How we Build with AI
This is part one of a three-part series on how HubSpot transformed with AI. Part two covers how we grow with Agent-first GTM. Part three is how we operate as an AI-first company.
Everything we build at HubSpot exists to help our customers grow. So when generative AI emerged, our engineering team didn’t just see a productivity tool; we saw an opportunity to build better products and get more value into customers’ hands sooner.
And when off-the-shelf AI tools hit their ceiling, we didn’t just look for better ones. We built the platform underneath them. That decision compounded faster than we expected. Because all of our AI is built on a shared foundation, every new capability we ship makes the whole system more powerful and customers get a more consistent experience across everything they use.
Today, we’re able to innovate at a pace that simply wasn’t possible before. 100% of our engineers use AI, and we’ve seen a 73% increase in lines of code written by our engineers.
We didn’t get here overnight. It took three phases, real infrastructure investment, and a willingness to build what didn’t exist yet. Here’s how we did it.
Phase 1: Productivity with Co-pilots (2023-2024)
In 2023, large language models had just crossed the threshold of being genuinely useful in a coding context. The best solution for using AI in engineering was to start with what was proven. At that time, it was code completion: a human writes code, and AI copilots suggest what comes next.
We rolled out a coding copilot and got to 30% adoption quickly. Then we pulled the incident data, compared teams using the copilot against teams that weren’t, and proved AI adoption did not negatively impact the reliability of the product.
With that data in hand, we removed the guardrails and gave everyone copilot access. Adoption shot past 50% overnight. This taught us a lesson in how we make decisions. Measure, prove, then scale.
By the end of Phase 1, 80% of engineers were using AI tools. We saw a 51% improvement in engineering velocity, meaning engineers were shipping working code to production significantly faster, and a 7% increase in lines of code updated per engineer. We proved AI could make every engineer faster without compromising product reliability.
Phase 2: Scaling with Coding Agents (2024-Mid 2025)
The next step was autonomous coding with agents. Our teams could prompt the tools to complete end-to-end tasks. The agents could read context, write code, run tests, and fix errors, all while the engineer reviewed and steered. We felt strongly this was the future of engineering and committed fully.
The real constraint came quickly. Off-the-shelf coding agents could not access internal build systems, our libraries, or verify that code actually worked in our environment. So, we built those agent integrations ourselves using MCP, a standard that allows AI agents to connect to external tools and systems, and deployed them to every engineer. To drive adoption, we …read more
Source:: HubSpot Blog
