How monday.com hit AI escape velocity by pausing its roadmap for 30 days

When a company pauses its entire product roadmap for 30 days to point 700 engineers at a single goal, it is a significant structural shift. For monday.com, that shift was the catalyst for transforming not only its product offerings but also its internal engineering culture, propelling it towards what its VP of R&D, Sergei Liakhovetsky, describes as “AI escape velocity.”

In a recent episode of the Dev Interrupted podcast, Liakhovetsky detailed how monday.com moved from treating AI as a series of isolated features to embedding it at the core of its platform. This journey involved fixing deep-seated technical debt, embracing a cell-based architecture, and orchestrating a high-stakes, month-long initiative that fundamentally changed how its technologists work.

The foundation: Trust as a currency

Before monday.com could integrate AI at scale, it had to address the underlying infrastructure that powered its platform. The company’s growth upmarket had revealed cracks in its performance and scalability.

“When we’re focusing on customer experience, we are looking at how actually our users will use the system,” Liakhovetsky explained.

This customer-centric focus drove the creation of mondayDB, a proprietary data management system that replaced the platform’s underlying technology. What started as a two-and-a-half-year project enabled monday.com to scale from handling thousands of items per board to millions. This was the first step in building a foundation of trust.

The company is now implementing a cell-based architecture designed to reduce the “blast radius” of incidents, ensuring that one noisy customer cannot take the entire system down. For Liakhovetsky, these foundational investments are about creating “highways” for future development.

“To build highways, it takes time, but in many cases, when you’re building this highway, you have unlimited speed later on,” he said.

This philosophy of balancing robust foundations with the need for speed became critical as the company turned its attention to AI.

The 30-day pivot: From features to transformation

Before the pivotal 30-day initiative, monday.com’s approach to AI was typical: teams developed features, but it was “not something that is transformational for anyone.” The leadership team, including Liakhovetsky, decided to change that by engaging the entire builders’ organisation in a concentrated effort.

The “AI month” was not a hackathon. From the outset, several key principles were established:

• Production-ready work: Everything built during the month had to have a clear path to production. This wasn’t about throwaway experiments.
• Honouring commitments: The leadership worked with dozens of teams to ensure that customer commitments were managed. In many cases, AI itself was used to accelerate existing roadmap items.
• Zero bureaucracy: To foster experimentation, the company removed barriers. A policy was established where anyone could request tools, but they had to become the champion for that tool. If usage dropped after two months, the tool was removed from the catalogue.

The decision-to-action timeline was remarkably short—just two weeks from conception to kick-off. Liakhovetsky emphasised that speed was critical.

“If you’ll start building the processes when you want to boost the mental mind shift, it’ll fail. Whoever is going to do like we did should make sure that from the point they decide they are all in on it, it should be immediately.”

The month saw 17 workshops, 22 speakers, and roughly 70 demos, creating an environment where engineers “started fighting” for the chance to present what they were building. The result was a suite of new AI products and internal tools that drastically improved efficiency.

A suite of AI solutions for every user

The AI month led to the creation of a cohesive ecosystem of products, each designed for different user intents and technical levels:

• Monday Magic: A tool that allows users to generate initial work solutions and boards using simple prompts, helping builders create reference implementations for things like a university library system or hospital shift schedules.
• Monday Vibe: An app builder that enables users to create custom applications on top of the monday.com platform, all without needing deep platform knowledge.
• Sidekick: A horizontal AI assistant that works across the entire platform. Users can prompt Sidekick to perform tasks like populating boards, managing items, or connecting different data sources.
• Agent Factory: A platform for building vertical, specialised agents that can handle specific workflows and roles. These agents can work in conjunction with Sidekick, sharing context to create a compound effect.

“Once we provide vertical agents on one side, solutions on another, and a horizontal copilot… we provide our users with the ability to decide how to work,” Liakhovetsky said.

This layered approach allows users of all technical levels to interact with AI in a way that fits their needs, while also providing monday.com with valuable data to improve context and user intent.

The new reality: GPU-bound reliability

The introduction of these AI capabilities did not just change the user experience; it fundamentally altered the demands on monday.com’s infrastructure. The platform’s shift from being CPU-bound to heavily GPU-bound introduced new challenges.

Liakhovetsky noted that agents interact with the system differently from humans, creating a “fan out” of API calls. This has forced the engineering team to rethink service-level objectives (SLOs), introducing concepts like a “fairness index” to prevent a single account from consuming all resources. Concurrency management and cost control have become critical.

“It’s a different type of SLOs here. When you are looking at the cost, at fairness index, at the concurrency and the agents utilisation, and how to deal with fan out here.”

This new reality also requires different security guardrails and API design, specifically optimised for agent use and the Model Context Protocol (MCP). The work on infrastructure is continuous, evolving in parallel with the AI products it supports.