In a recent fireside chat, two key voices at the forefront of enterprise AI shared the stage to unpack the evolving landscape of agentic AI: Matan-Paul Shetrit, Director of Product at Writer, and Sandesh Patnam, Managing Partner at Premji Invest.
While much of the generative AI hype has centered around flashy consumer tools and foundation models, this conversation turned the spotlight toward the enterprise workflows, compliance realities, and investment strategies shaping the future of practical AI deployment.
“With all things agents, I think it's top of mind for a lot of folks,” said Patnam, kicking off the session with a nod to the rapidly emerging world of agentic systems.
Before diving into technical challenges or product design, Patnam framed the discussion with an investor’s perspective. Representing Premji Invest, a long-term capital partner, he offered a contrasting approach to the fast-paced, bet-heavy world of venture investing in generative AI.
“You see a lot of investments being made today... but we think about things a bit differently.”
Premji Invest is looking for long-term value and companies that can win in the evolving enterprise AI stack. One of those companies is Writer, an enterprise-grade AI platform built to integrate securely and scalably into business workflows.
“Writer is one of our most promising companies; growing amazingly fast,” Patnam noted, setting the stage for a deeper conversation with Shetrit about what it takes to actually make generative AI work in the enterprise.
Inside Premji Invest
To understand why Writer stands out in Premji Invest’s portfolio, it's important to understand the firm itself.
Premji Invest is the captive fund of a $25 billion endowment started by philanthropist Azim Premji. Known for giving away the majority of his wealth to charitable causes, Premji created one of the largest endowments in the world, and the investment firm that supports it is mission-aligned for long-term impact.
“Our mission is essentially to give back to the endowment... a set of companies that are enduring and thriving for a long period of time,” said Patnam.
Premji Invest is a crossover investor, operating in both public and private markets. But unlike traditional venture capital firms that place many small bets hoping for one or two breakout wins, Premji Invest applies a long-term lens to every investment. Their approach is more like anchoring a company for the long haul, with deep conviction and sustained capital support, from private rounds all the way through public listings.
“We don’t think of it as a set of bets,” Patnam explained. “Each of these are partnerships.”
Their portfolio spans early to late-stage tech, life sciences, consumer, and financial services across both the U.S. and India. But what ties it all together is thematic, high-conviction investing with a focus on companies that demonstrate:
- A strong vision
- Efficient, sustainable growth
- Defensible moats
- The potential to compound value over decades
“So when I say Writer,” Patnam added, “it implies based on everything I just said, this is the kind of company we believe can sustain and thrive long term.”
Why full-stack matters in generative AI
When it comes to investing in generative AI, Premji Invest takes a thematic, cycle-informed approach, one rooted in lessons from past technology waves.
“People think of these as cycles,” Patnam noted. “First picks and shovels, then middleware, then applications... but in practice, it’s never that clean.”
Looking back at past inflection points, the rise of the internet, the shift to cloud, and the mobile explosion, Premji Invest observed a key pattern: innovation unfolds organically across layers, not in strict vertical order. Infrastructure, middleware, and application development happen in parallel, not sequentially. And the companies that endure? They typically operate across those layers.
Patnam criticized the early generative AI companies that had a myopic view:
“I’m just going to build models.”
“I’m just middleware.”
“I’m just building applications.”
That level of specialization, he argued, is fragile in a market evolving this quickly.
“When a cycle is moving as fast as this one, and you’re only focused on one layer, you don’t know where the disruption’s going to come from. It’s very hard to build an enduring company that way.”
Instead, Premji Invest looks for full-stack generative AI companies that have:
- A point of view on models (whether internal, open source, or hybrid)
- A grip on middleware and delivery mechanisms
- A deep understanding of the end-user workflow, especially in the enterprise
This multi-layer fluency enables better tradeoffs, faster iteration, and stronger defensibility. If a company can control and optimize decisions at the model, middleware, and application layers, it can deliver the most effective and efficient outcomes.
“The best companies have a full stack view... they can make the right tradeoff at any layer to deliver value,” Patnam said.
Enterprise AI is about workflows, not just chatbots
When most people think about enterprise AI, their minds jump to chatbots; interfaces that spit out answers, complete text, or summarize reports. But according to Patnam, this narrow view misses the true potential of generative and agentic AI in the enterprise.
“People think of enterprise AI as just a chatbot, just text-based models. But it’s much more than that.”
With the rise of agentic AI, AI systems that can perceive, reason, act, and iterate autonomously, the industry is finally beginning to see the bigger picture. And at the heart of that picture? Workflows.
“The secret sauce is workflows,” Patnam emphasized. “Most of what we do in enterprise is workflows: handoffs, creative decisions, deterministic queries, and data stitching to serve an end-user outcome.”
To bring that abstract idea to life, he shared a practical example from the world of wealth management:
- A wealth manager prepares personalized reports for hundreds of clients.
- They must pull specific client data from internal systems.
- They craft a point of view, marrying market trends with portfolio activity.
- They generate charts and visuals, align tone with brand, and assemble everything into a compliant, polished communication.
- Finally, the reports are reviewed and sent.
This is not a chatbot task. It’s a multi-stage workflow involving creativity, data retrieval, formatting, compliance, and human nuance.
Now imagine this entire process, from data pull to delivery, powered by AI agents working in tandem. That’s the vision.
“Imagine a world where all of this is being done by agents.”
In Patnam’s view, this is where the next generation of enterprise AI is headed: toward intelligent orchestration of real business processes, not just point-and-click interactions.
Sahar Mor
The rise of agents: From tasks to thinking
Many of today’s AI agents are task-specific, fetching data, generating a paragraph, or summarizing an email. But according to Patnam, the future of enterprise agents is far more sophisticated.
“This agent has to think. It has to reason. It has to make a decision.”
Enterprise-grade agents don’t just need to execute. They must:
- Query structured data with precision (no hallucinations)
- Make deterministic decisions where accuracy is critical
- Generate text or visuals where creativity and nuance are needed
This requires a careful balance between deterministic logic and generative capability, something few models, and even fewer systems, can do well.
“It’s a marriage of multiple things,” Patnam said. “These agents don’t just work in isolation; they must work in parallel and in synchrony.”
And this orchestration of intelligent agents, layered across workflows, is exactly where Writer comes in.
“Enterprise AI has to be something to do with that. That’s the fulfillment of the promise. And therein lies Writer.”
With a full-stack approach, Writer isn’t just layering generative models on top of enterprise software. It’s bridging the entire system:
- Models that are flexible but safe
- Middleware that translates between model outputs and business logic
- Applications that enable real, native workflow automation, from deterministic tasks to creative synthesis
Writer’s platform is designed to think, reason, create, and deliver, all within the guardrails and rigor required in enterprise environments.
“Think of applications that are guardrail-specific... and ones that require creative energy. Writer enables both.”
With that, Patnam passed the mic to Matan Shetrit, Writer’s Director of Product, ready to dive into how this vision becomes real inside enterprise workflows.
Inside Writer’s model philosophy: Full-stack by design
With the vision for agentic AI and enterprise workflows laid out, the conversation shifted to the technical heart of Writer’s platform: its models and system architecture.
Sandesh Patnam teed up the topic by pointing out the complexity of modern AI stacks; everything from custom models to orchestration frameworks like MCP and A2A that many in the audience likely use. But what sets Writer apart is its first-principles approach to system design.
“You have a very sort of first-principles system design view to model building,” Patnam noted. “And a familiar group of models that have done phenomenally well.”
Turning to Matan Shetrit, Director of Product at Writer, Patnam invited him to pull back the curtain on how those models are developed; models that, as he described, were “deepseek before DeepSeek.”
Shetrit responded:
“Thank you for that. I don’t know if I need to say anything after that intro; maybe we can all leave now.”
But joking aside, the timing was perfect. Writer had just launched a new model the day before.
This marked another milestone in the company's pursuit of tightly integrated, workflow-aware generative models, ones built not just to generate text, but to reason within enterprise contexts, stay grounded in company-specific data, and enable task automation at scale.
From this point forward, the discussion would dive deeper into model capabilities, framework integration, and the architectural decisions that make Writer’s AI stack production-grade.
Meet Palmyra: Writer’s high-context, enterprise-ready model family
With timing that couldn’t have been better, Shetrit revealed that Writer had just launched a new model the day before:
Palmyra X5, the latest addition to the company’s in-house LLM family.
“It’s the first 1 million token context window model on Amazon Bedrock,” Shetrit noted, adding that it runs at just $0.60 per million input tokens and $6 per million output tokens, with sub-300 millisecond tool calling.
These aren’t just technical bragging rights. They reflect Writer’s obsession with production viability in enterprise environments, where cost and speed are non-negotiable when agents are executing multi-step, real-time workflows.
The Palmyra line isn’t new; Palmyra X4 debuted in Q4 of the previous year. But even back then, Writer was already challenging the narrative that LLMs required $100M+ training budgets.
“X4 cost us $700K to bring to market,” Shetrit said. “X5? $1M in GPU cost. And that’s before DeepSeek and the $5M model hype.”
The message is clear: great models don’t have to be expensive. If you optimize for the right outcomes. And in enterprise AI, those outcomes are low latency, low cost per call, and compatibility with complex workflows.
AIHQ: The enterprise agentic platform
The launch of Palmyra X5 was also a continuation of AIHQ, Writer’s agentic platform designed for enterprise-scale automation. With a 1M-token context window and ultra-fast tool calling, Writer’s infrastructure can now support:
- Multi-hop agentic flows
- Memory-embedded reasoning
- Secure, enterprise-native tool usage
“For LLMs to be used in enterprise agents, you need speed and cost-efficiency,” Shetrit explained. “If it costs hundreds of dollars per call, it’s never going to work at scale.”
That’s especially true when agents are chaining steps across multiple internal systems, integrating with compliance tools, querying proprietary data, and generating polished outputs for external stakeholders – all in real time.
And real customers are already proving the value.
“We’ve got companies like Vanguard, Franklin Templeton, and Uber building real applications using AIHQ and seeing real ROI.”
Built for enterprise from day one
Unlike many startups that began as research labs or consumer tools before pivoting into enterprise, Writer was built for business from the start. That means its model architecture is about supporting:
- Compliance
- Data retention policies
- Legal safeguards
- Deterministic behaviors + generative creativity in balance
“We didn’t start as a research lab dabbling in enterprise,” said Shetrit. “This was the plan from day one.”
Writer’s enterprise-first model strategy
From day one, Writer has positioned itself not as a tooling layer or a wrapper around someone else’s model, but as a fully end-to-end AI platform built exclusively for enterprise.
That orientation affects everything: how models are trained, how updates are handled, and how reliability is guaranteed.
“We’re an end-to-end platform for enterprises and their AI needs,” Shetrit emphasized. “And everything we build drives toward that.”
Unlike large foundation model providers, Writer works within a different set of constraints, ones shaped by enterprise expectations and regulatory realities:
- No training on user data
- No distillation or silent model swaps
- No regressions that quietly break workflows
“It’s a real problem,” Shetrit said. “In the big labs, you can’t distill or quantize the model over time if you can’t train on user data. But that’s table stakes for us.”
He illustrated the point with a story he often shares with CIOs:
“You think you’re buying a Porsche. But two weeks in, someone sneaks in, steals your engine, and replaces it with one from a Kia. The car still runs, but now it takes 10 seconds to hit 60 instead of two... and you're paying the same.”
In mission-critical business flows, that kind of performance regression is a breaking change and a dealbreaker.
Writer's solution? Synthetic data and first-principles training methods enable them to evolve their models without touching user data, thereby preserving quality and latency.
These constraints, which might seem limiting, actually fueled creativity and precision, leading to:
- Smaller, more cost-effective models
- Highly specialized enterprise fine-tuning
- Self-adaptive, self-learning capabilities that stay within enterprise guardrails
“It pushed us to be very creative,” Shetrit said. “Both on the training data side and the cost structure side.”
And enterprise customers are feeling the impact in real outcomes, not just benchmarks.
Self-evolving models and the future of AI in enterprise
As the conversation wrapped up, Patnam posed a natural but pressing question: Where are large foundation model providers headed as agents become central to enterprise AI?
Shetrit didn’t hesitate; Writer had already been thinking about this.
“We published our opinion back in Q4, and there’s a paper coming out soon on what we call self-evolving models.”
According to Writer, the next frontier in enterprise AI will be about smarter, more adaptive systems that continuously learn and improve on the fly.
There are two major pillars to this idea:
- Reasoning vs. knowledge
- Knowledge can be retrieved externally through RAG, APIs, or delegated tools.
- Reasoning, however, is internal; it's about how a model interprets inputs, weighs options, and decides what to do next.
- Improving reasoning is the real frontier.
- Contextual adaptation
- The dream of a single, monolithic model that works across every workflow in a complex enterprise is, frankly, unrealistic.
- Even within Writer’s own teams, across just 2–3 product groups, needs and behavior vary widely.
- So why expect one frozen model to be a one-size-fits-all solution?
“This idea that a model could be a fit for every team out of the gate is not realistic.”
Instead, Writer’s emerging thesis is around models that evolve in response to real-world use, adapting based on agent feedback and workflow-specific interactions. These self-evolving systems would personalize themselves over time, within strict enterprise guardrails, delivering a tailored agentic experience for each user or team.
“Yes, agents don’t work well at first,” Shetrit admitted. “But they give feedback. And if your system can learn from that feedback, safely and efficiently, that’s how you create real enterprise value.”
Anthony Witherspoon
Guardrails, air Gaps, and the messy reality of model deployment
Writer’s vision of self-evolving models is powerful, but it’s not without its pitfalls.
Shetrit shared a candid story that exposed the limitations of current guardrail techniques: when one of Writer’s new models was released publicly for just 24 hours, it began to self-uncensor based on user input patterns.
“Like a lot of us, it became a worse person,” he said, half-joking.
This unexpected behavior underscored a hard truth: guardrails today are not sustainable.
“The way we do guardrails today needs to evolve. We need more than hardcoded filters or prompt-blocking.”
That’s why Writer now runs certain enterprise deployments in air-gapped environments; physically isolated systems that allow for experimentation without the risk of open-loop contamination.
Other companies are wrestling with the same problems. Shetrit pointed to efforts like SSI (Ilya Sutskever’s new company), which is focused on adaptive, behavior-aware models, as well as Anthropic and OpenAI, both of which are actively exploring models that evolve from user feedback. But in enterprise contexts, that kind of adaptation can’t come at the expense of predictability or compliance.
Deployment complexity: Version control for models
Beyond behavior, there’s the question of operational control. In enterprise settings, it’s not enough to have a smart model; you also need the tools to manage, roll back, audit, and govern it.
“Maybe the model goes off the deep end. You need to be able to revert to an older version.”
This requires a system architecture that mirrors the rigor of traditional software engineering:
- Version control for models and agent configurations
- Workflow reproducibility
- Failure detection and rollback mechanisms
- Security and isolation policies
Without this infrastructure, agentic systems are unlikely to move beyond the lab.
From prototype to production: What enterprise needs to get right
And that’s the reality for most companies today: 90% of AI projects stall at the prototype stage.
“Something about the system architecture and design isn’t translating to production,” Patnam noted.
So, what does it take to get there? Writer’s team believes the key is designing with agents and enterprise workflows in mind from day one, not slapping agents onto brittle foundations.
“To get from a 30% success rate on multi-hop agents to 95–99%, we need adaptive models, robust workflows, and full-stack infrastructure.”
That’s where Writer is investing, and that’s what they believe will define the next generation of enterprise AI platforms.
The real bottleneck: Change management, not just tech
You can have the best model. The fastest inference. The smartest agents. But if your organization isn’t ready for it, your project will stall.
“We’ve all seen what happened with cloud,” said Shetrit. “Whole businesses like Accenture were built around deployment and change management. The same is happening with AI.”
Too many companies still approach AI as if it's plug-and-play; offer an API, connect a model, and let the enterprise "figure it out." That, Shetrit said bluntly, doesn’t work.
A full stack that includes organizational change
From day one, Writer’s approach has included change management as a top-layer feature of the stack right alongside models, RAG pipelines, and guardrails.
“Either we work directly with companies, or we partner with firms like Accenture to handle knowledge transfer, co-design workflows, and build the muscle inside the org.”
Without this human-centered enablement layer:
- Projects fail to move beyond pilot
- Organizations don’t expand usage
- Larger contracts never get signed
Collaboration between business and IT is the missing link
Beyond deployment, there’s also the build process itself, which, for years, has been broken.
“Someone writes a product spec, throws it over the fence, and hopes IT builds it. That didn’t work before. It really doesn’t work with AI.”
The explosion of consumer-facing AI has only raised expectations. Shetrit recalled:
“Last year, my mom called and said, ‘Have you heard of OpenAI?’ I said, ‘Yes, Mom, that’s how I write your birthday notes.’”
This kind of mass adoption means enterprise users expect the same polish, personalization, and usability they see in consumer apps. The traditional separation between business users and IT just doesn’t cut it anymore.
AI supervisors: A new role for IT
With their AIHQ platform, Writer is pushing a more collaborative model: One where line-of-business teams and IT work together to design, build, and supervise AI workflows.
“We need to empower IT to become what we call AI Supervisors,” Shetrit explained.
These roles aren’t about building models from scratch. They’re about:
- Governing deployments
- Maintaining compliance
- Adapting workflows
- Ensuring reliability and scale
In other words, they’re the stewards of enterprise AI in production.
The future of enterprise AI: Blurring the lines between business and IT
As enterprise AI matures, the lines between business users and technologists are blurring and Writer is actively shaping that transformation.
“IT doesn’t want to be the one in the room saying no to AI,” Shetrit said. “But right now, they’re freaking out, they don’t have control.”
That’s why empowering IT to become AI Supervisors is so critical. It’s not about handing them responsibility for model development, but about giving them tools, oversight, and collaboration models that work at scale. From tool call validation to compliance controls, IT needs to feel confident that what’s being deployed won’t break workflows or the business.
“We work with heavily regulated industries, like healthcare, finance, and we’re not expecting citizen developers there,” Shetrit clarified. “But the separation between IT and the line of business is disappearing.”
In Writer’s view, the future of enterprise AI lies in tight collaboration:
- Engineers drive deeply integrated use cases
- Business users define workflows and goals
- IT ensures safety, reliability, and compliance
“That’s where the future of enterprise AI is going, and that’s what we’re pushing for.”
AI accelerator insider insider
Thank you for subscribing
Level up your ai accelerator insider career & network with ai accelerator insider experts.
An email has been successfully sent to confirm your subscription.
Follow us on LinkedIn
