George Mathew sees software moving from brittle rule engines toward AI systems that can reason about changing situations and business processes.
2
Agentic systems will combine foundation models, knowledge sources, and specialized reasoning to automate work, while some new products will be built from scratch around AI.
3
High-quality data remains necessary as models become more efficient, and companies must weigh managed platforms against the engineering cost of building their own systems.
Summary
George Mathew describes a shift from software built around encoded rules to AI systems that can reason, coordinate agents, and carry out work. He expects these systems to include software, implementation services, and synthetic labor such as developers, sales representatives, and process automation workers. Some agents will extend existing applications, while other products will discard old workflows and create new ones. Mathew also discusses DeepSeek, model distillation, and the prospect of compact domain-specific models that stay within tighter boundaries. He argues that high-quality data remains a central resource even as model training and inference become more efficient. The conversation covers build-versus-buy choices in ML infrastructure, custom software built above changing component layers, and the emergence of ambient agents that handle tasks in the background. Mathew is positive about the possibilities, but he also acknowledges concerns about alignment, model quality, and the uncertain economics of foundation-model companies.
Generational companies are built across long software cycles
George Mathew connects his idea of "generational outcomes" to the long arc of software, from client-server systems and the web to mobile and cloud infrastructure. Insight Partners has now reached its 30th anniversary and raised its 13th fund, but Mathew's larger point is about building companies that remain important through major technology shifts. He sees AI as the next major change in software and expects its effects to unfold over the next several decades. Software has become a basic layer for productivity, and he believes the next generation of companies will shape how people and organizations work with AI.
AI systems need the modern data stack and coordinated agents
Mathew says modern AI systems depend on high-quality data, which made the modern data stack an important foundation for current models. He describes a progression from the Transformer architecture and foundation models to machine learning operations, language-model operations, and agent-based systems. These agents can bring together several reasoning systems and knowledge systems for a particular problem. The system is therefore more than a model answering a prompt. It includes the orchestration needed to combine models, data, and actions into a working AI system for an existing or new use case.
Reasoning engines can replace brittle rule-based software
Most software has historically encoded business processes as rules. Mathew says those rules become brittle because someone must keep updating them as the business changes. A reasoning system could interpret what is happening and decide how to respond, potentially with human-like or better reasoning. He expects AI-first systems to combine software with services and labor. Examples include synthetic sales development representatives, developers, and process automation workers that handle routine work and exceptions. His investment in CrewAI reflects this direction, replacing manual robotic processing with systems that can reason about back-office tasks.
Existing applications will coexist with entirely new AI workflows
Discussing an agent that failed when moved from a Jira hackathon into a real Jira environment, Mathew says both integration and replacement will happen. An existing system can feed an agentic workflow, with a foundation model supplying reasoning and a retrieval-augmented approach supplying context. That can produce useful compound AI systems. In other cases, teams will decide that the old process should not be preserved at all. They may build a new system around the task instead of forcing an agent into an application whose data and workflows were designed for people.
Autonomous systems can work differently from humans
Mathew says he has moved from expecting mostly copilot products to expecting fully autonomous autopilots. These systems may not perform work in the same sequence a human would follow. He sees value in probabilistic models interacting with one another and producing unexpected approaches, especially for creative work. That freedom needs to depend on the task. Creative systems may benefit from diverse outputs and a sandbox for experimentation. A trade settlement or straight-through back-office process needs a tightly controlled model that handles exceptions while staying within its boundaries.
Efficient domain-specific models could challenge large model builders
Mathew says DeepSeek made the efficiency question hard to ignore. Model distillation and reinforcement-learning techniques can reduce a large model into a compact, domain-specific system that is cheaper to train and run. Such a model can provide high-quality reasoning while staying focused on its intended task. He does not assume that progress requires ever-larger models with vastly more parameters. Open-source models have often approached the capabilities of leading systems within months, which creates pressure for companies spending heavily on frontier models to show a lasting advantage.
Build-versus-buy decisions include the cost of engineering ownership
The conversation uses a community example where a team reported saving money by moving away from SageMaker AI and running more of its own infrastructure. Mathew says the visible platform bill does not capture the engineering effort needed to operate the replacement. Team size, maturity, regulatory needs, and the priority of the work all affect the decision. He has seen enterprises take pride in building bespoke ML operations pipelines, then face recurring decisions about whether to maintain or replace them. For teams without unusual data or regulatory constraints, he generally favors outsourcing infrastructure so ML engineers can focus on production AI systems.
AI makes custom software more practical above changing components
Mathew sees a new opening for custom software because AI can help teams build highly specific experiences without owning every underlying component. Data, pipelines, and other building blocks can keep changing while the organization iterates on an abstraction layer designed for its users. This differs from waiting for a package-software vendor to add a needed feature. He also points out that open-source projects often begin as solutions to a specific company's problem, then become general tools. Airflow came from Airbnb, and Spark came from work associated with LinkedIn and Databricks.
Ambient agents will handle small tasks before becoming a single companion
Mathew expects near-term AI use to involve many personal, enterprise, and hybrid agents that handle small, medium-sized, and large tasks. An agent could research and book a trip, process a repetitive task, or act across a user's applications. He does not expect the first version of this future to be one monolithic system that is always present. Instead, many agents may operate in the background and converge on the user's needs. A fully integrated, science-fiction-style companion may come later. For now, Mathew is more confident about ambient systems that quietly do useful work at home and in the enterprise.
"I actually see a lot more at least in the near term this emergence of massive amounts of personal Enterprise and almost High hybrid agents that are continuously doing little things medium-sized things big things on our behalf while we're talking here."George Mathew1:00:44
Who should watch
You are deciding whether to extend an existing enterprise application with agents or build a new AI-native workflow.
You run an ML platform and need to compare managed services with the people and maintenance costs of owning the stack.
You want a practical investor and operator view of model efficiency, domain-specific systems, and the next generation of AI software.
