That is half 1 of a 3-part collection about how HubSpot is remodeling with AI. Half 2 explains easy methods to develop with agent-first GTM. Half 3 is how we function as an AI-first firm.
All the things we construct at HubSpot exists that can assist you develop. So when generative AI arrived, our engineering group did not simply see a productiveness instrument. We noticed a chance to construct higher merchandise and get extra worth into prospects’ fingers quicker.
And once we reached the boundaries of off-the-shelf AI instruments, we did not simply search for one thing higher. We constructed a platform beneath it. That call turned bitter quicker than we anticipated. All of our AI is constructed on a shared basis, so each new function we carry makes the entire system stronger and offers our prospects a extra constant expertise throughout the whole lot they use.
Right this moment, we will innovate at a tempo that was beforehand not possible. 100% of our engineers are utilizing AI, and the variety of strains of code written by our engineers has elevated by 73%.
We did not get right here in a single day. It required three steps: actual infrastructure funding and a willingness to construct one thing that does not but exist. This is how.
Section 1: First Officer Productiveness Enchancment (2023-2024)
In 2023, large-scale language fashions have simply crossed the edge of being actually helpful in a coding context. The most effective answer for utilizing AI in engineering was to begin with what’s confirmed. Again then it was code completion. A human writes the code and an AI co-pilot suggests what occurs subsequent.
We rolled out coding co-pilot and shortly reached 30% adoption. We then extracted incident information and in contrast groups with and with out co-pilots to show that implementing AI didn’t negatively affect product reliability.
We used that information to take away guardrails and provides everybody co-pilot entry. Adoption charges exceeded 50% in a single day. This taught us a lesson in how we make selections. Measure, show, and scale.
By the top of Section 1, 80% of engineers had been utilizing AI instruments. Engineering velocity elevated by 51%. Which means engineers can now ship working code to manufacturing considerably quicker, leading to a 7% enhance within the variety of strains of code up to date per engineer. We have confirmed that AI can velocity up the work of each engineer with out compromising product reliability.
Section 2: Scaling with coding brokers (2024-mid 2025)
The subsequent step is autonomous coding by the agent. Our group can immediate instruments to finish end-to-end duties. Brokers might learn the context, write code, run checks, and repair errors, whereas engineers might assessment and supply directions. We felt strongly that this was the way forward for engineering and had been dedicated to it.
The true constraints got here shortly. Off-the-shelf coding brokers had no entry to inside construct methods, libraries, or the power to confirm that the code would truly work within the atmosphere. So we constructed these agent integrations in-house utilizing MCP, a normal that enables AI brokers to connect with exterior instruments and methods, and deployed them to all of our engineers. To speed up adoption, we hosted occasions that supplied a devoted area for engineers to be taught, experiment, and acquire confidence with new instruments. Agent utilization went from zero to 80% adoption in a single month.
The subsequent problem was scale. Engineers needed to run a number of brokers in parallel at evening with out supervision. So we constructed an agent execution platform on prime of Kubernetes infrastructure. All brokers run in remoted containers that replicate the precise HubSpot developer atmosphere. The agent compiles your code, runs automated checks, reads error output, and iterates by itself till the whole lot works. No human intervention required.
By the top of Section 2, 96% of engineers had been utilizing AI instruments, engineering velocity elevated by 60%, and the variety of strains of code up to date per engineer elevated by 48%. We began transport higher merchandise quicker by means of distributors. However that was only the start.
Section 3: Scaling with AI Platform (mid-2025 to current)
HubSpot’s platform strategy to product growth is how we have at all times created extra worth for our prospects. Once we constructed reporting and automation on the platform stage, we did not simply launch one function. We shipped that performance to all hubs on the similar time. That is how innovation turns into advanced.
In Section 3, we utilized the identical logic to our AI infrastructure. Somewhat than constructing each agent from scratch, we constructed a shared basis as soon as, together with how brokers entry information, what actions they’ll take, and the way they hook up with the remainder of HubSpot. All the things runs on it.
Consequently, all brokers are interoperable. They communicate the identical language, share the identical toolset, and draw from the identical context. Clients get a constant expertise regardless of which agent they use as a result of they’re all primarily constructed on the identical infrastructure. And since they’re all related, every new function will increase the worth of all the system. That is one thing that can’t be reproduced with a group of level options.
And it is made attainable by the best way we have augmented engineering with AI. 100% of our engineers now use AI, the variety of strains of code up to date per engineer has elevated by 73%, and the time to first suggestions on pull requests has decreased by 90%. Which means much less wait time and extra time to ship issues your prospects will truly use.
Why this issues: Improve buyer worth
The proper infrastructure accelerates the tempo of innovation. At HubSpot, our platform turns into extra highly effective with every agent we construct. Including context to the platform makes every agent much more efficient. For patrons, meaning merchandise that proceed to be higher, quicker and extra related.
Work that used to take months is now carried out in weeks, and people weeks translate immediately into new performance within the fingers of entrepreneurs attempting to succeed in the correct audiences, salespeople attempting to shut offers, and buyer success managers attempting to retain prospects. You do not want to consider the underlying platform. They simply expertise the outcomes.
