It is a section thought to be certainly one of a three-part collection on how HubSpot revamped with AI. Segment two covers how we broaden with Agent-first GTM. Segment 3 is how we serve as as an AI-first company.
The whole lot we assemble at HubSpot exists to lend a hand our shoppers broaden. So when generative AI emerged, our engineering workforce didn’t merely see a productivity tool; we spotted an opportunity to build upper products and get additional worth into shoppers’ palms faster.
And when off-the-shelf AI tools hit their ceiling, we didn’t merely seek for upper ones. We built the platform underneath them. That decision compounded quicker than we expected. On account of all of our AI is built on a shared foundation, every new capability we ship makes all the system additional tricky and shoppers get a additional consistent experience all through everything they use.
In recent times, we’re able to innovate at a pace that simply wasn’t possible forward of. 100% of our engineers use AI, and we’ve spotted a 73% build up in traces of code written via our engineers.
We didn’t get proper right here in one day. It took 3 phases, precise infrastructure investment, and a willingness to build what didn’t exist however. Proper right here’s how we did it.
Phase 1: Productivity with Co-pilots (2023-2024)
In 2023, large language models had merely crossed the brink of being actually useful in a coding context. The best answer for using AI in engineering was to start with what was showed. These days, it was code completion: a human writes code, and AI copilots counsel what comes next.
We rolled out a coding copilot and got to 30% adoption in short. Then we pulled the incident knowledge, when put next teams using the copilot against teams that weren’t, and proved AI adoption didn’t negatively affect the reliability of the product.
With that knowledge in hand, we removed the guardrails and gave everyone copilot get entry to. Adoption shot earlier 50% in one day. This taught us a lesson in how we make possible choices. Measure, end up, then scale.
By the use of the highest of Phase 1, 80% of engineers had been using AI tools. We spotted a 51% construction in engineering pace, that suggests engineers had been shipping working code to production significantly quicker, and a 7% build up in traces of code up to the moment consistent with engineer. We proved AI might simply make every engineer quicker without compromising product reliability.
Phase 2: Scaling with Coding Agents (2024-Mid 2025)
The next move was self enough coding with agents. Our teams might simply advisable the tools to complete end-to-end tasks. The agents might simply be told context, write code, run exams, and attach errors, all while the engineer reviewed and instructed. We felt strongly this was the future of engineering and devoted completely.
The actual constraint were given right here in short. Off-the-shelf coding agents might simply now not get entry to interior assemble strategies, our libraries, or check that code in reality worked within the surroundings. So, we built those agent integrations ourselves using MCP, an ordinary that allows AI agents to hook up with external tools and strategies, and deployed them to every engineer. To drive adoption, we organized events to supply engineers faithful house to be told, experiment, and assemble self trust with new tools. Agent usage went from 0 to 80% adoption in a month.
The next drawback was scale. Engineers wanted a few agents operating in parallel, in one day, without supervision. So we built an agent execution platform on best of our Kubernetes infrastructure. Every agent runs inside of an isolated container that replicates a real HubSpot developer atmosphere. Agents convey in combination the code, run automatic exams, be told error outputs, and iterate on their own until everything works. No human intervention required.
By the use of the highest of Phase 2, 96% of engineers had been using AI tools, engineering pace was up 60%, and lines of code up to the moment consistent with engineer had greater 48%. We’ve been starting to ship upper products quicker with agents. Alternatively that was only the start.
Phase 3: Scaling with our AI Platform (Mid 2025-Supply)
HubSpot’s platform strategy to product development has at all times been how we’ve created additional purchaser worth. When we built reporting and automation at the platform degree, we didn’t merely ship one function; we shipped that capability all through every hub similtaneously. That’s how innovation compounds.
We performed that exact same excellent judgment to our AI infrastructure in Phase 3. As an alternative of making every agent from scratch, we built the shared foundation once: how agents get entry to knowledge, what actions they can take, how they connect to the rest of HubSpot. The whole lot runs on best of it.
The outcome’s that all of our agents are interoperable. They talk about the identical language, percentage the identical toolsets, and draw from the identical context. A purchaser gets a relentless experience regardless of which agent they’re using on account of, underneath, they’re all built on the similar infrastructure. And because they’re all hooked up, every new capability we add makes all the system additional valuable. That’s something a number of stage solutions cannot replicate.
And it was made possible via how we’ve scaled engineering with AI. In recent times, 100% of our engineers use AI, traces of code up to the moment consistent with engineer are up 73%, and time-to-first-feedback on pull requests has dropped via 90%. That suggests a lot much less time in a position and overtime shipping problems shoppers in reality use.
Why this problems: Compounding purchaser worth
Having the correct infrastructure accelerates the pace of innovation. For HubSpot, every agent we assemble makes the platform additional tricky. Every piece of context we add to the platform makes every agent more effective. For customers, that implies the product helps to keep getting upper, quicker, and further hooked up.
What used to take months now takes weeks, and those weeks translate directly into new options inside the palms of marketers making an attempt to achieve the correct audience, reps taking a look to close provides, and Purchaser Excellent fortune Managers taking a look to retain shoppers. They don’t need to consider the platform underneath. They simply get to revel within the result.
0 Comments