It doesn’t.

What matters is:

1. User experience ownership

2. Data ownership

What doesn’t matter:

1. The model

2. The code to interface with the model

Let’s start with the things that don’t matter.

Subscribe now

Things That Don’t Matter

The Model

Countless open source models will continue to be created, and they will perform quite well. As closed source models get better, the open source models will continue to follow right behind.

In case you didn’t know, it’s quite easy big models (Claude, ChatGPT, Grok, Qwen, Mistral, etc.) to extract from the other big models. At one point, Grok would consistently respond that it was Claude Sonnet when prompted, which suggests that xAI used Claude Sonnet to fine-tune Grok. When it comes to fine-tuning models, you make use of question-answer relationships to better align human expectation with prompt generation. If one company is doing a good job, it’s not hard to take their outputs to tune your models.

Additionally, the transformer architecture is great, but it’s not the final form of machine learning. It has been an incredible auto-regression technique which has led to powerful results, but it’s only a stepping stone to something more powerful. Imagine a model that performs equally as well as the transformer, but without requiring nearly as much compute to train. Imagine an LLM that can be trained while you use it on your machine. The data center hype would pop in an instant. The world would change course.

The Code To Interface With The Model

Anyone can rebuild Claude Code. Anyone can rebuild OpenClaw. Anyone can make some variation of OpenClaw that focuses on security or a smaller bundle size or whatever.

You can either use an API aggregator like OpenRouter, or you can self-host a model on your hardware.

Self-hosting is only getting more powerful. Back to the point earlier that as model costs go down, as models become cheaper to train and run, the moat between models will diminish. If that’s the case, then the important part isn’t the model, nor is it the agent code that’s built on top of the model.

I’m not saying that building an agentic platform on top of models is a bad idea. I am saying that it’s the most cutthroat race to participate in right now, and differentiation is quickly lost in code.

It’s not only about the code though…

Things That Do Matter

User Experience Ownership

While the code doesn’t matter, having repeat users and product distribution are the things that do matter.

Most people are non-technical. Most non-technical people will become more capable of leveraging AI. Most people will not want to build the tools, so whoever owns an experience for non-technical people will do well for themselves.

Technical people will spend their time rebuilding a service that already exists. Non-technical people will spend their time using the tools as they’re provided.

If, and when, it becomes easy enough for non-technical people to self-host models on their computers, then big tech is in big trouble. If it gets so easy that non-technical people can completely rebuild technical tools with 1 prompt, then we will have a new set of problems to discuss. But I don’t think either of those scenarios will be the case for a while. Self-hosting models is fun for about 20 minutes, then you go back to using a model that actually solves the problem quickly (like Claude Opus).

There is a lot of room in the market for new software products to help non-technical people reap all of the benefits of AI though.

Data Ownership

How you think. The words you use. The things you shop for. The questions you have. The way you learn. The way you don’t learn.

All of this is getting codified, and more.

The world of consumption and advertising is eager for all this data. Whoever owns this data basically owns the psychology of humanity. Google was the best at search for the previous generation. The floodgates have opened, and the companies that own this experience will do anything to stay relevant until they become embedded in the fabric of technology. That’s why these investments are so big, owning the experience and the data comes with a price.

If distribution is what matters, this leak has helped give Claude some publicity without a doubt. Is it a PR stunt? I don’t know, but as the old saying goes:

There is no such thing as bad publicity.

Episode Timeline
00:00 Episode Preview
01:19 Introduction to AI in Enterprises
03:50 Transforming Business Processes with AI Agents
06:49 The Emergence of Semantic Engineers
09:22 Understanding Semantics in AI
12:17 The Role of Humans in AI Implementation
15:01 Measuring AI Success in Enterprises
17:55 The Future of SaaS and AI Integration
20:12 Identifying Processes for Automation
22:23 The Concept of Digital Employees
25:13 The Future of Work and AI
27:55 Trust and AI in the Workplace

Key Topics:

• The importance of fast decision-making in startups versus large corporates

• Leadership traits crucial for founders, especially risk-taking and decisiveness

• How AI can enable efficiency and decision-making at scale in organizations

• The distinction between AI agents (autonomous, open-ended tasks) and workflows (deterministic processes)

• Practical applications of robotics: from elder care companions to industrial automation

• The chicken-and-egg problem of training data for autonomous systems and the importance of limiting use cases for success

• Future of AI co-pilots as workplace mentors, coaches, and mental health aides

• Cybersecurity risks and trust considerations with AI agents in the workplace

• Global differences in robotics development and factors affecting innovation

• Practical advice for investors and founders on adopting AI technologies

Timestamps:

00:00 - Introduction to Pramod Gosavi’s background in engineering, product management, and VC
02:37 - Decision-making importance for startup founders vs. corporate leaders
04:11 - How team dynamics influence startup success
06:23 - The role of AI in transforming workplace productivity and decision-making
08:46 - Future of AI co-pilots: workplace mentors and mental health coaches
11:15 - Challenges and opportunities in integrating AI agents into business processes
13:23 - Rappers vs. workflows: understanding the future of AI tools
15:10 - Autonomous agents vs. deterministic workflows in automation
17:37 - Robotics applications: elder care, industrial automation, and limited use cases
20:19 - Global robotics innovation: US versus China
20:25 - Pramod’s preferred social platforms for connecting and sharing insightsResources & Links:

Connect with Pramod Gosavi:

• LinkedIn

While these job titles still exist, you’re finding them less and less in non-technical businesses. And the one’s that do exist interact with the cloud more and more frequently.

Much like the cloud consolidated physical infrastructure for the enterprise, agents are consolidating business processes for the enterprise. The very nature of the enterprise will evolve much like it did during the cloud era.

Business Processes Are Next

With the rise of agents, there are plenty of bets that 2026 is the year of the agentic co-worker. Just briefly consider Claude Cowork, OpenAI Frontier, and Ariso. The topic of performance reviews, employee sentiment, and what this means for the workforce will be saved another newsletter.

Today’s focus is: what happens to the departments, the divisions, the organizational functions that have spent decades encoding business logic into human workflows? What happens when this business logic becomes intertwined with humans and agents?

This is not a distant, speculative question. It is already happening. The open question is not whether your organization will automate its processes, but how deliberately you will do it, and whether you will be left with a coherent governance structure on the other side.

What Gets Automated First Matters

Agentic automation starts in the low-friction, high-volume workflows: lead qualification, invoice processing, customer support triage, contract summarization, data entry reconciliation, customer experience. These are the processes that organizations have long wanted to automate but found too complex or too fragile for traditional RPA.

Agents can reason over unstructured data, handle exceptions, and pass control to a human when needed. This makes them suited for exactly the kinds of messy, semi-structured business processes that make up the majority of enterprise operational work.

You cannot automate what you have not defined. Clear SOPs and business processes must remain framework-agnostic, human-readable, versionable, and governable.

The trap most organizations fall into is treating each automation project as a one-off engineering effort. A workflow gets built in LangGraph. Another gets built for OpenAI Assistants. A third is deployed as a python script in a lambda. Each team reinvents guardrails — or ignores them altogether. Business logic gets buried in framework-specific code only the original developers can touch. Business stakeholders lose visibility into their own processes.

This is the same mistake enterprises made with cloud before infrastructure-as-code became standard practice. Terraform did not compete with CloudFormation at the code level, it owned the declaration and the state. The agentic equivalent is emerging now: a process contract layer that is framework-agnostic, human-readable, versionable, and governable.

How Every Department Must Adapt

The shift to agentic automation is not purely a technology story. It is an organizational redesign. Each department faces a distinct set of pressures, opportunities, and risks. The departments that adapt deliberately will look very different from those that are disrupted reactively.

Finance & Accounting → From Reconcilers to Auditors

AP/AR, reconciliation, variance reporting, and expense classification are among the highest-ROI candidates for agentic automation.

The finance function evolves from processing transactions to auditing agent behavior and managing cost attribution across automated workflows. The CFO’s new question should not be “how do we process these invoices faster?”, it should be “how do we govern the systems that process them automatically, and how do we attribute their costs accurately?”

Legal & Compliance → From Reviewers to Policy Architects

Contract review, regulatory filings, compliance monitoring, and dispute documentation are already being addressed at scale.

Legal must be involved when defining what agents can access, what requires human approval, what constitutes a violation before one occurs. Policy becomes operational infrastructure, not a PDF in a shared drive.

The enterprise lawyers and compliance architects who will be most valuable in the agentic enterprise will do three things well:

1. Translate regulatory intent into operational controls

2. Maintain clarity on the rapidly evolving legal landscape

3. Hold the engineering team accountable if the implemented controls drift

Most regulatory frameworks were written for human workflows. Regulators are behind. Courts are behind. Compliance in 2026 means making defensible calls in a gray area, not following a clear rulebook. Every agent that touches a regulated record is a compliance event.

Human Resources → From Administrators to Experience Designers

Hiring, onboarding, benefits administration, payroll query resolution, and compliance training are ripe for automation.

Once agents are embedded in workflows, HR has to consider questions it’s never had to answer before. How do you measure individual performance when an agent contributed to the outcome? How do you structure compensation? How do you maintain culture and employee trust in an organization where headcount is shrinking but output is growing? What laws (union, state, federal, international) might play a factor into hiring and headcount?

On one hand, HRBP of 2027 will spend less time processing paperwork and more time designing the conditions under which people do their best work alongside automated systems. Whether this leads to headcount reduction is yet to be proven, but signs point this direction. However, most headcount reduction seems prioritize shareholder sentiment and cutting a bloated workforce at the time of this writing.

Sales & Revenue → From Pipeline Managers to Orchestrators

Lead scoring, outreach personalization, proposal generation, and CRM hygiene become agent tasks.

Sales leadership moves from managing reps through activities to orchestrating multi-agent pipelines that surface the highest-leverage human moments. The best salespeople will be the ones who know how to work with agents to amplify their reach while maximizing the moments that require genuine human connection.

Marketing → From Creators to Campaign Orchestrators

Drafting, personalizing, A/B testing, and distributing content are among the highest-volume, lowest-differentiation tasks in a marketing department.

Brand messaging is not one process to fully automate though. Knowing which message lands, which framing is off, and which campaign would embarrass the company in six months will all be agent-assisted processes. Marketing shifts from producing content to curating and directing the systems that produce it.

The marketers who thrive will be:

1. The ones who can brief an agent as precisely as they once briefed a copywriter

2. Those who understand which creative decisions still require a human with taste in the room

Operations → From Coordinators to Control Plane Operators

Procurement, customer services, vendor management, logistics coordination, supply chain, and quality control workflows are increasingly agent-executable.

Ops becomes the team that maintains process definitions, monitors agent behavior, and manages escalation when automation fails or behaves unexpectedly. This is a more technical role than traditional operations, and a more strategically important one.

Traditional silos will merge heavily across operations as technical team members embed themselves even closer to the people running the business processes. This area of any business stands a lot to gain as far as individual business advantage is concerned.

Data → From Stewards to Semantic Architects

Agents generate data as a byproduct of every action they take: summaries, classifications, extracted entities, decision logs.

This generated data quickly expands on the source records that it originated from. The data team has to decide what gets stored, where, and for how long, all while managing a new data challenge (read: risk) that traditional governance frameworks weren’t built for: semantic drift.

The business’ data is quickly becoming the semantic layer that enables all agentic operations to have relevance and coherence. Data attribution, traceability, and semantic architecture are a necessity when implementing quality agentic solutions. As the old saying goes, “garbage in, garbage out”. And while you will ultimately want to spend time fine-tuning base models to improve context, the starting place is to semantically structuring your data with tagging, ontologies, relationships, and more.

Cybersecurity → From Defenders to Control Plane Auditors

Code quality checks, application security testing, automated pen-testing, governance frameworks and automation, and modern identity solutions for agents are all under development.

But every agent that connects to a system, processes a credential, or executes an action in production is a new attack surface. This attack surface doesn’t get tired, doesn’t notice when something seems off, and doesn’t stop to ask questions when an instruction looks suspicious. The attack surface has changed from human principals with behavioral baselines to non-human identities that accumulate permissions, rarely get reviewed, and often have no off-boarding process.

Security’s job is shifting from approving & denying deployments, to defining what “normal” looks like for every automated process. This introduces a new class of security on top of the existing security stack, one that doesn’t feel solved in the slightest. CISO’s not only should be asking themselves “what are our users doing?”, but also “what are our agents doing, and how do we know?”.

IT & Engineering → From Builders to Technical Governance

Engineering covers a broad set of specializations that will fracture under agentic pressure.

Platform and DevOps teams become the governance layer: owning process contract standards, validating that implementations match specifications, maintaining infrastructure, and operating the observability infrastructure every other department depends on. But a meaningful subset of engineers won’t govern from a distance. Instead they’ll move closer to the business as forward deployed engineers, embedded in departments long enough to understand the actual work and trusted enough to automate it correctly.

The function that may be most disrupted isn’t engineering at all, it’s project management. The traditional PM role exists to coordinate handoffs between business stakeholders and builders. When a forward deployed engineer is already embedded in the department, when process definitions can be written and deployed in days rather than sprints, and when the feedback loop between “here’s the problem” and “here’s the automation” compresses to weeks, this layer of coordination overhead becomes friction. The organizations that move fastest will be the ones that shorten the distance between the person who understands the process and the person who can automate it. Eventually, that may even be the same person.

Closing Thoughts

So much more could be said, and there is far more nuance than this newsletter allows. One thing is for certain though — there is significant upside for the organizations who capture their business processes and automate away the low hanging, high impact processes. The amount of capital that will be unlocked during this era will eclipse the previous SaaS era in ways that we won’t understand until it happens.

What do you think is going to happen?

Hosting a podcast in tech means that things change RAPIDLY. But whether you’re hosting a podcast, or trying to keep your organization up to date on the latest information, you’ll know that it’s not easy to do. “The next greatest thing” releases and you must quickly discern if that is worth your attention or not.

So I opened up Claude and designed Up2Speed:

What it actually does

Up2Speed pulls from nine sources I actually care about. Hacker News, GitHub Trending, Dev.to, ArXiv papers, Lobsters, Wired, Ars Technica, TechCrunch, and podcasts. All in one dashboard. No more tab-hopping. And especially no more

Everything refreshes in parallel. If one source is slow or down, the others don’t wait around. You get what’s available, fast.

It’s early in development, but I already love it and use it daily.

The podcast problem

Podcasts have become one of my go-to sources of information for thought leadership and critical thinking when I can’t read. Driving in the car, mowing the lawn, shoveling the snow. I’ve also found that all of the news sources I pull from make no mention of podcasts that could be relevant to the topics.

So I made podcasts a first-class citizen. They sit right at the top in their own section. The Changelog, Syntax, Practical AI, Software Engineering Daily. They’re all there. And you can add your own feeds.

Filtering the noise

The real magic is in the tags. I built smart filtering that automatically categorizes content: AI Agents, LLMs, Robotics, Web Dev, Rust, Go, Security. You name it. Click a tag, and suddenly you’re only seeing what matters for your interests.

You can even create your own tags with custom keywords. Because my categories aren’t necessarily your categories.

It’s not as robust as I’d like (i.e. AI parsing), but it’s a start.

Why I’m sharing this

I originally built it for myself and my co-host / co-founder. I thought he’d think it’s cool, but his response hyped me up:

So I figure that if I’m drowning in information overload, other people probably are too. Product managers. CTOs. Podcast hosts. Anyone whose job involves staying current in a world that changes overnight.

It’s open source. You can run it as a web app, deploy it to Vercel, or even run it as a native desktop app. No accounts, no tracking, your settings stay in your browser.

Links to the code, example website, etc.

Code: https://github.com/intertwinehq/up2speed

Example website: https://up2speed-jet.vercel.app/

Subscribe for more free code projects, upcoming podcasts, interviews, and thoughts.