access to an LLM API will find that advantage fleeting.

Outlook: The Next 2–3 Years of Layered AI Absorption

Given these trends, the near future of work looks less like a sudden AI revolution and more like a series of incremental yet profound evolutions. We will likely witness an accelerated version of what enterprise IT has done for decades (automation, digitization), now turbocharged by AI’s ability to handle unstructured tasks. Some grounded predictions for the next few years:

AI as Colleague, Not Just Tool: The mindset will shift from “using an AI tool” to “working alongside AI colleagues.” Already developers talk about their suite of coding agents, and marketers consult AI content strategists. This will normalize. Expect job postings or team structures that explicitly incorporate AI agents (e.g. “AI campaign analyst” as a virtual role in a marketing team). Before AGI ever arrives, we’ll have narrow AI agents with semi-autonomy in specific roles – they won’t replace humans, but humans will delegate more and more sub-tasks to them.
Convergence of Agentic Systems: Right now, we have separate silos – a dev uses a code agent in the IDE, a salesperson uses a CRM bot, etc. In the coming years, these may start to connect. For instance, an AI in your project management tool might coordinate with the AI in your code repo to update deadlines based on development progress. This cross-agent communication could be facilitated by standards like MCP, effectively creating a fabric of specialized AIs that coordinate (with humans in the loop) to drive complex, multi-team workflows. We might see something akin to “digital workstreams” where a chain of AI agents passes tasks along (design -> marketing -> sales), overseen by human managers at key checkpoints.
Rise of Internal AI Platforms: Large enterprises, concerned with proprietary data and customization, will invest in in-house AI platforms. Rather than sending data to third-party APIs, they’ll deploy models on their cloud (or use player like OpenAI’s on-prem offerings via Azure) and use frameworks like Metaflow or LangChain to build bespoke agents that understand their internal systems. These “private copilots” will be a stepping stone to AGI-like capabilities in a narrow scope: a company’s internal AI might know everything about that company’s operations (all documents, all metrics) and serve as a powerful assistant, but it won’t generalize beyond. This could create a competitive gap between organizations that successfully build such internal AI leverage and those that do not. It also hints at the importance of knowledge integration: companies will pour resources into cleaning and structuring their data so that AI agents can utilize it effectively (since an AI is only as good as the knowledge it can draw on).
New Work Standards and Norms: With AI absorbing many tasks, we’ll likely develop standards for human-AI interaction in workflows. For example, just as “DevOps” became a standard practice, we might talk about “AIOps” – not in the sense of AI for IT operations, but operationalizing AI task handling (e.g., how to escalate from an AI to a human, how to audit AI decisions, how to version control AI prompts and policies). Already, experts urge treating prompts and agent instructions like code – with versioning, testing, and monitoring . We can expect methodologies to emerge around designing and maintaining these AI agents in production (a bit like how project management had to evolve with agile methods). Ethical and regulatory standards will also play catch-up: ensuring the AI actions (especially autonomous ones) are logged and transparent, bias-checked, and compliant with laws (like data protection). Before AGI, the challenge won’t be some rogue superintelligence, but very capable narrow AIs making mistakes at scale – so governance will be crucial.
Marketplace and Ecosystem Battles: Just as app stores became pivotal in the smartphone era, marketplaces for AI agents and plugins will be hotly contested. We see early signs: OpenAI has a plugins ecosystem, HubSpot launched a marketplace for Breeze Agents , and Salesforce’s AppExchange will surely include Einstein AI extensions. In the next couple of years, having a rich ecosystem (where third parties provide specialized agents or integrations) can decide platform winners. Users might choose their work software not just on core features but on what AI extensions are available for their niche needs. This dynamic could also mirror mobile vs web platforms – with debates on open ecosystems versus walled gardens. The MCP standard, if widely adopted, could allow interoperability (imagine an “agent” that can plug into both Microsoft Teams and Slack, or both Notion and Confluence), but vendors might resist anything that makes it easy to switch AI providers. So, expect some tussle in the standards arena.
Gradual Narrowing of the “Last Mile” to AGI: Each incremental improvement layers more AI into workflows, reducing the gap between what today’s narrow AI can do and the sci-fi vision of AGI. However, it’s likely that in 2–3 years, human experts will remain firmly in the loop for critical thinking, creative strategy, and final approvals. What will change is the scope of what one person can handle with AI help. A single developer might effectively manage a project that used to require a team, because AI agents handle documentation, testing, and devops. A marketer might run an entire multi-channel campaign (strategy, content, analysis) with AI doing the heavy lifting in each step. This productivity leap could be as transformative as the introduction of personal computers or the internet, but realized through a collection of narrow AI capabilities rather than a sentient general AI. In economic terms, we may see big gains in output per employee – and the nature of some jobs shifting to more oversight and orchestration (e.g. the developer becomes more of a “product manager” for the AI agents writing code).

In summary, the pre-AGI chapter of AI in work is defined by layered integration and specialization. Instead of one AI to rule them all, we’ll have many AIs – each specialized, each embedded in context – collectively absorbing a huge amount of operational toil. Like a sponge slowly filling up, these systems will soak in more and more of our routine cognitive labor, expanding what software can do automatically. The winners in this era will be those who figure out how to architect and govern these layers effectively, and those who leverage them to amplify human creativity and decision-making rather than trying to replace it entirely. By the time AGI looms on the horizon, our work environments will already be saturated with AI assistance, and perhaps we’ll find that a network of smart narrow agents, guided by human common sense, was the more practical revolution all along.